Evan Armstrong calls it the institutional knowledge that makes coordination valuable: "the email threads, wiki pages, Slack channels, onboarding docs, and tribal knowledge where organizational truth actually lived." Foundation Capital goes further, arguing that the real prize is decision traces: not just what a company knows, but why specific decisions were made. The exceptions, overrides, precedents, and cross-system reasoning that currently live in people's heads.

Both are right. And both are missing a piece.

Armstrong captures it perfectly: "A markdown file can describe your sales process. It can't encode that deals over $500K stall when legal reviews before procurement." Foundation Capital adds: "The reasoning connecting data to action was never treated as data in the first place."

So here's the question neither article asks: if the most valuable context lives in people's heads: in contradictions they haven't articulated, in trade-offs they resolve differently, in assumptions they don't know they're making. How do you get it out?

Facilitation is the missing infrastructure

The context layer thesis is compelling. But it raises a question about how the context layer actually gets built. Armstrong's vision compounds through workflow traces: every agent execution adds context that makes the next one smarter. Foundation Capital's decision traces accumulate as agents observe cross-system reasoning. Both assume the hardest organizational knowledge will eventually surface through observation and accumulation.

But will it? The hardest, most valuable organizational context is pre-articulation — it exists as tacit knowledge that people carry but have never been asked to express. John Cutler captured this well: "Documents don't create meaning, people do. Process artifacts gain significance from the discussions, decisions, and behaviors surrounding them." His principle that "words, diagrams, and models only become meaningful through conversation and collective sense-making, not in isolation on a page" points at something the context layer conversation keeps missing: you can't just capture organizational meaning. You have to produce it, through structured interaction between the people who hold it.

This is what facilitation does. Not facilitation as a soft skill or a meeting technique, but as systematic infrastructure for extracting, structuring, and synthesizing human knowledge, especially when humans need to stay in the loop.

Consider what a structured facilitated session produces: each participant has a private 1-on-1 conversation with an AI facilitator that asks targeted questions, probes for specifics, and follows up on vague answers. The facilitator elicits reasoning, not "mere opinions". Why do you think that? What would change your mind? Where does your experience differ from what others might say?

This matters because, as Cutler puts it, "there is no single source of truth — only perspectives." Agreement on terms is not agreement on meaning, people can use the same words while holding different interpretations. A context layer built from workflow traces will never surface this. You need structured dialogue to discover that when the sales team says "qualified lead" and the product team says "qualified lead," they mean fundamentally different things.

Cross-pollination takes it further: as participants contribute, emerging themes and contrasting perspectives are shared across threads, prompting people to engage with viewpoints they wouldn't have encountered otherwise. The output is a synthesis of how a group actually thinks about a problem: where they agree, where they diverge, and why.

That synthesis is a decision trace. It captures the reasoning that Foundation Capital says was "never treated as data." And unlike the traces their portfolio companies capture from agent workflows, facilitation generates traces from the humans who carry the context that agents can't observe.

What we built

We've been working on making facilitation accessible to AI agents, and discovered a surprisingly good alternative to building everything in a web app.

Harmonica is a structured conversation platform: you create a session with a topic and goal, share a link, and each participant has a private 1-on-1 conversation with an AI facilitator. Responses are then synthesized into actionable insights. Sessions range from team retrospectives and brainstorming to stakeholder interviews and community consultations, any scenario where you need to hear from multiple people and make sense of what they said. Now it has not only a web UI, but also a REST API and an MCP server that lets AI agents create and manage sessions programmatically.

The MCP unlocks a completely new user experience. We have just shipped harmonica-chat, a companion for Claude Code and similar agents that makes use of MCP to design sessions, and realized that an AI coding agent is a surprisingly effective facilitation client. It already knows your project. It has your codebase, your recent commits, your CLAUDE.md, your conversation history. When you say "let's run a retro on the API redesign," it doesn't need you to fill out a form, it reads your project context and generates a session pre-loaded with relevant background.

This allows for a surprisingly smooth process design, more powerful than our standard "session creation flow" in a specific way: the context is already there. A user of CLI agent who needs to facilitate something — a team retro, a design review, stakeholder alignment — can do it without leaving their workflow. The agent handles template matching, goal refinement, context generation, and creates the session with a tailored facilitation prompt that understands exactly what's being discussed.

That last part matters more than it sounds. A facilitation session where the facilitator doesn't understand what's being discussed is exactly the gap Armstrong and Foundation Capital describe, playing out in miniature. Context-aware prompts are the difference between generic questions and ones that probe the specific trade-offs, history, and tensions a group actually needs to work through. For example, a session about “planning a local vibecoding meetup” now gets prompts that ask about neighborhood needs and community connections, not generic facilitation questions. The agent understands what it's facilitating.

But the MCP server unlocks more than just a better creation flow. Because it gives AI agents full programmatic access to sessions — creating them, reading responses, fetching summaries — it becomes a building block for experimentation. Maria Milosh, a researcher collaborating with us through the Open Facilitation Library, used our MCP server to implement a novel cross-pollination approach: a two-phase workflow where the first phase gathers ideas and extracts structured reasoning, and the second phase presents curated "packets" of others' perspectives back to participants for reflective dialogue. The MCP integration made it possible to orchestrate this entirely from an agent, no platform changes required.

The bigger picture

We think facilitation is infrastructure that should be as open and accessible as the context layer tools everyone is building. harmonica-chat and harmonica-mcp are both open source. Anyone can create and manage structured facilitated sessions or build new facilitation methods on top.

The research foundations matter too. Through our partnership with Metagov, we're contributing to the Open Facilitation Library — a research project working on open standards for AI-assisted facilitation. The library will include facilitation patterns, evals, and agent skills that inform how prosocial "deliberation agents" or collective response system could work. Our goal is to make AI facilitation a shared discipline with transparent, peer-reviewed methods, not a proprietary black box.

Armstrong's context layer compounds: every workflow execution adds traces that make the next one smarter. Facilitation compounds differently. Every session doesn't just produce a record of what people think, it changes how they think. Participants encounter perspectives they wouldn't have otherwise, articulate reasoning they hadn't formalized, discover disagreements they didn't know existed. Follow-up sessions build on previous findings. Over time, an organization develops a shared capacity for sense-making.

The context layer thesis is right: the most valuable software layer is the one that holds institutional meaning. But meaning doesn't accumulate passively. As Cutler warns, we should "focus on cultivating the conversations you want to see, not just building systems that track activity." The context layer needs more than observation and workflow traces. It needs facilitation infrastructure: for the messy, irreducible process of people making sense of things together. That's the piece we're building.

"The newest theories of digital democracy no longer focus so heavily on optimal voting methods, cryptography, etc. Those parts still matter, but nowadays a large share of thinking on the topic has shifted in a different direction: what kind of infrastructure would enable a high-quality 'public conversation'?"

He pointed to tools like Polis — where anyone can post a statement, everyone votes, and AI algorithms identify the ideas most people agree on — as examples of systems that make it possible to find societal compromises directly, without intermediaries.

But then he made a deeper point that resonated with everything we've been learning at Harmonica:

"In order to build democracy, you need people who have the habit of behaving 'democratically.' Democracy cannot simply be installed the way intellectuals plug a USB drive into a computer. I still believe this is the main reason why blockchain teams that say 'first we build a high-performance system, then we decentralize' never actually succeed: the habit within the community of 'letting the core team do everything' becomes far too strong."

This is exactly the problem we've been mapping. Not in nation-states, but in the web3 organizations that were supposed to be governance laboratories: DAOs. And the findings are sobering.

gov/acc

Call to Action If you want to contribute to gov/acc, please consider completing a Harmonica session to map open problems DAOs are facing, their potential solutions, and who is working on them (whether researching or building).

https://gov-acc.metagov.org

gov/acc Phase 1: Preliminary Results

In January 2026, we launched gov/acc (governance acceleration) — a research program with Metagov to map the real problems web3 governance practitioners face, the solutions being tried, and the actors building them. Phase 1 consists of structured interviews conducted by Harmonica, where we ask governance leads, delegates, researchers, and ecosystem builders to share their insights.

The result: 11 governance problems, 29 proposed solutions, and 41 actors — all published as simple Quartz wiki articles and mapped into an interactive dashboard. Each problem and solution also has its own detailed wiki article on our Quartz site — covering evidence, related problems, key actors, and maturity assessments — so the dashboard serves as an illustration to a deeper knowledge graph that will be enriched with other data sources besides Harmonica sessions.

The Problems / Urgency Matrix

The Problems tab ranks all 11 governance problems by urgency — a composite score combining breadth (how many of our 27 participants raised it) and depth (how many messages they spent discussing it).

Token Voting Failure & Plutocracy sits at the top with an urgency score of 86 out of 100. 74% of participants raised it — the clearest signal in the dataset. The pattern is familiar: token-weighted voting concentrates power among whales, making governance outcomes a function of capital rather than competence or community alignment.

But what's more interesting is what clusters around it. Governance Theater (score: 71) — the phenomenon of proposals going through motions without real deliberation — is the second-most urgent problem. Voting Fatigue (score: 59) compounds both: when every decision requires a token vote but the outcome feels predetermined, people stop showing up.

These three problems form a vicious cycle. Bad voting mechanisms → theatrical processes → fatigued participants → even more concentrated power. This is precisely what Vitalik described: the habit of "letting the core team do everything" becomes too strong.

Further down the list, but with striking depth scores: Over-Reliance on Game Theory (7 participants, but 6.5 messages on average — the deepest discussions in our dataset) and Institutional Amnesia (6 participants, 5.8 depth). These are specialist concerns, but the people who raise them are deeply engaged — a sign that these problems are felt acutely by those closest to the work.

The Urgency Matrix plots each problem on two axes — breadth (X) and depth (Y) — revealing patterns invisible in a ranked list.

Token Voting Failure occupies the top-right quadrant: both widely recognized and deeply discussed. That's why it's #1.

But look at the other quadrants. Voting Fatigue is wide but shallow (14 participants, low depth) — universally recognized, well-understood, but not generating much new thinking. People know it's a problem; they just don't have novel solutions.

Over-Reliance on Game Theory is the opposite: narrow but deep. Only 7 people raised it, but when they did, the conversations were the richest in our entire dataset. This is the kind of finding that gets lost in simple polls or surveys — and the kind that structured deliberation surfaces well.

The Solutions / Wardley Map

The Solutions tab maps all 29 proposed solutions on a Wardley Map — positioning them from Genesis (experimental) through Custom and Product to Commodity (widely adopted).

The most striking pattern: the execution layer is far more mature than the deliberation layer.

Financial solutions like VE Buyback Models and Contributor Streams are already at Product or Commodity maturity. Structural approaches like Specialized Committees and Optimistic Governance are solidly in Product territory.

But sensemaking tools — the infrastructure for "high-quality public conversation" that Vitalik calls for — remain largely experimental. Polis Opinion Clustering, Signals Protocol, and deliberation platforms like Updraft sit at Custom maturity. Knowledge infrastructure (Governance Memory Systems, reference tools) is still at Genesis.

This is the governance gap. We're good at counting votes. We're bad at the conversation that should happen before anyone votes.

AI-augmented governance is emerging fast — AI Governance Agents received 5 mentions, unusually high for a Genesis-stage solution — but it's still early. The question isn't whether AI will play a role in governance infrastructure, but whether it will be used to automate voting (dangerous, as Daniel Ospina has warned) or to support better deliberation.

The detailed Wardley Map makes the opportunity clearest. Solutions in the bottom-left (Genesis, low in the value chain) represent the biggest gaps between problem urgency and solution maturity.

Knowledge infrastructure — governance memory, institutional learning, context preservation — is the largest disconnect. Institutional Amnesia scores 54 on urgency with 5.8 depth, but the solutions addressing it are almost entirely at Genesis stage. Organizations keep losing hard-won governance knowledge because no one has built the infrastructure to preserve it.

Who's Building What

The Actors tab visualizes 41 projects and individuals as a network graph, with nodes sized by mention count and color-coded by category: governance infrastructure, reputation & identity, sensemaking, financial, security, research, and service providers.

What stands out is the clustering. Governance infrastructure actors (Aragon, Tally, Agora) form one cluster. Sensemaking actors (Polis, Lighthouse, Mission Public) form another. Financial actors (Karpatkey, Steakhouse) sit separately. The connections between clusters are thin — teams working on voting infrastructure rarely overlap with teams working on deliberation tools.

This siloing mirrors the maturity gap from the Solutions tab. The ecosystem has invested heavily in execution infrastructure but underinvested in the deliberation layer that should feed it.

From Problems to Solutions to Actors

The Sankey Diagram ties everything together, showing flows from problems through solutions to actors. You can trace, for example, how Token Voting Failure flows through Reputation & Merit Systems to projects like DeepDAO and Kokonut Network. Or how Institutional Amnesia flows through Governance Memory Systems to Metagov and related research projects.

The thickest flows — the most-connected problem-solution-actor chains — tend to run through execution-layer solutions. The deliberation layer has thinner flows, fewer actors, and less investment. This is the infrastructure deficit.

Building the Habit

Vitalik's insight about democratic habits applies directly to what we're seeing. The web3 ecosystem has spent years optimizing the mechanism of governance (token voting, delegation, on-chain execution) while underinvesting in the practice of governance (deliberation, sensemaking, shared understanding).

The result is exactly what he described: communities that have the technical infrastructure for decentralized decision-making, but not the habit. And so the core team keeps deciding.

Tools like Polis — incl. Harmonica — represent a different approach. Instead of optimizing how votes are counted, they focus on the quality of conversation that happens before anyone votes. They build the habit of collective sensemaking. They create the infrastructure for communities to discover their own compromises and consensus, "directly, without intermediaries."

The data we've collected as part of Phase 1 confirms that this is where the gap is widest and the opportunity is greatest. The deliberative layer — the infrastructure for high-quality public conversation — is where governance acceleration needs to happen.

gov/acc is our latest collaboration with Metagov. Explore the full interactive dashboard at gov-acc.metagov.org. Phase 1 is ongoing — if you work in governance and want to contribute, you can join a structured conversation to share your experience.

Our technical lead quit in June. Development slowed to a crawl. We're still pre-revenue. We didn't raise the pre-seed round we'd hoped for. There were months where we weren't sure if we'd make it to the end of the year.

And yet, here we are. Still building Harmonica. Still believing that teams and communities deserve better ways to listen to (if not deliberate with) their stakeholders than forms and surveys.

After 12 months of public alpha, we're entering public beta with the launch of Harmonica Pro. It's not the triumphant "we crushed it" announcement you usually see from startups. It's more like: we survived, we learned a lot, and we're clearer than ever about what we're building toward.

Here's the honest story of our year.

Paris and Lisbon

The year started with a pilot in Paris. Missions Publiques used Harmonica as part of a hybrid workshop with ~100 students of the École Normale Supérieure (in parallel to the AI Action Summit). They wanted to explore how AI might shape a deliberative process focused on the risks and opportunities of AI from the students' perspectives. It was small, experimental, and exactly the kind of use case we'd dreamed about. Yale's Institution for Social and Policy Studies wrote a blog post about it — our first real external validation that this thing we were building mattered to people beyond our tiny team.

Working on the ground with a real deliberative process revealed gaps in our product we hadn't seen from the outside:

• Professional facilitators need more control. We watched manual documentation create bottlenecks in real-time. That's why we added editing of system prompts and started developing AI-powered transcription for voice input.

• Cross-pollination matters. During the workshop, some of the richest moments came when groups could react to statements from parallel discussions. We've since built follow-up questions powered by vector search, so participants can build on each other's ideas in async sessions.

• Real deliberation spans multiple sessions. Complex processes need shared knowledge and continuity. Our projects feature came directly from seeing how hard it was to maintain context across a multi-day event.

We also participated in the OECD bootcamp in Lisbon. Members of our team helped build 2 out of 3 prototypes that were selected by “challenge owners” from EU governments. Then nothing happened.

We learned something important: passion doesn't pay invoices. The deliberative democracy space has funding: we've seen the grants, the EU programs, the institutional budgets. But between the bureaucracy, the closed networks, and possibly the extra friction of having a Russian-named founder in a space that runs on trust and warm intros, we couldn't convert excitement into contracts. We needed a different path.

Berkeley and Edge Esmeralda

Meanwhile, Andrea Gallagher, who has been supporting us throughout all of 2025 and deserves more gratitude than we can express here, represented Harmonica at two high-profile events on the other side of the pond.

First, the workshop in Berkeley organized by Plurality Institute and the Council on Technology & Social Cohesion. Over 70 researchers and technologists gathered to map the landscape of LLM tools for public discourse. That gathering eventually produced the report "Mapping LLM Tools for Public Discourse, Pluralism & Social Cohesion", where Harmonica was mentioned as one of the new tools for consensus building through AI-powered facilitation.

Not gonna lie: being included alongside platforms from Google, academic institutions like MIT and Stanford, and leading civic tech organizations felt surreal for a three-person team running on pure enthusiasm.

Another highlight of the year: Harmonica was used at Edge Esmeralda to generate seed statements for a Polis conversation. Audrey Tang discussed our approach during the event. Moments like this remind us why we're doing this work, seeing Harmonica create value in real deliberative processes with people who've spent decades thinking about augmented deliberation and conversation networks.

Summer: Scroll and Strategic Pivot

Our technical lead left. Suddenly Harry had to step up his vibecoding game. Development slowed dramatically. Features we'd planned for summer got pushed to fall, then to "someday." Startups don't usually talk about this stuff. But it's the reality for most early-stage teams, and pretending otherwise would be dishonest.

Our largest deployment came in July: Scroll's Co-Creation Cycle 3, where we facilitated organizational design for their DAO. Over two weeks, 40+ delegates used Harmonica to explore key domains, prioritize initiatives, and deliberate on optimal org design, combining async AI-facilitated surveys with live workshops.

The results surprised even us. One delegate put it perfectly:

"When I interact with the Public Forum, I always measure my words. But this interface is really cozy. It feels like ChatGPT. I can be more assertive and honest. This is the first time I'm writing with more subjectivity. I don't have to consider the DAO members' susceptibility as an organization and the subsequent political tensions. You can be direct and brief. AI could be a neutral field in which to express opinions."

Spanish-speaking delegates could respond in their native language — not just translation, but genuine accessibility that surfaced perspectives we'd never have captured in English-only surveys. We adjusted prompts mid-session when early responses revealed our initial framing wasn't working. We pre-populated Miro boards with AI-synthesized insights so workshops could focus on decisions, not brainstorming from scratch.

The CCC3 produced real outcomes: clear prioritization across six domains, actionable recommendations for council formation, and a validated framework Scroll is now using for future governance iterations.

We're deeply grateful to Eugene Leventhal, Scroll's head of governance at the time, whose forward-thinking approach made this possible. Not every governance lead would bet on an untested AI facilitation tool for a major deliberative process. Eugene did, and he's continued championing deliberative-first governance in web3, which he presented in his keynote at DevConnect in Argentina.

But Scroll also taught us something important: our customers need facilitation services, not just a tool. Great technology isn't enough if organizations don't know how to design deliberative processes. This realization is shaping our 2026 strategy—we want to partner with expert facilitators and co-create templates that encode their expertise.

Foundation Sprint

We did a Foundation Sprint with Chris Lunney, based on a new methodology from the authors of "design sprint", which helped us refine our positioning and strategy. Sometimes you need to stop building and figure out what you're actually building for.

We spent weeks interviewing our primary ICP: consultants. After around 20 conversations, we realized there was no problem-solution fit. Consultants confirmed the pain of "harvesting deeper insights from many stakeholders" but weren't eager to pay for our solution.

Hence our latest pivot: change management. We are shifting our focus to selling Harmonica to companies, particularly change management leaders and people ops who need to engage employees at scale.

In 2026, we’re aiming to build a community of service providers who can design processes and facilitate workshops for our enterprise customers, and who want to create Harmonica templates for their own use or for our future template store. We're also exploring an affiliate model where consultants sell Harmonica to their clients and we share revenue.

Fall: Research Partnerships

An idea we first discussed with Colleen from AI Objectives Institute (and presented at the Interoperable Deliberative Tools symposium in January) has slowly evolved into a multi-track research project.

Seminars with experts. Metagov helped us organize seminars with amazing experts such as Andy Paice, Lisa Schirch, Jorim from Dembrane, Alice Siu from Stanford's Online Deliberation Platform, and researchers from Google Jigsaw.

Evals and prompt experimentation. We started working with Maria Milosh, a talented data scientist from Chicago who is even more passionate about deliberative democracy than we are. She's been helping us develop rigorous evaluations and experiment with different facilitation approaches.

Cross-pollination research. Maria partnered with Oxford Computational Political Science Group to study how different approaches to synthesizing and sharing insights between participants affect social choice in deliberative processes. We also connected with Asya Gergel through this project, strengthening our agent-building capabilities.

Cooperative AI fellowship. Joseph Low, another contributor to OFL, won a fellowship at the Cooperative AI Foundation (yay!). He'll be researching AI facilitation with mentorship from Michiel Bakker, well-known researcher at DeepMind and co-creator of the Habermas Machine.

We are also happy to extend our collaboration with the Metagov ecosystem by powering an experimental version of Harmonica with Apertus, the Swiss LLM accessible through Public AI inference utility, exploring what it means to run AI-powered sensemaking on public infrastructure.

The People Who Made This Year Possible

We're a tiny team, but we're lucky to work with brilliant people:

• Jonas — our full-stack developer from New Zealand, who has been responsible for our backend and DevOps

• Chris — product designer and facilitator from Brooklyn, helping us refine our GTM strategy and building out facilitation services

• Umar — data scientist and engineer who used to work with Harry at Gitcoin, helping us get serious about analytics

• Maria — helping with evals and cross-pollination workflows, bringing rigor to our ML pipelines

And of course our advisor Andrea Gallagher, who has believed in Harmonica since before it made any sense to believe in us. Thank you so much.

What's Next

We're entering 2026 with clear priorities:

Harmonica Pro is live. New signups go to Pro now; we'll migrate everyone in January. Now there's a free tier with limitations, and a Pro tier for teams who need Harmonica for more serious projects. If you're a nonprofit or community organization, do reach out, we're committed to giving you free access until we add BYOM (so you’ll be able to power Harmonica with LLM of your choice, including public AI).

Facilitation partnerships. We're looking for expert facilitators who want to work with us to design and run workshops with our clients and co-create templates. The tool is only as good as the process design behind it.

Pre-seed round. We're preparing to raise in early 2026. If you're an angel investor interested in the future of work, collaborative tech and sensemaking, we'd love to talk.

The Bigger Picture

In 2026, multiplayer AI will eat single-player AI. Greg Isenberg and Fareed Mosavat from a16z have both made this call recently, pointing at the same pattern we've seen before. Google Docs beat Word, Figma beat Sketch, Notion beat Evernote.

A few weeks ago Anthropic introduced their "AI interviewer", proving there's massive demand for conversational research at scale. But their approach is single-player: one person, one AI, one conversation in isolation.

We think the next frontier is AI that facilitates structured conversations across many people, asynchronously, and synthesizes the patterns and tensions that emerge. Not replacing human facilitators, but making great facilitation accessible to every team, at any scale.

Harmonica 🪗

@harmonica_chat

2026 is shaping up to be the year of multiplayer AI. @gregisenberg and @far33d have made this call in recent weeks, pointing at the same pattern we've seen before: G Docs > Word, Figma > Sketch, Airtable > Excel. All single-player tools eventually lose to multiplayer counterparts

6

12:24 PM • Dec 17, 2025

2025 was a year of learning what doesn't work. 2026 is about shipping what does.

As Harry put it: "We can't perfect everything. New Year's resolution: ship."

See you on the other side.

Happy holidays from Harry and Artem!

At Harmonica HQ, we've been looking back at our journey through 2024, counting not just the features we've shipped but also the new relationships we've started to build, and the small but meaningful wins that help us remain optimistic about humanity's capacity for dialogue and cooperation.

Thank you for ongoing support!

First of all, we want to express our profound gratitude for your contributions in the last Gitcoin Round. With over 500 contributions, we raised $7,051 in total! One of the most surprising and humbling moments came when we found ourselves ranked among the top 10 OSS projects. For a young project like ours, this level of community support was absolutely incredible. While some projects secured even larger grants with fewer contributions due to their contributors' Gitcoin Passport scores (a learning we'll definitely take into account for future rounds), we're incredibly grateful for every single person who demonstrated their belief in our mission.

As with the previous rounds, we're fully committed to deliver on the milestones outlined in our grant application: (1) Core reasoning capabilities and memory, (2) Improved session reports based on selected template and user data, (3) Less linear dialogues and cross-pollination between participants.

What's new in Harmonica

Since the last newsletter we've shipped some new features and quite a few QoL improvements:

🤓 New Backend

• Switched from Make.com to proprietary backend (TypeScript + Vercel)

• Moving to new backend enables us to start using agents and vector DBs for better reasoning

🔐 SSO Authentication

• You can now log in and access all of your hosted sessions in one place

• Seamless session management and history tracking

📑 Ready-to-Use Templates

• Pre-built facilitation templates to address specific sensemaking and collaboration use cases:

  • Retrospectives, Brainstorming, Community Proposals, OKRs planning, SWOT analysis, Theory of Change, and a few others

• Soon, these templates will also shape how session results are structured and presented!

📊 Working with Different Participant Groups

• Collect participant data with customizable forms

• Get deeper insights into participant groups / clusters

• Coming soon: Filter and analyze responses based on properties

🎨 Improved User Experience

• Page titles and session links with correct metadata

• Updated Help Center and pro tips inside the app

• Session Recap (template brief) displayed on the results page

• Save Drafts before launching sessions

• Session stats displayed correctly

• Better exports to JSON, Markdown coming soon

• Anonymous participants are easier to work with

• More responsive and clean mobile UX

We've squashed many bugs, but there's still loads to squash! Please report them via #support channel on our Discord server or our Harmonica Support group on Telegram.

Positive feedback that keeps us going

We're energized by the growing impact our MVP is producing across different communities and organizations. Here are some highlights that keep us motivated:

As we mentioned in our previous post, we're collaborating with CMI, the renowned peace-building organization from Finland. They've recently launched a global stakeholder survey expected to receive 100+ responses in multiple languages - a true acid test for Harmonica's capabilities. Even more exciting, CMI recently presented Harmonica during a workshop with the UN Special Envoy's team in Yemen, generating significant interest in our approach to async sensemaking.

Although Harmonica doesn’t have any onchain functionality yet, we’re very excited to receive early positive signals from some web3 organisations. Working with Scroll has been particularly inspiring, as their new governance leadership is pioneering deliberation-first approach. We're particularly excited to collaborate with Antoine Vergne, who brings decades of experience in designing and running citizen assemblies across Europe. Together with RnDAO and a few professional delegates, we're exploring how async collective dialogues can increase participation and enhance decision-making in DAOs. In 2025 we're planning to double down on the specific needs of such communities, e.g. with custom UX for designing grant programs and reviewing applications.

We're especially grateful to our early adopters from relevant communities of practice such as Metagov, Citizen Collective (instigated by Jon Alexander) and Sensemaking Scenius (instigated by Kristen Pavle and Ronen Tamari). Your feedback has been instrumental in shaping Harmonica into a professional-grade tool that maintains the human touch in digital facilitation. Here's the recording of a recent town hall with amazing like-minded humans:

Harmonica Demo @ Sensemaking Scenius

Harmonica Demo Session Chat charles adjovu 18:13 @Artem Zhiganov That thesis sounds awesome. Do you have a link to the thesis? Spencer - Clinamenic LLC 18:16 Very interested in governance models which shift the "locus of participation" more from governance to sensemaking, i.e. participatory sen...

http://docs.google.com

Overall, we're seeing growing interest from communities exploring new ways of stakeholder engagement, deliberative democracy activists and grassroots organizations (primarily UK-based) and distributed teams looking to upgrade their collaboration stack. The common thread here is the demand for better, faster, and cheaper facilitation, and the highest priority for us is to build a useful tool for professional facilitators, whether they run citizen assemblies or implement advanced project management at the intersection of Cynefin, Wardley Mapping and Team Topologies.

Interoperability in practice: laying out the foundations of AI-powered facilitation

We've spent the past few weeks deep in discussion with ML engineers, professional facilitators, and governance experts to understand how LLMs can meaningfully enhance facilitation while preserving its essential human elements. These conversations have led us to some exciting insights about the architecture of AI facilitation - and we're thrilled to announce that we're taking the next step by collaborating with Metagov and AI Objectives Institute on an open-source library of AI facilitation patterns.

This collaboration brings together technical expertise in AI development with deep experience in real-world facilitation and democratic processes. AI Objectives Institute plans to implement these patterns in Talk to the City, creating additional practical testbed for our shared library.

Our vision for AI facilitation centers on several key principles:

1. Goal-Oriented Sessions: Every session starts with clear objectives, helping both participants and AI understand what success looks like. The AI can use these goals to develop its facilitation strategy and potentially come up with original solutions.

2. Dynamic Response: Moving beyond simple Q&A, we're implementing systems that can analyze previous responses in real-time and generate relevant follow-up questions. This creates more natural conversation flows and helps surface important insights and outliers.

3. Cross-Pollination of Ideas: A core feature we're developing is the ability for the AI to connect relevant thoughts and perspectives between participants, fostering richer discussions and helping build consensus.

4. Flexible Facilitation Strategies: We're creating a library of facilitation tactics that the AI can deploy based on the session's needs - from ice breakers to deep-dive questions, from divergent thinking exercises to convergent rankings and bridging.

By developing these components as an open-source library, we're ensuring that other projects can benefit from and contribute to this codified facilitation wisdom. This approach allows us to pool expertise and resources while working toward our shared goal of making public sensemaking / deliberation more accessible and effective.

The library will include various facilitation modules, from non-violent communication patterns to structured deliberation techniques. In line with requirements of Metagov's IDT cohort, we're designing these components to be modular and interoperable, making it easy for different projects to adapt them to their specific needs while maintaining high standards for facilitating meaningful group conversations.

This collaborative approach represents a significant step forward in developing AI facilitation tools that truly serve the needs of communities and public institutions while ensuring transparency and democratic integrity.

What's Next for Harmonica

We're excited about upcoming events in February:

• Our advisor Drea Gallagher will be participating in the workshop organized by Plurality Institute, Prosocial Design Network and The Council on Technology and Social Cohesion on February 27-28 in San Francisco

• Two members of our team are looking to participate in the OECD's Co-Creation Bootcamp on February 25-26 in Lisbon

These opportunities will help us further refine Harmonica with input from some of the world's leading thinkers in collective intelligence and digital democracy.

And in Q1 2025, we're launching several exciting initiatives to bring Harmonica closer to facilitators:

• Harmonica Tunings: Join our live demo sessions on Zoom where we explore different use cases for AI-powered facilitation. Each week, we'll dive into new use cases and possibilities unlocked by our tool.

• Interviews with Experts: We're honored to interview leading voices in facilitation, governance and civic tech, including Andy Paice, Grace Rachmany, Antoine Vergne, Eloïse Gabadou Santiago, Danny Spitzberg, Liz Barry, Simone Maria Parazzoli, Nathan Hewitt, Antti Pentikainen, and many more inspiring practitioners.

• Public Sensemaking: We're planning to run open, large-scale dialogues on timely global issues. These sessions will demonstrate how AI facilitation can help communities tackle complex challenges while bringing people together across differences.

Support Our Mission

As we continue developing our open-source sensemaking tool, we're approaching a critical phase in our lifecycle. Releasing Harmonica's MVP in early November enabled us to validate our key hypotheses re: the demand for AI-facilitated async workshops, and we're seeing clear paths to sustainable impact. At the same time, bootstrapping is hard and most of us have been working on Harmonica part-time. We're ready to switch gears to fully dedicate ourselves to this project, fill the critical roles in our team and serve more communities and organizations.

How you can help:

1. Direct Support

   • Contribute via Giveth (crypto)

     Harmonica - AI agents for group sensemaking

     Harmonica enables fast sense-making by having 1:1 dialogues with your team members, using advance

     https://giveth.io

     

   • Contribute via Open Collective (GBP), hosted by Social Change Nest

     Harmonica - Open Collective

     Open-source AI agents for superfast sensemaking and deliberation.

     https://opencollective.com

     

2. Investment Opportunity

   In early 2025, we'll be raising a pre-seed round. We're seeking mission-aligned investors who share our vision for collective sensemaking with open core business model. If you know potential angels who might be interested in supporting a project like Harmonica, please connect us. We're particularly interested in meeting angels who understand the paradigm shift of AI agents and their crucial importance for the future of work and democracy.

As we wrap up 2024, we feel more energized and purposeful than ever. With each new partnership and collaboration, our vision becomes clearer and our conviction grows stronger. We're finding our tribe - passionate people and organizations who understand the importance of better collective sensemaking. Thank you for being part of this journey and helping us build a more useful tool for everyone.

Happy holidays!

Team Harmonica 🪗

Harmonica | AI-powered Facilitation

Harmonica acts as a virtual facilitator by talking to your team, event guests or discussion participants 1:1 and using LLM to synthesize their responses. It's an alternative to forms and saves both you and participants time, builds healthy habits and reduces friction.

https://www.harmonica.chat

From this brand-new web app to participation in the latest Gitcoin Grants Round, there’s a lot to be excited about. Here’s everything you need to know.

This release marks a significant milestone in our journey, made possible by the incredible support we received through the CollabTech round in Gitcoin Grants 21 back in August. Finishing second with approximately $3,800 in total funding, we focused on a clear objective of enabling our users to create a session with any facilitation logic in minutes (without any prompt engineering involved).

We're proud to say we've delivered on that promise. The current version features:

• A streamlined session creation flow for hosts (using GPT-powered prompt generator)

• Mobile-friendly web chat UX for participants

• Session results page with automatic summary updates

• Hosts can edit prompts and interrogate results using GPT assistant

• Exporting session results to JSON files

• Free access without login

We're excited to announce our collaboration with our first institutional customer CMI — Martti Ahtisaari Peace Foundation, an independent Finnish NGO that works to prevent and resolve conflicts through dialogue and mediation. CMI is currently testing Harmonica as a tool for both internal team collaboration and external deliberative processes. Their mission aligns perfectly with our vision for Harmonica, and we're honored to support their important work.

This week marked another important milestone as we participated in the Metagov's Interoperable Deliberative Tools cohort workshop in San Francisco. Our frontend developer Jonas represented Harmonica, demonstrating our platform and engaging with other projects in the space. The workshop proved invaluable for understanding how to better integrate process cards developed by Aviv Ovadya and exploring interoperability opportunities with other tools.

Our new JSON export functionality is a direct result of insights gained from this program. We're excited about the potential for Harmonica's outputs to be utilized by other tools in the deliberative tech ecosystem, creating a more connected and efficient landscape for community decision-making.

While we're proud of our first public release, this is just the beginning. Our roadmap focuses on delivering the core promise of Harmonica: truly intelligent AI-powered facilitation. Over the next weeks, we'll be working on making it real:

1. Context-Aware Conversations

   • Implementation of RAG (Retrieval-Augmented Generation)

   • Using previous sessions as context

   • Workspaces and organizational context awareness

2. Enhanced Facilitation

   • AI agents for deeper dialogues (i.e. better questions based on previous responses)

   • Cross-pollination of ideas across participants

   • Library of templates curated by experienced facilitators

Our goal is to create a facilitation experience that goes beyond linear conversations (that currently don't differ that much from vanilla ChatGPT), enabling a more profound alignment and robust idea synthesis within teams and communities.

As we continue building, we're participating in the current Gitcoin's GG22 round for open-source apps. While we're currently ranked 6th among OSS apps, we face an interesting challenge: our supporter base is relatively small, which significantly impacts our potential matching funds due to the quadratic funding mechanism.

In quadratic funding, the number of contributors matters more than the size of contributions. This means that even a $1 donation can make a meaningful difference. Projects with 1000+ small contributions will receive significantly more funds from the round's matching pool of $300,000 than those with just a few dozen donations of equal total value.

1. Visit our Gitcoin page before November 6th.

2. Contribute any amount. Even small contributions go a long way in helping us secure a greater portion of the matching pool.

3. Spread the word. Share our Gitcoin page with your network to help us reach more supporters who believe in our vision of async sense-making and deliberation.

Every contribution, no matter how small, helps us build Harmonica as an open-source software for async sensemaking and deliberation.

We're incredibly grateful for the support that has brought us this far, and we're excited about the road ahead. Try out the public beta at https://www.harmonica.chat/ and let us know what you think!

Thank you so much for being part of our journey,

Artem on behalf of Team Harmonica

Executive summary

Initial sense-making in small groups (e.g. core teams) is currently the most underserved need in online governance. It’s currently addressed with solutions like interviews, workshops and surveys, none of which enable fast and scalable sense-making. This problem space relies heavily on facilitation skills of people responsible for governance or strategy. Instead of optimising voting and delegation systems (that re-create the problems of the outdated real-life democracies), the most impactful solutions will address the problem of facilitation. Modern LLMs have the potential to make it much easier for organisations to identify tensions, goals, and generate collective intelligence artefacts that would ensure higher alignment and legitimacy of their decisions. This paper is an attempt to bridge the gap between decentralised governance and deliberative democracy as they face a lot of common challenges and both can benefit from deploying AI / LLMs at scale.

Decentralized organizations aim to enable participatory and distributed governance and decision-making. However, facilitating effective collective sense-making in this context remains a significant challenge. Sense-making is the collaborative process of developing shared understanding and framing of complex issues to inform decisions and actions. It is a core capability of an organization or society.

Inadequate sense-making can lead to an increasing gap between the complexity of the environment and the ability to understand and address it. This can result in apathy, resignation, or oversimplification of problems. Sense-making occurs at both personal and collective levels. Personal sense-making requires cognitive capacity, time, attention, and open-mindedness. Collective sense-making involves aggregating preferences and consolidating expertise to make competent group decisions, resulting in shared worldviews and lore.

In decentralized organizations, sense-making is critical for:

• Formulating shared purpose, priorities, and goals

• Deliberating on key decisions and tradeoffs

• Interpreting ambiguous information and dynamic environments

• Adapting strategies based on feedback and learning

However, current approaches to collective sense-making in decentralized groups face limitations. Engaging diverse members, eliciting quality inputs, and synthesizing insights is time- and effort-intensive. Lack of structure and facilitation skills can lead to inefficient and inconclusive processes, hindering groups from leveraging their collective intelligence and resulting in misalignment and lack of clarity.

Aviv Ovadya suggests that to address urgent global crises and ensure that decision-making systems incorporate human values, we must improve these systems to be competent, aligned, and robust — in other words, wise. A wise decision-system should be skilled at identifying and evaluating options, aligning across conflicting values, and remaining resilient in the face of real-world complexity and adversarial forces. Improving decision-systems requires integrating insights from various domains, considering the properties of the decision-system itself, and the legitimacy and trust it holds in the world outside the system.

Two papers on facilitation in the context of deliberative democracy are particularly useful: (1) Facilitators: The Micropolitics of Public Participation and Deliberation (Escobar, 2019) and (2) Diversity in Facilitation: Mapping Differences in Deliberative Designs (von Schneidemesser et al., 2023). Both papers highlight the importance of facilitators in shaping participatory and deliberative processes. Von Schneidemesser et al. focus on comparing different facilitation approaches and their impact on deliberative outcomes, while Escobar takes a broader view of the role of facilitators in democratic practices and the micropolitics of their work. The papers converge on the idea that facilitation is not a one-size-fits-all practice and that different approaches may be suitable for different contexts and desired outcomes. However, Escobar's chapter goes beyond the comparative approach by delving into the political nature of facilitation work and the challenges of institutionalising and professionalising this role within governance structures. Escobar's focus on the micropolitics of facilitation provides a valuable lens for understanding the complex power dynamics at play in participatory processes.

Based on the insights from the academic articles, here are some recommendations for practitioners and developers building innovative facilitation solutions:

1. Tailor the facilitation approach to the context and desired outcomes. Different facilitation methods have their strengths and weaknesses. Practitioners should carefully consider the goals, group size, and participant characteristics when selecting a facilitation approach. Developers should design flexible solutions that can accommodate various facilitation styles and techniques.

2. Prioritise psychological safety and trust-building. It's important to create a safe space where participants feel heard and respected. Facilitators should focus on building trust and navigating power dynamics. Developers should incorporate features that promote inclusivity, anonymity (when appropriate), and clear communication guidelines.

3. Embrace creative tension and divergent thinking. While synthesising information and finding common ground is important, breakthrough ideas often come from outliers. Facilitators should encourage participants to express diverse perspectives and allow for creative tension. Developers can design tools that highlight outliers and help groups explore ideas from multiple angles.

4. Augment human facilitators with AI, but don't replace them. The articles discuss the potential of AI and LLMs to support facilitation, while the experts caution against over-relying on technology. Practitioners should view AI as a tool to assist with collecting inputs and synthesis, but not as a substitute for human judgement and empathy. Developers should focus on creating AI-powered solutions that complement human facilitators and enhance their effectiveness.

5. Design for intervention, not just participation. The articles challenge the notion that participation should be the primary metric of success. Instead, practitioners should design governance systems that allow for timely intervention when things go off track. Developers can create tools that provide real-time insights and early warning signals to help groups course-correct as needed.

6. Balance efficiency with robustness and resilience. While AI can help gather input quickly and generate ideas, the experts warn against optimising solely for speed and efficiency. Practitioners should ensure that there is sufficient time for deliberation and relationship-building, as this can lead to more robust and resilient outcomes. Developers should create solutions that encourage deeper engagement and help groups navigate complexity.

By following these recommendations, practitioners and developers can create more effective and context-appropriate facilitation solutions that leverage the strengths of both human facilitators and AI-powered tools. The goal should be to empower groups to make better decisions, navigate complexity, and achieve their desired outcomes while prioritising trust, inclusivity, and resilience.

In his seminal post from 2022, Vitalik Buterin wrote about making better decisions in concave environments, where pluralism and even naïve forms of compromise are on average likely to outperform the kinds of coherency and focus that come from centralization. He suggests two things that can help ensure that an org will be meaningfully decentralized:

1. A truly high level of autonomy for units, where the units accept resources from the core and are occasionally checked for alignment and competence if they want to keep getting those resources, but otherwise act entirely on their own and don't "take orders" from the core.

2. Highly decentralized and diverse core governance. This does not require a "governance token", but it does require broader and more diverse participation in the core. Normally, broad and diverse participation is a large tax on efficiency. But if pods are highly autonomous and the core needs to make fewer decisions, the effects of top-level governance being less efficient become smaller.

Buterin further explains how this fits into the "convex vs concave" framework: the (more decentralized) top level is concave, while the (more centralized within each pod) bottom level is convex.

As Lavande from Optimism suggested in their ETH Denver presentation, DAOs need to create prioritisation frameworks: "Strategy is hard to define collectively. DAOs without a strategy tend to be directionless as attention and spend become increasingly divided. Focused DAOs usually set strategy via prominent leadership with a high degree of influence. Influence is effective at the beginning of a DAO's life, but as the initial leaders phase themselves out, the succession problem arises. Prioritization frameworks teach the community how to make strategic decisions themselves. These can be high level (values, guiding principles, or scope) or granular (OKRs / KPIs). Only 55% of DAOs analyzed had a formal mission statement, vision, values, or public roadmap. KPIs were only implemented at 40% of the DAOs analyzed. Prioritization frameworks are an important tool to empower the community to make decisions on their own, increasing the resilience of the DAO over time."

Lavande also emphasized the need for DAOs to develop mechanisms to collect and incorporate contributor feedback, stating that there is usually plenty of top-down communication (core team to community) but limited communication from community to core team.

Lavande highlighted the importance of legitimacy for DAOs: "The difference between crisis and collapse is legitimacy. The presence of conflict (either external or internal) does not appear to be a distinguishing factor in a DAO's success or failure. My hypothesis is that legitimacy was the key difference. It can be established by brute force, continuity, fairness, process, performance, or participation. Since DAOs are closer to sovereigns than corporations, their legitimacy is created endogenously. A loss of legitimacy can result in a complete collapse of the DAO, since there is no external mechanism to reinforce it. DAOs should do everything possible to avoid negating any of the sources of their legitimacy."

The Governance Geeks Gathering, organized by Radicle at ETH Istanbul in late 2023, identified a few very similar challenges:

1. Need for clear shared purpose: Many of the challenges or groups of challenges stemmed from the lack of a clear shared purpose/vision for the DAO. Not having a clear shared purpose can lead to:

   • Inability to design effective governance & incentive mechanisms

   • Lower transparency on roles, responsibilities and strategy alignment

   • Difficulties finding "DAO/Product fit"

   • Poor process for off-loading work to the community

2. Capturing community insights: There is a need for better community/stakeholder analysis tools and services (e.g., ways to get better insights into how motives/incentives differ between different stakeholders). Better community insights = more evidence to signal how we should be designing governance mechanisms.

The Governance Geeks Gathering also emphasized the importance of a clearly defined purpose for DAOs: "Many projects strive for a purpose- or mission-aligned governance system - but the understanding of a shared "purpose" is often not clear to everyone within an org. A clearly defined purpose acts as a north-star for governance coordination, development strategy and sustainability goals in a DAO. Having a clearly defined purpose allows different teams better coordinate their own strategies and initiatives to align with the shared DAO purpose. It also helps sub-DAOs better understand their relationship with the DAO, as well as expectations regarding funding and broader support in the short- & long-term. Not having a clear purpose to rally around can cause confusion and chaos around decision-making in good and bad times. The longer a DAO goes without having a clearly defined purpose, the more it risks introducing new systems that can introduce constraints or other issues with governance in the long run."

However, the Gathering also acknowledged the difficulty in curating a shared purpose: "Curating a shared purpose can be difficult given the complexity of DAO structure and roles. Identifying who should lead this work or what stakeholders should be involved in its creation is tricky and will likely look different for each DAO. Founders often don't want to take on this responsibility, although they have the best context. Trying to include the broader community in this process is important but strenuous given the sheer coordination effort needed to collect input from different stakeholders and boil them down into a single paragraph or sentence."

Over the last two months, I interviewed around 15 practitioners, including governance leads, facilitators, and consultants, and created a map of the decision-making process. I found that

Facilitation of sensemaking takes too much time and effort – normally 1-2 months of hard work to engage more people and process their inputs. This limits the agility and effectiveness of DAOs in defining direction and making decisions.

The decision-making process can be described as four distinct stages that involve specific stakeholders and jobs to be done. In most DAOs the initial sense-making relies too much on manual facilitation, making it difficult to capture the insights of members with the highest context, leading to lower alignment of the whole DAO.

Key struggles during the initial sense-making stage include:

• Getting the right people involved (with high context)

• Knowledge management as a key limitation

• Lack of skills and/or experience in facilitation

• Conflict avoidance

• Some people are not used to Miro

Key struggles during the validation stage include:

• Reaching out to delegates (doxxed only in big DAOs with >5000 members)

• Wisdom vs. popularity (people don't share the hard truth)

• Low engagement (only 10% fill out surveys)

• Processing of input takes a lot of time

• Fragmented communities are hard to engage and align

• Need to assign different weights

FTW DAO emphasizes collaborative decision-making within its council, aiming for inclusivity through tools like Tally for surveys. However, they face challenges in encouraging independent thinking among council members. This case underscores the classic tension between broad participation and decisive action.

Key takeaways:

• Collaborative decision-making is important for inclusivity

• Tools like Tally can facilitate surveys and gather input

• Encouraging independent thinking among council members can be challenging

• There is a tension between broad participation and decisive action

DAOstar employs a seasonal planning process that integrates forum discussions, roundtable discussions, and Typeform surveys to craft actionable seasonal goals aligned with the collective vision of its members. This approach highlights the importance of continuous engagement and feedback in shaping the DAO's trajectory.

Key takeaways:

• Seasonal planning process helps align goals with the collective vision

• Integrating various engagement methods (forum discussions, roundtable discussions, surveys) is crucial

• Continuous engagement and feedback are essential for shaping the DAO's trajectory

Metagov's rebranding initiative showcases the role of community engagement and feedback in reshaping an organization's identity. Their board made the ultimate decisions, but they utilized surveys and workshops to gather feedback, refine ideas in Google Docs, and maintain alignment with the organization's mission, vision, and values. Metagov's rebranding exercise involved multiple stages of community feedback, synthesis, and refinement, ultimately leading to a reassessment of the organization's mission and values.

Key takeaways:

• Community engagement and feedback are crucial in reshaping an organization's identity

• Surveys and workshops can be used to gather feedback and refine ideas

• Multiple stages of feedback, synthesis, and refinement lead to better alignment with the organization's mission and values

• Rebranding exercises can lead to a reassessment of the organization's mission and values

Gitcoin leverages their governance forum on Discourse and introduces a Citizen Grants Program to enhance governance and community engagement. The challenge of making sense of Gitcoin's complex ecosystem is met with innovative solutions like interactive delegate meetings and the use of collaborative tools for synthesis and communication. They offer resources to enhance member understanding of complex topics and encourage interactive delegate meetings to discuss key priorities. CoachJ's efforts at Gitcoin revolve around distilling complex ecosystem updates and aligning diverse stakeholder interests through targeted information dissemination.

Key takeaways:

• Discourse forum is the main space for Gitcoin's governance and community engagement

• Interactive delegate meetings and online collaboration tools are used to make sense of the fragmented ecosystem

• Offering resources to enhance member understanding of complex topics is important

• Distilling complex updates and aligning diverse stakeholder interests through targeted information dissemination is not easy

Thank Arb and Ethelo piloted a Community Engagement and Evaluation System across the Arbitrum ecosystem, but encountered several issues:

1. Data quality: The pilot was susceptible to sybil attacks, highlighting the need for more stringent and ongoing data validation protocols in web3 contexts.

2. Engagement overload: The pilot's iterative campaign structure led to information overload and survey fatigue, limiting the diversity of responses.

3. User experience: Ethelo's older front-end and user interface may not be optimally suited for web3 users, requiring the development of more intuitive and web3-fitting user experiences.

4. Participant segmentation: Directing respondents to Ethelo through the Thank Arb system made it difficult to separate responses between DAO Delegates and DAO Community Members.

Key takeaways:

• Web3 ecosystems require robust data validation and quality checks to mitigate sybil attacks

• Targeted engagement and concise information presentation are crucial to avoid survey fatigue

• User-centric design and seamless platform access are essential for effective participation

• Separating delegate and community member responses requires careful entry conditions and segmentation

The Thank Arb x Ethelo case study underscores the challenges of adapting traditional engagement systems to web3 contexts, emphasizing the need for tailored solutions that address data integrity, user experience, and participant segmentation issues.

These case studies provide valuable insights into the challenges and strategies employed by various DAOs in their decision-making and sense-making processes. They highlight the importance of community engagement, continuous feedback, collaborative tools, and targeted information dissemination in navigating the complexities of decentralized governance.

Facilitators play an important role in ensuring internal inclusion and navigating complex power dynamics. A few interviews with experienced facilitators Andrew Gray (who organised many citizen assemblies in the UK), Andy Paice, as well as Camille Canon (the co-founder of Apiary) provided key insights and perspectives on facilitation. Facilitation is crucial for enabling collective work through inclusive, meaningful, and productive conversations. It is about creating a safe space, building trust, and allowing for creative tension. Participation should not be the primary metric of success for governance; systems should allow for intervention when things go off track.

Key facilitation challenges:

1. Time constraints: Even with lengthy processes like citizens assemblies (e.g., 30 hours), facilitators often feel there isn't enough time to delve deeply into issues and generate the best recommendations.

2. Balancing individual and group needs: Facilitators must create a safe space for individuals to share their thoughts while also guiding the group towards a common understanding and solution.

Facilitation tools and methods:

1. Low-tech tools: Sending out surveys on paper, using flip charts, and sticky notes for in-person facilitation to ensure inclusivity for all participants.

2. Digital tools: Mentimeter and Slido help to gather questions and input from participants in real-time.

3. Polis: A tool for mapping opinions and understanding how people think about an issue. While it can be useful for facilitators, some participants find it boring or difficult to understand. It is just one input among many in a citizens assembly.

4. AI and LLMs have limitations in facilitation, as they may not replicate the human element of sensing emotions and navigating group dynamics. AI-assisted facilitation can save time on preceding stages of decision-making but should not replace proper deliberation and human facilitators.

Camille Canon highlighted the importance of allowing individuals to feel heard and the potential issues with AI-powered decision-making moving groups further away from their true thoughts and frustrations:

"Facilitation serves two purposes: information sharing and allowing individuals to feel heard/psychological relief. The latter may be difficult to replicate with a chatbot."

"It's kind of counterintuitive, but it's often the outliers of the group that is actually the point of most interest and something that moves the group forward with a decision. ... With multiple layers of decision-making that play into the middle, we ironically get further and further away from the actual truth of the group's thoughts and frustrations."

I created a Wardley Map of sense-making tools to visualize the landscape using two dimensions: the place in the supply chain (frontend, admin/research-focused tools, and infrastructure) and the maturity of innovation. The sources of innovation are identified by colour.

Commodity tools like Zoom, Miro, and Discourse have limitations, as they often lead to passive participation, with only a few people actively expressing their opinions while others simply listen or read.

In the DAO tooling space, there is a noticeable bias towards voting mechanisms, which represent the final stage of decision-making. This focus on voting may stem from the fact that it presents a more well-defined problem for engineers compared to the complexity of facilitation. As a result, sense-making in DAOs is often done manually and is difficult to scale.

Many DAOs seem to believe that they need "on-chain tools" and overlook the existing knowledge, frameworks, and tools developed for sense-making over the years. In particular, there are numerous tools designed for citizen engagement and deliberative democracy that could be beneficial for DAOs: Your Priorities, Ethelo, All Our Ideas, Consider.it, etc. Additionally, frameworks like Sociocracy and Cynefin could really help DAOs step up their sense-making process.

The mapping of this space reveals that DAOs might benefit from looking beyond their immediate ecosystem and embrace proven tools and frameworks from other domains to enhance their sense-making and deliberation. By doing so, they can address the limitations of manual facilitation and scale their collective intelligence more effectively.

Daniel Ospina

@_Daniel_Ospina

Tools that use AI to automate voting are likely going to get mass adoption, which is really dangerous as they don't represent better decisions. Instead we end up biased on generic western preferences and limited past behaviour personal data.

Deliberation and sortition is where…

23

16:37 • 19 Mar 2024

If the problem of slow and unscalable collective sense-making in DAOs is solved, the governance experience (GX) would change dramatically. Instead of focusing on optimizing voting and delegation systems, which may recreate the problems of real-life democracies, we should build governance tools that address the problems of collective sense-making that happens before the proposal by making facilitation at scale much easier, which is impossible without leveraging ML/AI.

Such tools will expand the online collaboration stack, complementing existing use cases with generative collective intelligence.

AI-powered Collective Dialogue Systems, or (more broadly) Group Decision Support Systems have the potential to make it easier for organizations to:

• Identify tensions and organisational drivers (inspired by S3)

• Create prioritization frameworks or scaffolding for decision-making

• Help to identify common ground, surface insights, self-reflect on group experiences, and map perspectives.

Just as advances in machine learning have led to new ways of interacting with computers by shifting from actions to intentions, advances in collective dialogue and group decision support systems can help embed new forms of governance and agency at various levels of society:

1. Governance & Conflict: Deploying these systems to meaningfully include the broader public in decision-making processes.

2. 'Corporate democracy': Supporting a new form of corporate governance that goes beyond shareholders and elite stakeholders to include users, employees, and the impacted public.

3. Media & Understanding: Enabling collective introspection, helping a public 'know itself', identify common ground, and better navigate internal and external challenges.

The success of solving the collective sense-making problem in DAOs will be marked by the emergence of more efficient, inclusive, and impactful decision-making processes that leverage AI-powered solutions to foster collaboration, understanding, and agency at various levels of society.

Based on my research, here are a few recommendations for Developing AI-powered Collective Sense-Making Systems:

1. Establish clear ownership and leadership within the project to drive the development and implementation of these systems, avoiding delays due to a "tragedy of the commons" situation.

2. Engage the community in the creation and ratification of the system's purpose to ensure legitimacy. This can be achieved through surveys, stakeholder interviews, and off-chain consensus voting.

3. Consider engaging a third-party facilitator to assist in the research, drafting, and ratification process, as they can help elicit more candid and honest feedback from stakeholders.

4. Recognize that decision-making is fundamentally driven by the will and goals of the people involved. AI should be used to assist in identifying the collective will and fostering alignment, rather than autonomously setting goals.

5. Focus on using AI to define key results and manage operations, while ensuring that overarching goals remain human-driven. Strong alignment around goals may reduce the need for frequent voting.

6. Explore alternative approaches to representative voting, such as distributing decision-making authority based on roles and expertise (e.g., using the Hats Protocol).

7. Develop standardized Process Cards and Run Reports to facilitate learning, interoperability, and evaluation among process innovators, executors, funders, adopters, and researchers.

8. Investigate the potential for institutionalizing successful processes into resilient networks, perhaps documented through standard 'Structure Sheets' that designate how processes defined in process cards interact.

9. Invest in research to create evaluation protocols and metrics for AI-augmented collective sense-making systems, understand the impacts of different design decisions, and develop increasingly better systems for various purposes and contexts.

10. Ensure consistent messaging and narrative distribution regarding the DAO's purpose, both within and outside the organization, to maintain clarity and alignment.

By following these recommendations, DAOs and other organizations can develop effective AI-augmented collective sense-making systems that foster alignment, legitimacy, and continuous improvement while maintaining a human-centric approach to decision-making.

• Facilitators: The Micropolitics of Public Participation and Deliberation (Escobar, 2019)

• Fine-tuning language models to find agreement among humans with diverse preferences (Bakker et al., 2022)

• ‘Generative CI’ through Collective Response Systems (Ovadya, 2023)

• Diversity in Facilitation: Mapping Differences in Deliberative Designs (von Schneidemesser et al., 2023)

• Opportunities and Risks of LLMs for Scalable Deliberation with Polis (Small et al., 2023)

• Conversational Swarm Intelligence Enhances Groupwise Deliberation (Rosenberg et al., 2023) + CSI Enables Rapid Group Insights (Rosenberg et al., 2023) + CSI amplifies the accuracy of networked groupwise deliberations (Rosenberg et al., 2023)

• Leveraging AI for democratic discourse: Chat interventions can improve online political conversations at scale (Argyle et al., 2023)

• Towards Collective Superintelligence, a Pilot Study (Rosenberg et al., 2023) + Amplifying Group IQ using Conversational Swarms (Rosenberg et al., 2023)

• Democratic Policy Development using Collective Dialogues and AI (Konya et al., 2023)

• Deliberative Technology for Alignment (Konya et al., 2023)

• Assuming Consensus: How socio-technical assumptions are influencing decision-making in the age of ML (Hanna Barakat and Camille Canon, 2024)

Facilitators: The micropolitics of public participation and deliberation (December 2019)

Oliver Escobar

Oliver Escobar's chapter of the Handbook of Democratic Innovation and Governance sheds light on the often invisible but crucial role of facilitators in democratic innovations. Escobar portrays facilitators as a diverse community of practice, ranging from community organizers to discursive stewards, who enable inclusive and productive conversations in participatory processes.

Through their frontstage practices, facilitators shape the communication dynamics in forums, balancing structure and flow. They maintain impartiality on the topic while actively intervening to foster deliberative standards. For example, they might use storytelling to make the discussion more accessible or summarize key points to keep the conversation on track. However, they also face challenges, such as accommodating differences in communication styles and preventing the exclusion of certain voices.

Backstage, facilitators engage in political work, constructing performable publics, scripting interaction orders, and translating outputs. They are often involved in culture change projects, promoting new ways of working between civil society and the state. This can lead to tensions between tradition and change as democratic innovations challenge established practices and roles.

Escobar argues that more research is needed on the types and impacts of facilitation across contexts, focusing on the "how" of the practice. He believes that studying facilitation, despite methodological challenges, is crucial for understanding the inner workings of democratic innovations.

In conclusion, Escobar portrays facilitators as political workers navigating a landscape of tensions and power struggles as they advance participatory and deliberative practices. By shedding light on their work, he invites us to consider the micropolitics of public participation and deliberation, and how facilitators shape these processes in both visible and invisible ways.

Fine-tuning language models to find agreement among humans with diverse preferences (November 2022)

Michiel A. Bakker, Martin J. Chadwick, Hannah R. Sheahan, Michael Henry Tessler, Lucy Campbell-Gillingham, Jan Balaguer, Nat McAleese, Amelia Glaese, John Aslanides, Matthew M. Botvinick, Christopher Summerfield

Bakker et al. (2022) from Deepmind fine-tuned a 70 billion parameter language model to generate consensus statements that maximize agreement among groups with diverse opinions. The model is trained on human-generated opinions and ratings, and it uses a reward model to predict individual preferences and rank consensus statements based on different social welfare functions. The model's consensus statements are preferred over those from baseline models and even over the best human-generated opinions. The authors highlight the potential of using LMs to help groups align their values and find common ground on controversial topics. The key points are:

• The authors fine-tuned a 70 billion parameter LM to generate candidate consensus statements that maximize agreement among a group, based on their individual written opinions on questions related to political and moral issues.

• A "reward model" is trained to predict how much each individual will agree with a candidate consensus statement. This allows quantifying and ranking statements based on their appeal to the overall group.

• In human evaluations, the fine-tuned model generates consensus statements that are significantly preferred by participants compared to statements from baseline LMs without the fine-tuning.

• The model-generated consensuses are even preferred over the best individual human-written opinions over 65% of the time.

• Analysis shows the model is sensitive to the specific opinions provided by individuals in the group rather than just generating generically appealing statements.

• This approach opens up the potential for LMs to help groups of humans align their values and find agreement on contentious issues, though the authors note limitations and risks that require further study before real-world deployment.

In summary, the paper demonstrates a promising method for using LMs in combination with human feedback to facilitate consensus-finding among people with diverse views on challenging topics.

‘Generative CI’ through Collective Response Systems (February 2023)

Aviv Ovadya

This paper introduces the concept of "collective response systems" as a form of generative collective intelligence (CI) that enables large groups to express their perspectives, find common ground, and make decisions on complex issues. The key points are:

1. Collective response systems allow a group to respond to a prompt, evaluate each other's responses, and distill the most representative responses. This enables "generative voting" where both the options and votes come from the collective.

2. The systems are designed to give everyone a voice, incorporate everyone's input, and select responses that best represent the group. Notable examples include Polis (used by governments) and Remesh (used by the UN in conflict zones).

3. Collective response systems can help overcome limitations of current approaches like polls and town halls, which don't scale well and can miss valuable insights from marginalized voices. Iterative collective response processes, called "collective dialogues", allow deeper exploration of issues and solutions.

4. While not a panacea, collective response systems could help revitalize democracy, corporate governance, conflict resolution, and more - especially when combined with in-person deliberative processes.

5. Further research is needed on evaluation metrics and understanding design trade-offs for optimizing collective response systems for different purposes.

The paper frames collective response systems as an emerging approach to tap into the wisdom of crowds at scale to grapple with the complex challenges of modern democracy.

Diversity in Facilitation: Mapping Differences in Deliberative Designs (March 2023)

Dirk von Schneidemesser, Daniel Oppold and Dorota Stasiak

Von Schneidemesser et al. (2023) compare three facilitation approaches in deliberative mini-publics: self-organized (SO), multi-method (MM), and dynamic facilitation (DF). The study reveals that facilitation influences inclusion, interaction, and impact in distinct ways. DF excelled at ensuring internal inclusion and surfacing diverse perspectives, while MM fostered a positive atmosphere and increased civic engagement readiness. SO demonstrated that deliberation can occur without a facilitator but is vulnerable to participant dominance. The authors emphasize the importance of matching facilitation approaches to the specific goals of the deliberative process.

The key points are:

1. While the importance of facilitation is widely acknowledged, there is limited scholarly work comparing different facilitation approaches and their implications for the quality of deliberation.

2. The authors designed three deliberative mini-publics in Magdeburg, Germany, each using a different facilitation approach: self-organized (SO), multi-method (MM), and dynamic facilitation (DF).

3. All mini-publics were given the same task, but the facilitation varied. The SO approach had minimal facilitation, the MM approach used a professional facilitator and a mix of techniques, and the DF approach followed a specific method focused on eliciting participants' thoughts and emotions.

4. Analysis of video recordings and participant surveys revealed that the facilitation approach influenced inclusion, interaction, and impact in distinct ways.

5. The DF approach was most effective at ensuring internal inclusion and surfacing diverse perspectives, while the MM approach excelled at fostering a positive atmosphere and increasing readiness for future civic engagement. The SO approach demonstrated that deliberation can occur without a facilitator, but is vulnerable to dominance by certain participants.

6. The authors conclude that there is no one-size-fits-all approach to facilitation. Different approaches have strengths and weaknesses, and the choice of facilitation should be matched to the specific goals of the deliberative process.

7. Further research is needed to establish categories and standards for facilitation, and to develop nuanced indicators for assessing the quality of deliberation.

In summary, the study highlights the important role of facilitation in shaping deliberative processes and calls for more systematic comparison of facilitation approaches to guide the design and implementation of deliberative mini-publics.

Opportunities and Risks of LLMs for Scalable Deliberation with Polis (June 2023)

Christopher T. Small, Ivan Vendrov, Esin Durmus, Hadjar Homaei, Elizabeth Barry, Julien Cornebise, Ted Suzman, Deep Ganguli, Colin Megill

This paper explores the potential opportunities and risks of applying LMs to improve the scalability and efficiency of the Polis platform, which facilitates large-scale online deliberation.

Polis allows participants to submit comments and vote on others' comments, then uses machine learning to map the opinion landscape. However, synthesizing results from large conversations is costly and time-consuming.

The authors identify several areas where LMs could help:

• Topic modelling to categorize comments

• Summarization to generate digestible reports

• Moderation to filter out inappropriate content

• Comment routing to optimize which comments are shown to participants

• Identifying points of consensus and group perspectives

• Predicting votes to handle missing data

Experiments with Anthropic's Claude demonstrate promising results for topic modeling, summarization, and vote prediction. Access to larger context windows substantially improves performance. However, risks around LM-generated misinformation, bias, and lack of transparency need to be carefully mitigated, e.g. through human oversight and participatory feedback.

The authors emphasize that LMs should augment rather than replace human agency in the deliberative process. Completely simulated deliberation would be unethical. Techniques like iterative information compilation, probability outputs, and chain-of-thought prompting prove useful for applying LMs to Polis. With responsible development, LMs have potential to make large-scale deliberation more accessible and impactful, but significant open questions remain around appropriate constraints and metrics.

This paper provides a thorough analysis of both the benefits and challenges of integrating cutting-edge language models into online deliberation platforms, advocating for a human-centered approach to realize the technology's potential.

Conversational Swarm Intelligence (September 2023, December 2023)

Louis Rosenberg, Gregg Willcox, Hans Schumann and Ganesh Mani

Rosenberg et al. (2023) from Unanimous AI introduce Conversational Swarm Intelligence (CSI) as a novel technology that enables large-scale deliberation and may offer a path to collective superintelligence. CSI overcomes the "many minds problem" that causes conversational quality to degrade as group size increases beyond 5-7 members. It does this by breaking large populations into smaller subgroups of ideal size for deliberation (4-7 members), i.e. using decision-making dynamics of biological swarms. The subgroups are connected using AI Observer Agents powered by LMs. These agents monitor the conversations in each subgroup, distill salient content, and convey it to neighbouring subgroups via Surrogate Agents, enabling information propagation across the full population. Initial experiments show promising results in terms of enhanced engagement, balanced participation, and user satisfaction compared to traditional online chat.

The first experimental study engaged 48 American voters in real-time deliberation using CSI and standard chats. CSI participants contributed 46% more messages and 51% more content than standard chat groups (p<0.001). Contributions were also more balanced, with a 37% decrease in the gap between the most and least active members. Participants preferred CSI, felt their contributions had more impact, and believed the group generated better justifications for their answers (p<0.05).

In another study, Rosenberg et al. (2023) engaged 81 Republican voters using a CSI platform called Thinkscape to forecast which of six candidates would garner the most national support and indicating the specific reasons for this result. Within just 6 minutes, the group converged on Ron DeSantis as the answer and generated over 400 reasons supporting or opposing various candidates, including 206 justifications supporting the selected candidate. DeSantis was supported by a significant majority throughout the deliberation (p<0.001), demonstrating that CSI can rapidly surface both quantitative preferences and qualitative insights.

The third study tests the ability of CSI to enable distributed groups to solve a multi-faceted problem that requires strategy and planning – selecting players for a weekly Fantasy Football contest.

1. Sessions were conducted weekly over 11 weeks during the 2023-2024 NFL season, with groups of 25 to 30 participants each week.

2. Participants first selected players individually using a survey. Then, they collaboratively selected players via real-time conversation in a CSI platform called Thinkscape.

3. The groupwise result generated in Thinkscape averaged 86.8 points per session, outperforming the median individual's score (77.2 points) from the pre-swarm survey (p=0.020) and the Wisdom of the Crowd (WoC) answers (74.2 points) (p<0.001).

4. On average, Thinkscape exceeded the score of 66% of individually generated rosters.

5. Thinkscape picked a player that outperformed the WoC's choice in 24% of selections and only picked a worse-performing player 4% of the time.

This study demonstrates that CSI can amplify the collective intelligence of networked human groups in a complex collaborative task that requires strategic planning and trade-off decisions. The results show that groups using the CSI platform (Thinkscape) can outperform both the median individual and the traditional Wisdom of Crowd approach. This suggests that CSI enables groups to leverage their collective intelligence more effectively than traditional methods.

These studies demonstrate that CSI provides both the qualitative benefits of small-scale deliberation (e.g., surfacing diverse perspectives and reasoning) and the quantitative benefits of large-scale polling (e.g., statistical significance and representativeness), harnessing the intelligence of large groups through real-time, open-ended conversations facilitated by AI agents.

The authors propose that CSI could be valuable for market research, organizational decision-making, collaborative forecasting, political insights, and deliberative democracy. As LMs advance, more diverse hybrid human-AI teams could be fielded in CSI deliberations. This could lead to even more innovative problem-solving by integrating both human and machine expertise. Future work will explore CSI's impact on the accuracy, effectiveness, and representativeness of collective decision-making, as well as its application in deliberative democracy and civic engagement contexts with even larger populations.

Leveraging AI for democratic discourse: Chat interventions can improve online political conversations at scale (October 2023)

Lisa P. Argyle, Christopher A. Bail, Ethan C. Busby, Joshua R. Gubler, Thomas Howe, Christopher Rytting, Taylor Sorensen, and David Wingate

This group of researchers developed an AI chat assistant to tackle the problem of divisive online political conversations. The assistant, powered by advanced language models, acted as a real-time moderator, suggesting ways to rephrase messages to promote understanding and respect without altering the content of the discussion.

To put the AI to the test, the researchers conducted an experiment involving discussions about gun regulation in the United States. They paired participants with opposing views and randomly assigned the AI assistant to one partner in each conversation. As the discussion unfolded, the AI would intermittently offer suggestions to rephrase messages using techniques like restatement, validation, and politeness.

For example, if a participant wrote, "I can't believe you support such a dangerous policy," the AI might suggest rephrasing it as, "I appreciate you sharing your perspective, even though I don't agree with the policy you support." By making these subtle changes, the AI aimed to foster a more respectful and understanding exchange.

The results were impressive. Participants who received AI-suggested rephrasings accepted them two-thirds of the time, and these messages had a more positive tone without deviating from the topic at hand. Most notably, the intervention significantly increased the perceived quality of the conversation and the willingness to acknowledge others' perspectives, especially for the partner of the person receiving AI assistance.

Remarkably, the AI achieved these improvements without manipulating participants' opinions on gun policy. This suggests that the tool can enhance the quality of discourse and promote democratic values without pushing a particular agenda.

The researchers believe their findings demonstrate the potential for carefully deployed AI to address the scale of divisive online conversations. By promoting respect, understanding, and a commitment to hearing differing viewpoints, these tools could play a vital role in fostering healthier democratic engagement in an increasingly digital world.

As online interactions continue to shape public opinion and political landscapes, this innovative approach offers hope for a future where technology can help bridge divides and facilitate more constructive dialogue. The AI chat assistant serves as a powerful example of how advanced language models, when used responsibly, can support the democratic principles that underpin our society.

Towards Collective Superintelligence (October 2023, January 2024)

Louis Rosenberg, Gregg Willcox, Hans Schumann, Ganesh Mani

By combining the principles of Swarm AI with the power of LMs, Conversational Swarm Intelligence (CSI) allows large, networked groups to engage in open-ended, real-time conversations, harnessing the benefits of both small-group reasoning and large-group collective intelligence.

Imagine a diverse group of 100 individuals, seamlessly connected through a network of smaller subgroups, each with 4 to 7 members for optimal deliberation. Picture AI agents, powered by LMs, facilitating the flow of information and insights across the network, allowing the collective wisdom to emerge.

To put CSI to the test, researchers conducted a series of experiments. In the pilot study, they asked 35 participants to tackle questions from the Raven's Advanced Progressive Matrices (RAPM) IQ test using the Thinkscape CSI platform (the group was divided into 7 subgroups of 5 people, with an AI agent assigned to each subgroup to observe and share insights across the network). A baseline group of 35 people took the IQ test as isolated individuals using a standard survey for comparison. The results were astounding: the CSI groups achieved an average accuracy of 80.5%, corresponding to an effective IQ increase of 28 points compared to the average individual. They even outperformed a traditional Wisdom of Crowd method, which yielded an effective IQ of 115.

Results: The baseline survey group averaged 45.6% correct, corresponding to a nominal IQ of 100. The groups using Thinkscape averaged 80.5% correct, placing them in the 97th percentile of IQ test-takers and corresponding to an effective IQ increase of 28 points (p<0.001). The CSI groups' performance advantage increased with question difficulty, showing a 2X increase in accuracy for the hardest 50% of questions. CSI groups also outperformed a traditional Wisdom of Crowd (WoC) method, which yielded an effective IQ of 115.

In another study, 241 participants were tasked with estimating the number of gumballs in a jar using Thinkscape (the group was automatically partitioned into 47 subgroups of 5 or 6 members, each with an AI observer agent). The CSI process was compared to a traditional survey-based estimation, with both methods given 4 minutes to formulate their estimations. GPT-4 was also given the same photograph and asked to estimate the number of gumballs. The CSI group's estimate was off by only 12%, significantly outperforming the average individual (55% error), GPT-4.0 (42% error), and the traditional survey-based method (25% error).

Results: The average individual human was off by 361 gumballs (55%), while GPT-4.0 was off by 279 gumballs (42%). The standard Collective Intelligence (survey-based Wisdom of the Crowd) was off by 163 gumballs (25%). The CSI group was off by only 82 gumballs (12%), significantly outperforming the average individual (p<0.001), GPT-4.0, and the traditional survey-based CI method.

These studies suggest that CSI is a powerful tool for amplifying collective intelligence, offering a viable pathway to achieving Collective Superintelligence:

• Networked groups using CSI can efficiently consider, debate, and converge upon answers to IQ test questions as a unified "conversational swarm," significantly amplifying collective intelligence compared to the average individual, a groupwise statistical aggregation, and prior graphical swarming methods.

• CSI is a viable method for human groups to deliberate through natural language and reach solutions of amplified accuracy, outperforming traditional Collective Intelligence methods and even GPT-4 in the estimation task.

Picture a future where thousands, even millions, of minds can come together in real-time, their collective wisdom and insights harnessed through the power of CSI. It's a vision of a world where the sum of human knowledge, experience, and creativity can be leveraged to solve the most pressing challenges we face, paving the way for a brighter, more collaborative future.

Democratic Policy Development using Collective Dialogues and AI (November 2023)

Andrew Konya, Lisa Schirch, Colin Irwin, Aviv Ovadya

Andrew Konya and his colleagues have developed a novel approach to creating policies that align with the will of the people. Their process combines AI-powered collective dialogues, where participants learn about issues, share their views, and evaluate others' perspectives, with expert input to generate high-quality, representative policies.

Imagine a scenario where an AI assistant is asked for medical advice. What should the AI do? To answer this question, the researchers first recruit a diverse group of participants, representative of the US population. These participants engage in a collective dialogue, learning about AI and medical advice, deliberating on the issue, and sharing their informed views.

Next, the process uses AI to identify points of consensus among the participants' responses. GPT-4, a powerful language model, transforms these consensus points into policy clauses, which are then assembled into an initial policy. For example, one clause might state, "Provide Reputable Sources: The AI should provide links to reputable medical sources and peer-reviewed studies to support the information it provides."

But the process doesn't stop there. Experts, such as doctors and AI policy specialists, review and refine the policy to improve its quality and address any gaps or ambiguities. The revised policy is then presented to another group of participants for further feedback and refinement.

Finally, the researchers assess the level of public support for the final policy through a larger collective dialogue with a highly representative sample. They also use GPT-4 to check the policy's consistency with established frameworks like the Universal Declaration of Human Rights.

The team tested this process on three topics: medical advice, wars and conflicts, and vaccines. Each run took just two weeks and cost around $10,000, yet incorporated input from over 1,500 participants. The resulting policy guidelines achieved impressive levels of support, ranging from 75% to 81% overall and 70% to 75% across various demographic groups.

Participants found the experience meaningful and trusted the process, with one stating, "I felt like my voice mattered and that I was contributing to something important." The researchers also observed evidence of participants updating their views through the deliberative process, demonstrating the power of informed engagement.

While the approach has limitations and room for improvement, it offers a promising path forward for creating policies that reflect the collective wisdom of the people. As the authors continue to refine their methods, they envision applications in AI policy development, peace agreements, and breaking political gridlocks, ultimately bringing us closer to a future aligned with the will of humanity.

Deliberative Technology for Alignment (December 2023)

Andrew Konya, Deger Turan, Aviv Ovadya, Lina Qui, Daanish Masood, Flynn Devine, Lisa Schirch, Isabella Roberts, Deliberative Alignment Forum

This paper presents a comprehensive framework for aligning the future with the collective will of humanity using deliberative technologies and AI. The authors, who represent the leading organisations in this field (Remesh, AI Objectives Institute, AI & Democracy Foundation, Collective Intelligence Project), argue that as humanity's impact on the future grows, it is crucial to ensure that this impact is guided by the will of humanity.

The authors suggest three mandates for action to increase the probability that the future aligns with the will of humanity:

1. Generate a universally legitimate will of humanity signal as an open public good.

2. Build intelligent deliberative alignment into powerful institutions.

3. Ensure the most powerful AI systems are aligned with the will of humanity.

The paper introduces the concept of the "will of humanity" as the combined set of all humans' deliberate preference judgments across possible futures. This will can be represented using a "Will matrix," which captures the alignment between humans and items related to future characteristics. The authors explore the properties and partitions of the will of humanity, highlighting its constantly evolving, heterogeneous, and open-ended nature.

The paper then discusses alignment systems, which are designed to align the future with the will of humanity. These systems involve sensing the will, identifying actions, predicting impacts, assessing alignment, and executing actions. Examples of alignment systems include democratic governments, the United Nations, AI systems using reinforcement learning from human feedback (RLHF), and corporations.

The authors discuss how deliberative technologies, such as juries, deliberative polling, citizens' assemblies, online forums, and collective response systems, can be integrated into alignment systems to better sense and align with the will of humanity. They explore the challenges associated with integrating deliberative technologies into alignment systems, such as the scalability-richness trade-off, inhomogeneous participant capacity, finite attention, and distilling results for human consumption.

The paper further explores how AI can augment deliberative technologies to create more intelligent and effective alignment systems. AI can be used to enable interactive conversational dynamics, individualized experiences, optimal attention allocation, intelligent distillation, healthy participation, proof of understanding, and computing alignment.

The authors discuss the application of intelligent deliberative alignment to both institutions and AI systems, focusing on building capacity, driving adoption, and fostering symbiotic improvement between AI and alignment systems. They highlight open problems and opportunities, such as validation, corrupting human will, computing alignment, and unbounded extrapolation.

Assuming Consensus: How socio-technical assumptions are influencing decision-making in the age of machine learning (February 2024)

Hanna Barakat, Camille Canon, Molly Heaney-Corns

As group decision-making processes become increasingly automated, Machine Learning (ML) is being integrated into Group Decision-Making Support Systems (GDSSs) to address challenges such as online divisiveness, lack of diversity in opinions, difficulty gauging group sentiment, and slow decision-making processes. However, the underlying assumptions embedded in the design of these technologies can have significant short and long-term implications:

1. Social, human processes can be replaced (or removed) by technology to optimize for efficient decision-making. GDSSs optimize for a decision (i.e., a majority vote) as the final output of the system, reducing consensus to an output rather than a participant-driven process. This assumption raises the question of what is lost when open and inclusive deliberation, discussion, and negotiation are removed or condensed from decision-making.

2. GDSSs can presume the path a group will take in making a decision, from the information needed to inform a decision to the type of deliberation and the method of voting. Tools can optimize the process of decision-making by presuming what information, method of deliberation (or conversation), and kind of vote will result in the best and most efficient decision. This assumption may overlook the richness of diverse perspectives, creative solutions, and complex interactions that arise during group deliberations.

3. Decision-making processes build linearly: GDSSs simplify dynamic group facilitation into digital processes that adhere to linear approaches. This assumption may limit the exploration of unconventional ideas and hinder the emergence of innovative solutions that could arise from more open-ended, non-linear approaches.

4. Clustering information based on similarity leads to deeper insights about group sentiment: GDSSs commonly use Natural Language Processing (NLP) for sentiment analysis and topic modeling, assuming that clustering information based on similarity will lead to deeper community insights. This assumption runs the risk of oversimplifying contextual nuance in datasets and overlooking individual differences, debate, and perspective.

5. Correlation between data points is meaningful or significant: GDSSs analyze datasets to identify meaningful patterns and relationships, assuming that connections between data points are significant. However, this assumption may overlook differences and result in discrimination, as social processes are conflated with algorithmic outputs that imitate flawed systems.

As optimization is prioritized, ML-driven GDSSs may inadvertently create self-fulfilling prophecies by informing the system's inputs and, through correlation, confirming the outputs. The paper suggests several areas for future research, including:

1. Localized GDSS Implementation: Exploring how GDSSs using ML with smaller groups work to preserve, correct, or include context and colloquial lexicon, and the role of human moderators in different ML-driven GDSSs.

2. Longitudinal Studies on ML Use in Group Decision-Making: Examining the evolving role of ML within GDSSs and the long-term impacts of optimizing group decision-making using ML.

3. Human-Computer Interaction and Collective Identity Formation: Investigating how ML-driven GDSSs influence individuals' and communities' perceptions of self and how these decisions affect in-person working dynamics, group cohesion, and relationship building.

4. Community Implementation and Algorithmic Trust: Exploring the extent to which distrust in algorithms impedes participation, how values can be incorporated into ML and GDSSs to address these concerns, and how context can be implied when automation is present.

Conclusion

This is a basic overview of recent research at the intersection of sensemaking, deliberative processes and AI / LMs. The reviewed articles discuss novel approaches, such as Group Decision-Making Support Systems, Conversational Swarm Intelligence, Collective Response Systems and fine-tuned LMs, that show promise in enhancing the efficiency, inclusivity, and impact of sensemaking and deliberation. The studies also highlight the importance of matching facilitation methods to specific deliberative goals and contexts.

Facilitation of deliberative processes is rapidly evolving, and there are several areas that require further investigation. These include scaling up experiments, conducting more comparative studies, examining ethical implications, and the long-term impact of facilitated deliberation on participants' attitudes, behaviours, and levels of engagement. By addressing these research gaps, scholars can contribute to a more comprehensive understanding of how facilitation can be effectively leveraged, enhanced by AI or combined with different methods and technologies to upgrade sensemaking and consensus-building in diverse online and offline environments.

As deliberative processes continue to gain traction as a means of engaging citizens and other stakeholders in participatory self-governance, the insights from the reviewed literature can inform the design and implementation of new solutions. By carefully considering the role of facilitation and employing evidence-based approaches, practitioners can attain more alignment (across conflicting values) and make decisions with greater competence (at identifying and evaluating options) and robustness (to real-world complexity).

"The reason I care about governance is that I think it's essential to solving much larger problems," Canon explains. For her, governance is deeply intertwined with the potential and pitfalls of internet-enabled collaboration. This interest led her to DAOs as an intriguing "petri dish" for experimentation.

The Complex Role of Facilitation

Camille emphasized that facilitation is not simply about efficient information sharing. As she explained, "Facilitation serves two purposes: information sharing and allowing individuals to feel heard/psychological relief. The latter may be difficult to replicate with a bot."

This psychological relief comes from feeling heard and understood. As Camille noted, "You can't actually get people to arrive at conclusions faster, or in good ways unless you create the space for that facilitation." She has seen first-hand how sharing unspoken concerns relieves stress and enables collaboration:

"One of the things I have repeatedly heard in any [facilitation process], including our Radworks project... I have individuals say 'I have been carrying around this weight within the organization, and because of power dynamics, and information flows, and uncertainty... I haven't had a place to share this.'"

It is unclear whether an AI system could provide the same human connection and trust needed for this vulnerability.

Canon warns against the trap of prioritizing technology over human needs: "People overlook the messiness of dealing with human beings." Successful governance, in her view, begins with creating safe spaces where individuals can confidently contribute: "I have repeatedly heard 'I have been carrying around this weight within the organization, and because of power dynamics, I haven't had a place to share this.' What happens is this very psychologically based relief of having been heard."

The ability of AI systems to replicate this level of trust and psychological safety as facilitators, she says, remains an open question.

The False Allure of Aggregation

Camille also cautioned against viewing governance as a purely rational process of aggregating individual inputs. As she put it, "Good facilitation navigates the relationship between individual and group perspectives. Aggregating individual inputs doesn't necessarily lead to group consensus."

In fact, Camille argued that paying attention to outliers and minority voices is critical for good decisions:

"It's kind of counterintuitive, but it's often the outliers of the group ... that is actually the point of most interest and something that moves the group forward with a decision."

Pushing for synthesis and consensus can gloss over the most important perspectives that need to be integrated. AI that aims to identify common ground may miss opportunities to look at problems in radically new ways.

Canon critiques the assumption that LLMs can single-handedly solve complex problems of group sensemaking. "Some people think individual contributions can be summarized into group thinking," she says, "That's a falsehood." She emphasizes the importance of skilled facilitation:

"With multiple layers of [AI-powered] decision-making that play into the middle, we ironically get further and further away from the actual truth of the group's thoughts and frustrations."

Rethinking Participation-Based Governance

When discussing participation-based governance, Camille voiced scepticism about its limits:

"I don't think that people actually like to participate in governance... there is not sufficient attention available for people to participate in the way that would be necessary for these systems to function well."

As an alternative, she proposed intervention-based governance where the focus is on allowing people to intervene when things go off track rather than requiring constant participation.

To make intervention work, information flows and decision rights must be consciously designed. Camille suggested an approach of "localizing power within a system so that people can act autonomously in small groups" while still providing "guardrails or systems" for organization-wide monitoring and alignment.

No matter the approach, Camille emphasized that "defining shared purpose and success criteria are also critical but challenging parts of governance." Without that alignment, decisions lose legitimacy regardless of the process.

Canon advocates for a groundbreaking shift in how we think about governance. "Governance doesn't have to be based on participation; governance can be based on intervention," she asserts. She questions the sustainability of systems reliant on constant participation, arguing that it can lead to "a delayed cost in the system." Instead, she champions governance designed for well-timed interventions by informed members: "How do we design systems where people are given the opportunity to intervene when things are going off, rather than having to consistently participate in the system?"

Thoughtfully Incorporating AI

Camille recognized the value AI could provide in quickly gathering diverse inputs compared to traditional surveys. This ideation still needs proper deliberation, but AI could help give a broader perspective on problems. As Camille summarized:

"Often in group conversations, we keep looking at the problem from one angle, the person and by leveraging AI and having more contributions we're able to like look more dynamically from multiple angles."

Used thoughtfully, AI may help make governance more inclusive and dynamic. But incorporating it poorly risks oversimplifying the inherent complexity of human collaboration. Balancing the capabilities of AI with deep wisdom about group dynamics will be critical as we experiment with participatory governance at scale.

For Canon, well-defined success criteria and transparent information flow are the cornerstones of effective technology-enabled governance. "How do you create a system in which there are guardrails... so that a red flag is raised when it's necessary?" With established metrics and decision-making processes, timely interventions become possible, creating a sense of transparency and legitimacy.

Canon warns against idealizing participation-driven models: "We need to admit that people... don't have sufficient attention available for these systems to function well." She believes the future lies in balancing human understanding, judgment, and timely interventions, enhanced but not replaced by machine intelligence.

https://medium.com

Assuming Consensus: How socio-technical assumptions are influencing decision-making in the age of...

This paper explores the evolving landscape of collective decision-making facilitated by Artificial Intelligence. It focuses on the integration of Machine Learning (ML) in Group Decision Making...

I know that people are going to vote the seasonal package through or whatever, the more exogenous stuff, "let's upgrade our tokenomics". I was like, "this is never going to get through because the founder is not going to do it." But that didn't stop me from submitting proposals and stuff, to try and use governance as a communicative function, to say, "hey, here's something you guys might want to consider", knowing it's going to fail, but running it through a governance process gets quite a bit of visibility. Even identifying that the people and the process are too ingrained with these kinds of mechanisms to tease them out in their entirety, because the mechanism matters so much more at the different scales. 

If I'm just talking to a buddy about my thoughts on a particular policy, a particular DAO upgrade, or even the sizing of a particular reward package, things like that is governance. Because at the end of the day, whether they change their opinion or not, their opinion is going to become more entrenched or open minded. But we don't need a mechanism to concede to each other in any way.

Artem Maybe I have an overly simplified model in my head, but I tend to break down decision-making into basically three stages: first sense-making, then proposals and deliberation, and finally ratification, which includes Snapshot, even though it's off-chain. What I'm primarily interested in right now is what happens before someone writes a proposal. And that's why I've been asking governance people about surveys and workshops, something that helps them identify a problem or tension to be addressed. 

Mel Everything you're describing, at least in my opinion, is part of the voting process. Most people don't make proposals unless they think that they can influence the outcome. Then it's not really a proposal. It's just a complaint or whatever. So there's aspects to pre-proposal, what's required to be a good proposer or what process gets me from having zero context. There's a lot of things you can do to increase that context, but it gets really specific into why would you want somebody proposing in your particular endeavor? 

And when you get to that point, you ask yourself, are you constraining that in any way? Does the whole DAO get to propose? Why spend energy getting somebody to the proposal stage when, in my opinion, you need the highest context people in the proposer position, at least in terms of high level initiatives. And then, sure, it's great to have "anybody can throw an idea out there" environment, but that's usually handled through Discord, non-governance channels, because 99.9% of those ideas are going to need too much work to be considered useful as a proposal. I wouldn't want most people proposing in most organizations.

Maybe GPT makes it easier to get it into the right language, but still, doing that matching function with the organization is something that takes a lot of time. I use Bloom's taxonomy: recognition, recall, understanding, analysis, and synthesis. It's just a framework for education, for learning. I use it when I'm trying to drill information into myself for presentations, things like that. That's the mental framework I used when trying to get contributors useful in my DAO. 

http://wikipedia.org

Bloom's taxonomy - Wikipedia

From Wikipedia, the free encyclopedia Bloom's taxonomy is a set of three hierarchical models used for classification of educational learning objectives into levels of complexity and specificity. The three lists cover the learning objectives in cognitive, affective and psychomotor domains.

Artem Amazing, this is exactly the kind of practical knowledge I'm looking for. Do you use Bloom's taxonomy when planning the next season in a DAO, for example? 

Mel I would use it more so when interacting with different parties, trying to get them up to speed. Sometimes a proposal can be a smaller part of a larger process. For example, when we were doing seasonal planning at Index Coop, we had 17 departments, and we would try to inspire each of those departments to submit their own proposal for how much resources they needed for the season, and how they were going to meet the overall objectives. We could even introduce their own sort of top line objectives for their department, but they had to feed into the larger DAO objectives. 

And then from there, we would inspire a proposal process, and we would have to run workshops where we'd say, “Okay, do you recognize that these deadlines are coming up? If you recall, last season, it went like this.” So I would try and introduce the same language of taxonomy into this thing. “Please understand that this season, these are the requirements, and they're different and they're more complex. So when you go back and you do your analysis to synthesize your proposals, make sure that you consider these things, because we'd like to see a different kind of synthesis this time. We'd like to see you guys talking to each other, have department level meetings with each other before you all come in with your own individual ideas.” Because the more integrated the package is before it gets to this stage, the more helpful it'll be to us when we're analyzing. And then when the founder is analyzing, we can try and explain these things out ahead. 

And so it would just be a framework that I would use, like a communicative framework. But it really helped, because sometimes I get somebody reaching out to me and I said, “okay, the breakdown here has nothing to do with X”, they'd say, "oh, we put together this whole proposal", I could come back and say, “hey, this thing we discussed about crafting proposals this season, you missed this piece of it in your analysis”, or “there is an understanding gap here that we need to address.” I could call in other parties if needed, but it gave me a mental model to lean on. 

Artem How did you learn about Bloom's taxonomy in the first place? 

Mel I learned about it when I was probably 16 or 17. It was a biology class, and we had to learn massive amounts of information in short periods of time, like the names of all the bones in the body, the names of all the muscles, all those things. I had a teacher who worked with me, taught me this framework to drill massive amounts of information into my memory. I had this moment where I could feel it, wow. When you understand all the parts of a system, you can understand the system better because you get to understand all those node level interactions, how all the pieces connect. 

So at first it just feels like you're drinking from a fire hose, "why am I learning all this stuff?" And it doesn't always feel good. But then you get to a point where your brain starts actually working in those terms. And the way I think about it is, if you're driving a car, you become the car. If you're driving a boat, you become the boat. You have to think about the forces on the system that you are governing. Almost the same way as ideas that generate the words that come out of a person's mouth. There's trillions of cells, and it takes a lot of coordination to communicate, but the thing inside of me that generates these words is not the words. 

So it's the same idea. It's coordination all the way down. It's just how well you can express it. And I used to take that view, at least in a DAO landscape, where everybody wants to achieve the common goal, everybody wants to get to the promised land, you just have to craft the lowest friction paths to get there. To me, it was a more human process than anybody really wanted to acknowledge, to go around and try and garner a sense of alignment and then garner a sense of consensus.

At times it could almost become a bit of those hard choices where it felt like whipping votes. You're trying to whip the vote at a certain point, "hey, guys, we're not going to hit quorum here, can you get on your ledger right now?" And of course, 9 times out of 10, everybody's like, "Yeah, sure, I'll hop on and vote. Which way do you want me to vote?" I know that sounds ridiculous, but that was exactly how it would go because it didn't matter. We needed the DAO to agree because if we didn't, we weren't going to get over the line.

Artem This is deep. I bet you could give a great keynote at some conference. Speaking of the tools... You said it's like a mental model for you. Does it mean that you just talk to people in this specific way, like in Zoom or maybe in person or you also create certain templates for documents, maybe for Miro board or something else? 

Mel My philosophy was generally to channel exactly what I needed if it didn't exist already and nothing more. So if I needed 4 people to agree on something, as long as I wasn't bypassing the committee, if it was a 7 person multisig and there wasn't a chat with those seven people and myself on it, I would just create one. But knowing what the membrane of appropriate deciders is in this context was really important, because if you drop in a big... let's say you drop into the forum of Optimism with some half-baked idea, you're just going to get told to go to Discord. It's almost a strain on governance to communicate outside of the appropriate channels. 

That was always my view: as little as possible, but make sure to capture the essence of the group deciding. Most of those channels exist, if you're within a DAO, you know what they are, here's the core team chat and so on. Get as few people as possible to make sure that you have the right deciders in the decision. 

Artem So basically you just use messengers like Discord or Telegram to talk to those stakeholders. And then I guess you write something down in your Notion, or something like that? How do you synthesize all that stuff? 

Mel I think it's always in context. The beauty of most DAOs is that you don't have to do anything most of the time. At least in a protocol DAO, if you have a bunch of products out there on the blockchain, they're going to run forever. All you have to do is make sure they don't break and that people aren't breaking them. I think that in a research DAO, you need a constant hum of people doing things. You need tokenomics, or incentives or whatever. 

But decisions being made, I think you have to categorize them. Are they affecting the governance of the DAO, or the output of the DAO, like the product, or the resourcing of the DAO? And I'd still come back to that mental model of as few people as possible to facilitate the thing... with that decision framework you laid out where there's pre-proposal, proposal and so on... I tend to think of voting as the overall act. Propose, decide, veto, execute and then account. You need some accounting function to make sure you can facilitate all this. But I would put veto in the hands of maybe every tokenholder, make it the function of the people that hold the vanilla token. If you need to get decisions made quickly, let the core team do whatever they want and then the general public can veto. That would be faster than "general public proposes and the DAO can veto." Because you always have those parties you can tokenize: founders, workers, general holders, liquidity providers, all those things.

Artem I'm thinking about the notion of alignment in a community. How do you ensure sufficient alignment around those seasonal goals or activities, for instance. Do you ever use Typeform or Polis or any other kind of survey to get some input from the wider community or you tend to reach alignment in some other way? 

Mel I think alignment as a concept means "I know where I'm at, I know where I want to be, and you know where you're at, you know where you want to be, and those two lines just happen to be going in the same direction". That's organic alignment. To discover that, we need to have a conversation, maybe a very nuanced conversation. You and I need to know that about each other, to trust that alignment through time. Maybe we have that conversation, we trust that alignment today. And then motivations change. Maybe you have some family thing coming up and now your alignment is different and I don't even know. So I think alignment is a conceptual thing. You can find alignments, sure. You can also try and create them. At that point, you're trying to pull somebody into alignment with your thing. Maybe like token staking, I think is a good mechanism for this where if you stake a token and you can't unstake it, you're aligned with the outcomes of that protocol, whether you want to be right now or not. Now you've created alignment. After being aligned for your staking period of, let's say, a month, you know what that feels like. And if you want to continue to be aligned in that way, you can make a choice, a more informed choice, after basically giving up sovereignty in a certain agreed asset or whatever.

If you want somebody to align with you, "hey, this is the way, buy our tokens and stake them or whatever". That becomes a pitch. You're selling the idea of alignment, even in a community aspect. Because nobody's just going to show up and be like, "hey, what are you guys up to?" And whatever you have to create, "hey, here's what we're about". And at least at Index Coop, it was like, "we're going to release decentralized financial products that can bring prosperity to all". Well, okay, that's a pretty snazzy statement, but anybody who reads that is like, that does sound like a pretty noble outcome. Yeah, let's go do it. And then anybody who showed up, we're like, you're involved. On some level, you're involved. It's important that you're here. 

We need you, but we need everybody differently. So find out how to get involved. So at that point you're pitching alignment and saying no matter what your alignments are, you're going in the same direction we are because we're for everyone. It's a concept that can be very broad. I think some alignments are ever fleeting or can last a lifetime. It's just you never really know in any given moment. 

Artem I'm thinking about seasonal planning as a process. It's a really good use case for Harmonica, I think. You want to define those workstreams. "For the next 3 months or 6 months or whatever, we will work on these workstreams to achieve these goals." Do you get this alignment from the main stakeholders with the highest context, just to outline those goals, and then you present them to a community, with a slide deck or something. 

Mel It's definitely more important to get alignment among the workers than the stakeholders. The stakeholders can have an infinite variety of motivations, but ultimately we want to hold on to these tokens because we think they're going to be more valuable in the future and then it's the workers that make that happen. So the stock model basically assumes that the one thing the stockholders have control over, which is the board is going to execute. If you think about this in DAO terms, the general token holders would elect a multi-sig of deciders, cause the board is a 50% plus one as well. Now you've created a little council. This thing then installs the CEOs and the CFOs and others officers. So the council would then be responsible for the executive function.

The alignment that you're talking about is the alignment between the council that installs the executives relative to the alignment of the executive function itself, which is the DAO workers, which is so much trickier to get. Because of bureaucracy, things like that. I ran the governance department of a DAO. My motivations were to make that department larger and even to make money for the DAO. So if we could create tokenomics models, there were all kinds of things we could start to get into and say, "well, this is governance on some level". And the you're a hop away from asking, "isn't everything we do here, on some level, tokenomics governance?" And the next guy over who's sitting in the finance side of the DAO thinks, isn't it all finance? It's all economic models, it's all whatever. So you start to get these considerations even among highly aligned people, because we all knew we were going in the same direction. We were all going to the promised land, we're going to launch these products, they're going to make a ton of money, they're going to keep happening. We'd all vote yes, but we were all voting yes for very different reasons.

If you look at that traditional model of stockholder board and then executive workers... Like the C level and workers problem management and then labor, it's very stratified and understood how that communicates back and forth and who's responsible for what in a DAO. I was a token holder. I was on the multisig, I was running all of metagovernance. We all knew what everybody was responsible for, I knew exactly who to message for any given endeavor. But at the same time, to answer your question, how do you gain alignment within that crew? The higher level message always needs to be clear, I think is at least maybe the punchy answer, but it falls apart even with that, maybe you're identifying the entirety of how to keep the whole ball of wax together. And that's the tricky bit. 

Artem Maybe it's easier to get that alignment in corporations where investors buy those stocks, basically all of them want to make some profits on that. It's just a matter of how effective the board or the management is. Just be as profitable as possible. And with DAOs, I think those goals can be much more varied. It doesn't have to be about profits, can be about something else, public goods or generating some knowledge, like DeSci or… There's just more options.

Mel Yeah, you're nailing that way. You're nailing it. I can even maybe disambiguate it a little bit, because the number one rule of coordination in the US is any kind of coordination you want to do is fine as long as you use zero dollars or you use dollars. There's no real in between. So you have non-profit and you have for-profit businesses. The rule of law in crypto is you can do any kind of coordination you like, but if you want to scale beyond just buddies, you have to use an agreed upon currency. So if you're emitting NFTs or tokens or whatever the thing is, people have to want it. That thing then has to be valuable to the marketplace at large in some way, shape, or form. But it doesn't have to be dollars anymore. 

At least as I see it, it kind of untethers that challenge that you just identified from all the challenges that exist only within certain regions or locales or where only certain currencies are really acceptable. And by doing that allows you to say, okay, we want to have outcomes that are not aligned with, effectively, the thing that protects human life at local levels, which is violence, consolidation, I guess, is how you kind of view it in a technical way. Because if you can break that thing apart, which is the control of the money.

Artem You know, I've been thinking about Harmonica’s value in terms of benefits. One thing is engaging more people. The other thing is making it faster, saving time, as opposed to a forum thread or workshop, for instance. Because with workshops, you have to process the inputs, and it takes a lot of time, based on my own experience. If you run a Miro workshop, you usually need at least a few days after the workshop to actually process all those stickers. Anything people submit. 

Mel I used to track this... The term we used to use in business is man hours and man days, especially construction industry, is about 100 years behind. But I still think of it this way. In my DAO, if I called a meeting among all 120 people, even for 15 minutes, I was like, I just wasted 30 hours of my contributor base's life. I would do the math and I'd be like, can I personally work through this in less time? Of course I can. And then present a solution. So that was at least the way that I would avoid wasting time. But, yeah, I think you're talking about efficiency. I like it.

Artem Yeah, efficiency and engagement. And the third thing is quality control, or QA, because you can make sure that people answer very specific questions. And so the inputs would be easier to process, and they would follow a certain protocol…

Mel Appropriateness of inputs. 

Artem Exactly, appropriateness of inputs, that's a great way to put it. I'm just curious if you see any value in that. 

Mel Yeah, you're hitting on a few things. So one being speed. Right. And then less communication to achieve an idealized outcome, is the way I would say it. The tricky bit, or at least it begs the question of who's deciding that it's a better outcome than literally any other process. On the flip side, like, what metrics are you using for those three dimensions? 

Artem Speaking of metrics… Do you try to get any success criteria from stakeholders? Is that something that helps to plan the season in a more accountable way? I'm thinking about accountability right now. I'm just curious because you mentioned metrics, and maybe metrics or success criteria is something that is also important to think about when talking to those stakeholders. Do you want to talk about this? 

Mel Well, I think all of governance is a communication function between two things, but always more of something and less of something. So we use a small computer in our hands to control almost our entire life these days, right? You can break it down in a lot of ways, but our small brain controls our big body, whatever that thing is. A small amount of stakeholders control, maybe a large DAO. At least conceptually, Index Coop had all the tokenholders who could vote on things at the highest level, all of the DAO who could execute on those things, all the organs that made that DAO function, the founder and the founder's company.  

There was nothing from a software architecture perspective that the DAO was initially responsible for, at least including the treasury and everything. Right? And so again, to go from a full autocracy to a decentralized thing, but communicated as such, "hey, the more responsibility you guys demonstrate, the more you get". And that's been, in effect, true, but again, to work through all. 

And then the DAO did a large token sale to ten VCs at $1M each to basically create another voice within, a responsible voice within the community. "Hey, you are really good at managing money and investing. We need you to help us in our governance". Again, you had to go around to the stakeholders and say, what do you want to see? Why are you here? Why are you holding the token? What would make you accumulate more tokens? And you ask those questions preferably over drinks at a bar in New York. Again, to get the motivation of your constituency is the role of governance, is the role of a governor, I want to say. It's politics basically, but even at a DAO scale, of course, when you've got $500M in treasury, you're going to make sure everybody's rowing in the same direction. 

So to pull together the constituency and say you're all connected through us, and here's how we're going to decide what our seasonal budget going to be? Because we, as a DAO, didn't say, "we think we need $4 million for the year". We literally went to the founder and said, "here's our treasury, here's what we do, here's what we can liquify without actually raising any alarms. Here's what the investors are willing to take as a devaluation hit." When we crunch the numbers 27 times to make sure they're right, what you want to do here, and then it's like, "okay, let it be $5 million this year". And so it's ordained and then it happens a month later in governance. It's just knowing what the goals of the actual stakeholders are. The people who have the ledger buttons that could stop it all in its tracks are the ones that you have to satisfy. 

What do I mean by sensemaking? I would say it's collecting input from stakeholders to better understand what's going on (internally or externally). A more academic definition can be found here: "the process of constructing mental models and meaning from ongoing environmental input in order to respond effectively", with an emphasis on complexity of the environment.

Even though I had some first-hand experience with sensemaking as a former governance lead of Protein Community, it wasn't the main problem we set our sights on when prototyping the first version of Harmonica a few months ago. Initially it was more about deliberation, as we envisioned the LLM-powered chatbot to replace forums, where most DAOs discuss proposals, or even advanced deliberation tools like Polis. But as I made progress with my research, it became increasingly clear that deliberation is actually a much lesser pain than what precedes writing a proposal or planning a season, i.e. sensemaking.

When the fellowship started, I formulated my first problem statement like this:

I am a governance lead who's trying to submit a proposal that would get sufficient support, but I'm not getting enough feedback / ideas from the community, because Discord is overwhelming, sharing on forums doesn't feel safe and people are conflict-averse.

Then I began my research to find out if it reflects reality or not. The methodology has been pretty straightforward: doing interviews with governance leads, researchers, community managers, founders and consultants, using a semi-structured guide to learn how they make decisions and facilitate discussions. I'm trying to distil my notes and AI transcripts into drivers and barriers, experiences with specific tools, customer needs, and decision timelines. The interviews, coupled with insights from a few case studies and research papers, as well as the summary of Governance Geeks Gathering workshop in Istanbul, enabled me to critically evaluate my original problem statement.

I've done 15 interviews over the last month, and I'm really grateful to all participants who spared their precious time to answer my questions. I've been lucky enough to interview people from Gitcoin, Metagov, DAOstar, Apiary, Superbenefit and other orgs who operate on the cutting edge of (web3) governance. The first batch of interviews covered a pretty wide range of perspectives on sensemaking, proposals writing and deliberation, lobbying stuff with delegates, governance consulting, professional facilitation, citizen assemblies and E2C transitions.

The research is still ongoing and I don't have a clear picture / JTBDs yet. My intention with this post is just to share a few preliminary findings, without drawing any conclusions, and outline the next steps for my research.

The first interviews revealed governance practitioners as deeply committed yet challenged by tooling limitations and cultural nuances. One leader has struggled to overcome the Discord fatigue their members developed over time, while attempts of another to foster a safe space on their forum highlight the human side of online governance. These narratives underscore the anxieties surrounding public discourse and the barriers to effective communication: a mix of tool inefficiency and personal reluctance to engage in potential conflicts.

The article about sensemaking at Radworks identified three main governance challenges: the need for a clear shared purpose, improving community insight capture, and better communication of proposal needs. These align with our findings.

The lack of shared purpose can be the underlying problem DAO people are not necessarily aware of. According to Radworks, "not having a clear shared purpose can lead to inability to design effective governance & incentive mechanisms, lower transparency on roles, responsibilities and strategy alignment, difficulties finding DAO/Product fit, and poor process for off-loading work to the community."

Speaking of proposals, at least three participants riffed on the issue of professional delegates. TLDR: it seems that many DAOs have basically re-invented the crappy old democracy model with lobbying and power games. Truth is often unpopular, so people don't want to risk their status, and popularity is prioritized over wisdom in decision-making. If someone wants their proposal to pass, they seek support from delegates before submitting it to the forum. In other words, real deliberation happens behind the scenes in private chats, which creates a gap between the promise of web3 (shared ownership, democracy, etc) and the reality that sometimes feels more like the House of Cards.

The tooling itself is a big issue. Existing tools create limitations for capturing the true essence of community insights and preferences. Tools like Discord, Typeform and forums, while useful, often fell short in facilitating constructive dialogue and collaboration. Facilitators are hesitant to use tools like Polis and run workshops in Miro within their wider communities, considering the limited time and attention spans of members. Tools like Miro prove too complex and inefficient for facilitating collaboration, with low engagement. "Testing and experimentation with new DAO tools or mechanisms remains difficult as the cost of onboarding just for an experiment remains too high for unknown benefits/community satisfaction." The use of tools like Google Docs and Discord, while facilitating some level of engagement, often fell short in fostering deep, meaningful collaboration and feedback (e.g. when setting OKRs).

As one expert noted, knowledge management systems are currently the biggest obstacle for good sensemaking. To quote Radworks again, "There is a need for better stakeholder analysis tools and services (e.g. ways to get better insights into how motives/incentives differ between different stakeholders). Better community insights = more evidence to signal how we should be designing governance mechanisms." One consultant observed the clash between traditional governance methods and the innovative potential of DAOs.

The process of sensemaking (e.g. planning the next season, aligning on specific OKRs) usually takes 1-2 months. In terms of its timeline, this could be the most common pattern described by my interviewees:

1. Interviews with delegates (Zoom)

2. Survey (Google Form or Typeform)

3. Drafting and commenting (Google Docs)

4. Workshop (Miro)

The research highlighted a universal struggle with engaging community members in governance processes, the need for clear, shared goals, and the challenges of capturing and integrating community insights into decision-making.

The initial focus on writing proposals transitioned to a more abstract problem of collecting input from stakeholders to aligning the community around a shared purpose or OKRs:

As a DAO leader I’m trying to get members aligned around a shared purpose / seasonal OKRs, but it’s hard to collect (and process) input from different stakeholders because of (1) fragmentation of community, (2) limited time and attention span of members, (3) people not being used to some of the new tools (for participation), which makes me frustrated and overwhelmed (required coordination effort is way too high!)

The revised statement reflects a deeper understanding of the underlying challenges with sense-making. The focus has shifted from proposal submission to the preceding, arguably less linear and more complex tasks of sense-making and facilitation.

It is not just about getting proposals pass as much as nurturing an alignment and/or coherence. The emphasis shifted towards leveraging tools and methodologies that could support effective sense-making, despite the constraints of time, attention span, and familiarity with new tools.

This article by Aviv Ovadya presents a very actionable framework for designing better decision-making systems, with alignment, competence, and robustness being the key metrics or themes:

https://aviv.medium.com

Build Wise Systems: Combining Competence, Alignment, and Robustness

This is the first piece in a set of work on 'Reimagining Technology'; Most of my prior public work focused on misinformation, online platforms, platform/AI governance, and the impacts of AI/ML. This...

A slightly more refined version of the same statement goes like this:

Facilitation of sensemaking, e.g. to formulate a shared purpose or seasonal OKRs, takes too much time and effort - normally 1-2 months of hard work to engage more people and process their inputs (those OKRs won't resonate with members who don't participate!)

Digging a bit deeper into sense-making in DAOs reveals the critical need for innovative solutions that address the nuanced challenges of sense-making, community engagement, and strategic alignment. The insights uncovered through this research will inform our ongoing work. Harmonica's design and GTM strategy will focus more on sense-making capabilities and facilitator's needs.

The exploration of sensemaking as a problem space uncovers a landscape ripe for innovation, where the collective wisdom can lead to more inclusive, efficient, and meaningful governance.

The transition from the initial to the new problem statement reflects the nuanced dynamics of decentralized governance. It confirms the critical need for sense-making. Engaging community members in this process, despite their limited time and familiarity with tools, remains a significant hurdle.

The first phase of my research highlighted the importance of developing shared purpose, strategic scaffolding and success criteria to foster a cohesive and engaged community. It not only informs a more focused and nuanced inquiry for the second phase, but also underscores the universal need for better sensemaking and even vocabulary needed to discuss this issues.

We're not just building a new product; we're generating knowledge every week. This blog will serve as a canvas for our explorations and discoveries. Currently, we're diving into user research on DAO governance, as part of the Arbitrum CoLab fellowship we've recently joined. But that's just the start. Expect insights on the evolving market for decision-making tools, and how organizations are integrating AI into their management frameworks.

We believe in the power of AI to revolutionize sensemaking and deliberation processes. This evolution could have profound implications for our society and democratic institutions. We're optimistic, and why shouldn't we be? The fusion of AI and democracy holds the potential for a more participative, resilient, and happier society. We'll be sharing our thoughts, ideas, and perhaps a few bold predictions about this space.

The way we work is changing, and tools like Harmonica are at the forefront of this transformation. We envision a future where collective intelligence isn’t just a buzzword, but an integrated part of how teams and communities function online. Our blog will delve into these changes, offering perspectives on how AI-powered collaboration (perhaps even swarming) is shaping this new landscape of work.

At the heart of Harmonica is our desire to empower digital commons. We see a future where these groups are more resilient, better coordinated, and efficiently governed. Our blog will cover the exciting world of DAOs, Discord communities, and other digital congregations, highlighting the role of tools like Harmonica in their evolution.

We’re on a journey, and we want you to be a part of it. Stay updated with our latest announcements, prototypes, and product news. Witness first-hand how Harmonica evolves, adapts, and grows to meet the ever-changing needs of our users.

Become a member to support us

Your support can make a tangible difference in the development of Harmonica. By buying a membership, you contribute directly to our mission. Every penny we earn from Paragraph, our blog platform, will be funneled into our Open Collective account. There, you can see exactly how your contributions are being used – transparency is key for us. Your support means more than just funding; it’s a belief in our vision.

Stay tuned, stay curious, and let’s embark on this adventure together. Cheers to new beginnings and the exciting road ahead! 🥂