Author : Christopher Choi (@christopher_c), Damon Lee (@i_naiveai)

Fission is an AI research and development team that specializes in model optimization—making advanced AI models more efficient, scalable, and accessible to a broader range of users. Rather than pouring endless resources into ever-larger models, we focus on streamlining computing requirements and enhancing model performance through smarter, more targeted approaches. Our guiding principle is simple yet profound: the best AI emerges when technical innovation is combined with high-quality, iterative human feedback.

To achieve this, we design workflows and tools that invite user participation throughout the AI lifecycle. From initial data refinement to ongoing feedback loops, we’ve learned that human insight—especially when harnessed at scale—often trumps brute-force computation. By refining AI models through collective user input, we aim to craft systems that are both robust and grounded in real-world needs.

As we delved deeper into model optimization, we discovered a critical tipping point: no matter how elegant or efficient our algorithms became, they still depended on authentic human feedback to truly excel. While data labeling had been the traditional approach, we realized that continuous, context-rich evaluation—from domain experts, enthusiasts, and everyday users—was the key to fine-tuning AI in a more dynamic, impactful way.

However, gathering and managing this feedback via centralized methods proved costly and time-consuming. It was also difficult to maintain transparency and fairness at scale. That challenge led us to explore more decentralized structures—platforms where users could offer feedback and be rewarded in ways that felt natural, equitable, and community-driven.

That’s where DeSci (Decentralized Science) entered the conversation. DeSci promised an open, collaborative environment for knowledge-sharing and verification, yet it quickly became clear that advanced AI was needed to handle large data streams and real-time inputs. Combining DeSci’s ethos with AI-driven workflows gave rise to DeSAi—a synergy merging decentralized collaboration with iterative optimization.

In this article, we share how our journey in AI model optimization naturally led us to embrace user-driven feedback at scale, eventually aligning with DeSci principles to form DeSAi. One remaining challenge was how to govern and incentivize these decentralized feedback loops efficiently and fairly. After evaluating various governance models, we settled on Futarchy, which uses prediction markets to guide resource allocation and decision-making.

Ultimately, we want to articulate why user feedback is not just a ‘nice-to-have,’ but the linchpin for genuinely optimized AI. We also explain how combining DeSci, AI, and Futarchy can open the door to a vibrant, sustainable, and impactful research ecosystem. Our hope is that this discussion inspires fresh ideas, invites collaboration, and expands what AI can achieve when fueled by collective human insight.

Modern AI systems have grown to a scale few anticipated—think of large language models (LLMs) with billions of parameters or vision transformers processing massive multimodal datasets. While these models exhibit notable capabilities, they also highlight structural issues:

1. Resource Intensiveness: Training top-tier LLMs demands enormous GPU/TPU resources, favoring only the most heavily funded institutions.

2. Data & Quality Bottlenecks: Gathering (and curating) comprehensive datasets can be logistically daunting, with the risk of poor generalization if hastily compiled.

3. Diminishing Returns: Adding parameters yields marginal performance boosts compared to the steep rise in infrastructure, time, and R&D costs.

These factors often concentrate AI advancement in a small number of elite labs, hindering community-driven or smaller-scale innovations. From Fission’s vantage, chasing bigger models forever is no longer tenable.

Instead of pursuing ever-larger architectures, Fission focuses on model optimization that balances performance and accessibility. By using methods such as knowledge distillation, parameter pruning, and domain-focused fine-tuning, we find that smaller, specialized models can match—or surpass—monolithic systems when properly contextualized. This underscores the importance of relevance: a well-targeted, efficiently designed model can outperform a generic behemoth.

However, we realized that no matter how well we compress or refine a network, authentic human feedback remains critical for genuine success. Consistent, context-rich feedback from real users—be they experts or casual participants—is vital for aligning AI with real-world demands. Model size or clever compression only solves part of the problem; the rest involves harnessing collective human insight to detect errors, refine logic, and adapt to nuanced tasks.

Reinforcement Learning from Human Feedback (RLHF) completes this puzzle by embedding user dialogue into training loops. The AI asks clarifying questions, receives critiques, and integrates new suggestions over time. This approach:

• Enhances Accuracy & Adaptability: Subtle domain requirements or shifting contexts can be incorporated quickly.

• Unearths Biases or Gaps: Human spot checks can reveal flaws not apparent in static benchmarks.

• Strengthens Ethical Alignment: Real stakeholders guide moral and situational reasoning, beyond curated datasets.

Scaling RLHF beyond small experiments, however, proved complex—collecting and verifying large amounts of feedback demands robust incentives and transparent coordination. Centralized methods are either too slow or too opaque, but DeSci’s ideal of decentralized collaboration offer an achievable solution. Combining DeSci’s open ethos with bespoke AI pipelines leads us to DeSAi, with Futarchy's governance and incentive structure completing Fission's system to ensure unrivaled fairness, alignment, and perpetual evolution.

While standard AI optimization focuses on efficiency and resource usage, privacy-first contexts—notably Fully Homomorphic Encryption (FHE), Multi-Party Computation (MPC), and Trusted Execution Environments (TEE)—add further complexity. These technologies promise data confidentiality even during processing, but at the cost of significant overhead. In Fission’s view, unoptimized AI models are simply infeasible in these privacy-preserving environments, especially when tackling anything beyond small or highly specialized tasks.

1. Massive Computational Slowdowns

   • FHE’s Slowness: Homomorphic encryption operations can be orders of magnitude slower than plaintext equivalents, pushing practical AI usage to the limit. Large LLMs, for instance, remain far out of reach for real-time or high-volume scenarios.

   • MPC’s Overheads: Distributing computations across multiple parties drastically increases communication and coordination complexity, raising latency. While smaller ML models or stats-based tasks can run in production, deeper neural networks face steep performance drops.

   • TEE Resource Limits: Although faster than FHE, secure enclaves provide minimal computing and memory resources, making bigger architectures or real-time inferencing difficult without heavy pruning or quantization.

2. Memory & Storage Constraints

   • Encrypted Data Bloat: FHE and certain MPC schemes inflate data significantly; naive model designs cause memory usage to spiral.

   • TEE’s Constrained Enclaves: The tight memory caps in TEEs hamper large, complex models. Even if an enclave can store the model, throughput and responsiveness often suffer.

3. Latency & Real-World Responsiveness

   • FHE’s Latency: Encrypted inference can multiply the computation time by hundreds to thousands, ruling out latency-sensitive tasks.

   • TEE’s Bottleneck: Although enclaves run at near-native speeds, the overhead from secure context switching and limited resources still hinders real-time AI at scale.

Through our expertise in knowledge distillation, parameter pruning, and domain-focused fine-tuning, Fission aims to lower the resource threshold at which FHE, MPC, and TEE-based AI can operate. While we don’t claim to eliminate all overhead, these optimizations can:

1. Shrink the Model Complexity Reduced parameter counts or domain-specific architectures make encrypted computations or enclave operations somewhat more feasible.

2. Enhanced Domain-Specific Focus By tailoring AI to narrower, domain specialized tasks, we avoid the cost of unneeded generalization, a crucial step when every operation is slowed by cryptographic layers or secure enclaves.

3. Integrate Human Feedback Instead of retraining from scratch, iterative feedback (not limited to RLHF) can refine the model on-the-fly, minimizing repeated heavy computations in encrypted or secure environments.

Importantly, these measures serve as enablers rather than a cure-all. Large LLM-style networks remain too big for typical FHE or TEE constraints, and real-time use cases often can’t yet absorb the latency. However, incremental efficiency gains can make more specialized tasks or smaller-scale AI feasible—particularly as the hardware and protocols continue to evolve.

Current Reality: For many large-scale or real-time applications—like GPT-scale language models or complex image processing—FHE, MPC, and TEEs remain prohibitively slow or resource-intensive. Even advanced solutions run 100–1000× slower than plaintext, making it impractical to offer latency-sensitive features.

Where It Shines:

• Smaller Domain Models: Stripped-down classifiers or specialized NLP tasks can thrive with heavy pruning and domain-specific tuning, turning “private AI” from theory to limited but impactful reality.

• High-Value, Lower-Volume Use Cases: Medical diagnostics, private financial analytics, or specialized industrial monitoring can justify performance trade-offs for strong privacy guarantees.

• Future DePIN-Style Devices: As specialized hardware (e.g., next-gen GPUs with TEE features) and advanced cryptographic schemes mature, we may see truly private AI extended to decentralized physical infrastructure networks, bridging on-device inference with privacy-by-design.

Road Ahead:

• Cautious Optimism: The field moves quickly, with new “lightweight” FHE or specialized MPC protocols emerging. Tailoring model design to these constraints—rather than fighting them—could unlock new frontiers in privacy-protected AI.

In short, Fission embraces a realistic stance: FHE, MPC, and TEE solutions for AI remain constrained, but not impossible. Through rigorous model optimization and incremental improvements, we believe privacy-preserving AI can support domain-specific tasks today—and possibly break new ground as technology evolves.

Labeling raw data is fundamental to AI development—transforming unstructured information into training-ready inputs. But while it’s effective initially, such labeling is often treated as a one-off process. In practice, AI models benefit from continuous, contextual refinement, where human users regularly assess and adjust outputs. At Fission, we discovered:

• Static vs. Evolving Needs: Traditional labeling can be rendered obsolete if objectives change. Iterative feedback aligns AI outputs with new goals in near real-time.

• Surface vs. Subtle Insights: Simple labels capture obvious distinctions. Ongoing user feedback can reveal deeper nuances (tone, cultural context) that static labels can never convey.

• Higher-Order Reasoning: Users can flag logic errors or inconsistencies, driving the model to adapt beyond mere classification.

This nuanced evaluation—often via RLHF—treats humans as active collaborators, shaping the AI toward more reliable, meaningful performance.

Despite its merits, iterative feedback poses a significant question: how to gather, validate, and integrate vast user inputs efficiently? Centralized methods typically require a single entity to manage large contributor pools, fund compensation, and audit authenticity. Such an approach is:

• Costly: Maintaining infrastructure and paying contributors can be expensive, skewed by top-down funding.

• Time-Intensive: Heavy managerial oversight slows data pipelines, limiting quick iteration.

• Limited in Diversity: Centrally curated platforms may inadvertently filter out or ignore essential perspectives.

These issues become glaring in advanced model contexts demanding wide-ranging, high-quality feedback. A more distributed model is needed—one that expands participation without exploding overhead.

The typical, centralized approach to handling vast amounts of user feedback—employing dedicated staff or contracting large labeling farms—rapidly becomes both expensive and slow. As participation scales, so do overheads in quality assurance (QA) and authenticity checks. We set out to design a more automated, consensus-driven process that could verify and integrate user contributions without a central gatekeeper.

At Fission, we break down user feedback into modular “micro-tasks,” each requiring incremental checks—ranging from simple binary judgments (e.g., “Is this output correct?”) to more nuanced domain validations (e.g., “Is this financial forecast logically sound?”). Our system then relies on a distributed consensus layer where multiple participants, each with a stake or reputation in the ecosystem, reach agreement on the correctness of each micro-task result.

• Staked Participants: Individuals or small groups commit tokens or reputation collateral, signaling they will evaluate tasks honestly.

• Aggregated Scoring: Feedback from different stakers is combined, filtering out outliers or coordinated attacks. The network then consolidates these partial QA results into a final consensus.

• Predictive/Automated Checks: Machine-driven checks (e.g., heuristics or smaller verification models) further scrutinize the aggregated results to catch inconsistencies, spam, or sabotage attempts.

Instead of funneling all user submissions through a single manager or slow editorial pipeline, this approach distributes QA across many semi-autonomous verifiers. Drawing on a synergy of human oversight and lightweight automation, we can:

• Reduce Managerial Overhead: No large staff is required to manually screen and compile thousands of user comments.

• Empower Diverse Expertise: Domain specialists can “plug in” to certain micro-tasks, ensuring that high-level inputs—such as legal or financial feedback—are vetted by relevant experts.

• Achieve Real-Time Scaling: As the user base grows, so does the pool of staked validators, enabling parallelized, continuous QA without significant slowdowns.

To further motivate participants, we incorporate futarchic incentives: each verifier can stake on the veracity or long-term value of proposed changes, aligning rewards with accurate judgments. By tying gains to outcomes rather than subjective opinions, the system automatically penalizes spam or shallow claims, while surfacing genuinely valuable insights.

1. Market-Driven Quality: Overstated or deceptive inputs lose money for those backing them.

2. Collective Ownership: Everyone from domain experts to curious enthusiasts can partake, transforming QA into a community-driven project.

3. Iterative Model Refinement: As micro-tasks are accepted, the model is updated in near real-time or at scheduled intervals, bridging large-scale feedback with a fluid, decentralized governance.

Ultimately, this consensus-based QA framework supplants the traditional, centralized workflows with a flexible, community-powered mechanism. By blending user-friendly micro-tasks, automated checks, and futarchic incentives, Fission ensures continuous, high-quality feedback loops that scale with user engagement—without compromising fairness or speed.

DeSci aims to decentralize authority over scientific discovery, championing open data, inclusive participation, and transparent governance. Yet, in the realm of large-scale AI:

• Data & Real-Time Feedback: Traditional DeSci structures don’t inherently solve the computational complexities or near-instant feedback loops AI demands.

• Sustained Incentives: A volunteer-based approach can falter unless contributors see tangible rewards for their input.

• User-Friendly Interfaces: Many blockchain-based or decentralized protocols can be cumbersome, deterring non-technical participants.

DeSAi emerges from the understanding that DeSci’s collaborative ideals need advanced AI pipelines to handle substantial data flows and iterative improvement. By uniting the two, we create an ecosystem where:

• Data Streams Flow Freely: Community members supply domain insights and real-time user feedback, powering continuous updates.

• Models Evolve Contextually: AI pipelines process these inputs near-instantly, adapting to new discoveries and niche requirements.

• Transparency Thrives: On-chain governance logs data, decisions, and model changes for universal verification.

In essence, DeSAi supercharges DeSci’s openness, applying it to demanding AI tasks that benefit from both scale and community-based oversight.

Fission anchors DeSAi on two complementary pillars—one grounded in efficient architectures and the other in continuous, people-powered refinement:

1. Model Optimization: We minimize compute overhead for AI, enabling smaller teams or resource-limited domains to participate in decentralized projects. Rather than infinitely scaling parameter counts, our emphasis is on techniques like parameter pruning, knowledge distillation, and domain-specific fine-tuning. This keeps models agile, interpretable, and better suited for large-scale collaboration where hardware constraints and network demands can be major bottlenecks.

2. Iterative Human Feedback: While advanced model designs offer a strong foundation, human feedback loops—ranging from casual user hints to expert critiques—ultimately shape an AI’s real-world utility. Reinforcement Learning from Human Feedback (RLHF) is one proven methodology, but our approach isn’t confined to RLHF alone. For instance:

• Simple Upvote/Downvote Mechanisms: Community members can quickly signal approval or spot obvious errors.

• Structured QA: Domain specialists contribute nuanced clarifications—whether in legal, medical, or financial contexts—enabling targeted improvements.

• Micro-Task Verification: Curated, bite-sized evaluations ensure broad participation while maintaining robust quality checks.

As our user base expands, so does the volume and variety of feedback. A globally distributed pool of contributors offers deep domain knowledge, cultural awareness, and creative problem-solving beyond what a purely algorithmic approach could achieve. By uniting efficient architectures with iterative feedback, we maintain a dynamic cycle of improvement—one that’s agile enough to pivot when user needs change, and powerful enough to push AI’s boundaries through genuine community insight.

Hence, Fission’s DeSAi framework transforms an aspirational DeSci concept into an accelerated discovery engine, fueled by AI shaped by diverse stakeholders. This synthesis upholds DeSci’s ideals of transparency and inclusivity while adding the computational heft to handle large-scale, continuous research in AI optimization.

Within crypto and Web3, “hyper-financialization” often calls to mind unfettered speculation. Yet Fission believes it can be leveraged constructively, turning market enthusiasm into transparent, outcome-oriented governance. Futarchy serves as the mechanism that aligns economic incentives with real, data-backed progress—rather than hype.

Futarchy typically unfolds in two steps:

1. Define Key Metrics: The community agrees on a quantifiable goal—model accuracy, real-world usage success, or user satisfaction.

2. Bet on Proposals: Participants stake tokens on whether a suggested improvement will enhance that metric. Verified outcomes reward accurate bettors; incorrect guesses lose their stake.

By centering resources on verifiable performance instead of promises, Futarchy transforms financial speculation into a community filter. Hollow claims quickly lose credibility and capital, while solid ideas gain wide support.

In the DeSAi framework—where user feedback loops (such as RLHF) are crucial—Futarchy ensures:

• Outcome-Focused Upgrades: Market-based consensus directs energy toward meaningful improvements.

• Resilience Against Gaming: Artificially inflated metrics can be shorted by informed participants, driving a self-correcting process.

• Sustainable Evolution: Participants who believe in a model’s long-term potential can reallocate or hold rewards, reinforcing community-led growth.

Thus, Futarchy harnesses hyper-financialization—the inflow of capital seeking ROI—and directs it toward genuine, metric-proven gains in AI.

While hyper-financialization often conjures up fears of bubble-like speculation, Futarchy channels it into a system that prioritizes legitimate breakthroughs and penalizes baseless claims:

• Reality Surfaces Quickly: If a metric is gamed or inflated, on-chain data eventually contradicts it.

• Collective Accountability: Savvy participants profit by uncovering weak proposals, reinforcing honest, data-driven progress.

• Integrity Matters: Project owners or researchers proposing improvements must deliver real results if they hope to gain from the predictive markets.

In this way, hyper-financialization—managed under a futarchic framework—acts as a filter for authenticity, ensuring speculation aligns with actual model improvements.

From Fission’s viewpoint, leaning into hyper-financialization fosters economic sustainability for large-scale AI endeavors, guiding capital to promising initiatives. By letting market dynamics and predictive markets surface genuine progress, we unite community-driven research with iterative AI optimization:

• Transparency at Scale: On-chain governance logs each prediction, outcome, and model update.

• Accelerated Innovation: Sound proposals gain backing, fueling expansions in data coverage and RLHF enhancements.

• Collective Ownership: Rather than a few influential backers dictating direction, anyone with insight can guide AI’s path.

Put simply, Futarchy transforms hyper-financialization into a supportive engine—turning pure speculation into communal intelligence that advances DeSAi’s broader goals.

While Fission’s foundation lies in model optimization and human-centric AI, we also partner with enterprises, providing on-premise deployments and continued technical support. This professional service model sustains our broader DeSAi aspirations. Our aim is to reconcile commercial viability with the iterative, open feedback loops needed to evolve AI responsibly.

We chart a series of phased objectives:

1. Alpha Integration of RLHF Pipelines

   We’re piloting user feedback interfaces that allow domain experts—and eventually broader groups—to submit clarifications directly into training loops.

2. On-Premise AI Deployment

   We refine a licensed toolkit for enterprise use, balancing data privacy with aggregated insights for model refinement.

3. Adaptive Governance Modules

   We experiment with futarchic governance in tasks such as prioritizing new features. Components may be shared under specific agreements, based on our evaluation of open vs. permissioned usage.

By demonstrating tangible improvements—like local inference or RLHF capacity—we build trust through concrete value.

Enterprise partnerships ensure financial stability, but we also want everyday users shaping AI outputs:

• miniApps: A lightweight app invites volunteers to test AI outputs, flag errors, or offer refined answers. Participants might earn small perks (e.g., recognition, possible token incentives) while the model gains real-world perspective.

• Community Challenges: Periodic themed events—like “Healthcare QA Challenge”—unite domain experts, letting them guide the model’s growth. Over time, futarchic governance could help allocate resources or rewards for these challenges.

By combining commercial engagements with grassroots input, we reinforce the notion that robust AI emerges from diverse collaboration—encompassing industry, academia, and everyday contributors.

Fission’s research scope is firmly anchored in the AI realm. Alongside enterprise solutions, we pursue co-studies with universities and labs:

• Academic Pilots: Graduate researchers test advanced RLHF techniques, while Fission supplies resource-friendly architectures and partial funding. Results may be published, offering external validation.

• Sector-Specific Innovation: Where a lab excels in specialized AI topics (e.g., multi-lingual RLHF or interpretability), we co-develop solutions and integrate them into our broader pipeline.

This collaboration ensures well-tested, academically grounded enhancements rather than isolated in-house ideas.

Progress will be shared in increments:

• Technical Updates: Performance metrics or user satisfaction scores from on-prem deployments, plus futarchy trials.

• Enterprise Case Studies: Real pilot outcomes demonstrating local AI gains, along with continuous user feedback loops.

• Academic Collaborations: Papers or prototypes revealing how RLHF and futarchy can expand AI’s boundaries.

Through enterprise deals that finance robust AI deployments, academic ventures exploring advanced solutions, and mini-apps engaging everyday users, Fission follows a practical, stepwise strategy. We invite:

• Enterprises seeking on-prem AI to maintain data sovereignty and feed anonymized feedback into the ecosystem.

• Universities & Research Labs for specialized RLHF or futarchic governance studies, grounding DeSAi in peer-reviewed rigor.

• Enthusiasts & Experts to join test apps, share domain insights, and shape a governance model that benefits a larger AI community.

Gradually, we aim to prove that a deeply human-aligned, decentralized AI vision is both pragmatic and achievable, rooted in real products, real partnerships, and a deliberate, incremental roadmap.

From our foundational work in model optimization to the evolving concept of DeSAi, a common theme emerges: meaningful AI progress depends on human collaboration, open research structures, and data-driven governance. We believe that community-sourced feedback channeled through advanced AI pipelines and guided by prediction-market incentives can produce more equitable and impactful innovations than any single stakeholder could manage alone.

To that end, we’ve laid out a pragmatic roadmap: on-premise enterprise deployments that sustain research and development, mini-app engagements inviting everyday users to refine model logic, and academic partnerships that anchor our endeavors in scientific rigor. Step by step, we aim to demonstrate tangible value rather than make sweeping promises.

Ultimately, DeSAi is an actionable approach for AI that evolves under real-world conditions and transparent incentives. Whether you’re an enterprise seeking localized AI capabilities, a researcher exploring new optimization techniques, or someone interested in shaping AI outputs in a collaborative environment, we invite you to join us. Together, we can ensure the future of AI remains human-aligned, ethically grounded, and open to genuine participation all while moving forward with a deliberate, incremental build-out of this new paradigm shift.

Hub | 𝕏 | Discord | TG (Ann) | TG (Community)