You can check it out on iOS via TestFlight now and see it in action on the Subset landing page. Or keep reading for more on each of these additions...

Saving, sharing and searching should be fun. For most, fun is intertwined with indicators of progress. Subset now has some. In the list view, tap the three bars in the top-left to slide out a tray. At the top, you'll see your stats.

There's an indication of your current streak status. One save, share or search a day activates or preserves it—you can also miss a day (only one, and only one missed day per streak) and still maintain it. Underneath the streak, you'll see your save, share and search activity for today and an indicator comparing it to yesterday's counts—totals for yesterday and the last 7 days are there, too.

The counts work like so: a save is counted when you save something via the share extension or do an import, a share is counted when an item is shared and a share record is created, and a search is counted when you enter a search query and take an action (view item detail, open, share, archive, delete).

We've had some sharp search syntax from the jump:

• Full text search (e.g. "biology")

• Phrase-specific search (e.g. ""biology"")

• Field-specific search

  • "t" / "title"

  • "u"/ "url"

  • "e" / "excerpt"

  • "n" / "privateNote"

  • "s" / "share"

• Date-specific search

  • "c" / "created"

  • "m" / "modified"

However, it only worked for single queries. Now, compound queries are possible.

• Search for multiple phrases in specific fields—e.g. "t:cafe url:youtube".

• Do searches involving created and modified dates using YYYY-MM-DD format—e.g. "c:2024-12 u:newyorker", "m:2025-01-16 u:gwern logic".

• Use "s" / "share" to search a named app from a share record, any share notes, what's included, and (if linked) the URL you linked the record to—e.g. "s:whatsapp menu", "c:2025-01 s:linkedin u:wiki".

• See all your archived items and those with share records by typing "a:1" / "archived:1" and "s:1" / "share:1" into the search field, respectively.

Compound queries need to have their sub-components separated by a space, and they're all ANDs—"c:2024-12 u:newyorker e:macroecon" equals "Created December 2024 AND has "newyorker" in the URL AND has "macroecon" in the excerpt". Fancy operators (NOT, OR etc.) and things like date ranges and relative dates are pending.

Often URLs that are saved or copied come with parameters—"https://socialmedia.com?s=1343&p=dw2f", for example. Sometimes, there's fragments too—"https://en.wikipedia.org/wiki/Numerical_tower#In_Scheme".

Occasionally, parameters and fragments have meaning. But most of the time they're used by products and services to capture or generate some information about the entity grabbing the URL and doing things with it. This is fine when all you're doing is storing it. But if you're sharing it or frequently accessing it, it's nice to remove the nonsense. So we added the ability to clean 'em up.

Say you've saved a post from X / Twitter and it has some parameters. In the Subset share extension and the detail view in the app, a "Clean URL" button will appear. It will bring up a dialogue that allows you to strip all the parameters and fragments, or select specific ones to remove. You'll be shown the original, what's being removed and the new URL, before confirming the action and saving the new version. And if the title contains the removed elements (which it will if the title is derived from the URL), they'll be scrubbed from there, too.

The things we save have a disproportionate important in the digital realm. They're harbingers of meaning. So the ability to access and preserve them matters. Thus, basic import and export mechanisms.

The export feature is simple. It allows you to manually exfil your Subset as JSON to an arbitrary destination of your choice. In the background, we also do continual back-ups to your device's filesystem, for safety's sake.

The import feature is for safety's sake, too. It allows you to re-import previous exports, should something happen. It also allows you to import from other systems, though only in a specific JSON format for the moment (available upon request).

On successful import, you'll see a summary of what changed.

We already have native sharing, the ability to save something in Subset and share it with other people on other apps, creating a searchable record in the process. But wouldn't it be helpful if you could just share directly with your peers, instead of having to route it through another service?

Yes, yes it would...

Say you've saved something to Subset. Now, you can share it to X, maybe adding an excerpt or special note. That share gets recorded and becomes searchable within Subset. Later, you can link the outcome (the X post URL) to the share record. You can get Subset on iOS via TestFlight and try this out now. Or you can see it in action on the Subset landing page.

This isn't an ideal approach, by any means. But right now, it's the best way to match share actions to outcomes. To see why, think about how many ways the average interested person shares things. They post on Farcaster or Bluesky or X. They send messages in work Slacks and community Discord channels. They push links in emails and DMs and group chats. And they often want to reference the things they've posted and sent and pushed. But because this sharing activity traverses public, private and semi-private spheres it remains mostly illegible.

There are four ways to make it legible as a whole, though, to connect share actions (e.g. post this link and a witty comment to X) with their outcomes (e.g. the URL of the X post):

• Platform integration

• Protocol interop

• User labour

• Agents

Platform integration means playing nicely with likely-adversarial entities that have a vested interest in maintaining and further enshittifying the web. Even if these entities are neutral or benevolent, it still relies on APIs that are never completely transparent about their inherent capabilities and tradeoffs. Further, even if the entities are neutral/benevolent and they provide rigorous APIs, it creates dependencies on non-1st party resources that have to be continually tracked and maintained and that scale non-linearly as the user base increases.

Protocol interoperation is an alternative. Protocols are designed to promote interoperation and thus represent way less risk than integration with traditional entities and their platforms. And although protocol developments must still be tracked and maintained, this is simpler than platform integration maintenance—interop with every protocol is plausible; integration with every platform is not. Unfortunately, most of the digital realm is still composed of platform fiefdoms, not protocol-ish freedom-loving infrastructure. Protocols as they are only account for a small portion of total sharing activity.

User labour—the thing we've opted for, for now—is next up. Instead of platform integration or protocol interop, the load of mapping can be shifted to a user in the wake of a share. This requires minimal overhead because the user is already auth'd on their device and they're able to navigate enveloping platforms or protocols at will. However, mapping does introduce a severe context switch and flow-break that's unlikely to be sustained in the long-term.

If I'm reading an epic paper in Safari, it's simple enough to punt it to Discord. It's just that getting the link to the outcome from Discord requires me to open the Discord app, nav down to the location of the message, get the URL, and map it to the relevant share record in the Subset extension. The ideal Subset user is a prolific sharer whom, whilst not ecstatic about such a quest, would undertake it in some cases on a consistent basis. But it's still a whole lot of excessive friction.

The final possibility—agents—is where it gets interesting. We don't mean everything agents. We mean Unix-like, Boydian burglar agents that have:

• A close alignment to a singular user or entity and their actions

• A deliberately constrained set of capabilities and protocol-savviness

• A mission-driven approach to objectives

• An ability to imaginatively manipulate the space they traverse

To circumvent excessive user labour, an UBB agent could be introduced to perform the same task. The UBB agent doesn't need to touch all the things a user does—it just needs to spin up, receive and execute a mission, then self-destruct. The mission an agent receives in the wake of a sharing action essentially resembles:

• Materialise post-sharing action

• Receive mission packet with target app/service and contextualising info

• Go to target app/service and use contextualising info to locate outcome

• Return with and confirm successful mapping

• Dematerialise post-mapping

This can occur in the background whilst the user continues in the flow of their original activity and may include arbitrary agent-user touches (e.g. to authenticate or query). Existing precedents for these behaviours exist on more permissive devices and OSs (e.g. mobile agents on Android) as research or academic projects, and they're not present at all on more restricted ones (e.g. iOS). Agentic tech just isn't consumer ready.

Thus, the state of the game: spinning up performant platform integrations for all possible sharing activity isn't viable; there's no set of protocols that adequately covers potential user sharing activity anyway; and agentic tech is neither sophisticated enough nor sufficiently empowered to map on behalf of users in an autonomous manner. So we're left, for now, with user labour. But that's okay—most of the labour is in the mapping.

The sharing? That's real easy.

And with Subset's latest version, it just became searchable.

Can a neural model running in real-time simulate a complex game at high quality?

In this work we demonstrate that the answer is yes. Specifically, we show that a complex video game, the iconic game DOOM, can be run on a neural network (an augmented version of the open Stable Diffusion v1.4 (Rombach et al., 2022)), in real-time, while achieving a visual quality comparable to that of the original game. While not an exact simulation, the neural model is able to perform complex game state updates, such as tallying health and ammo, attacking enemies, damaging objects, opening doors, and persist the game state over long trajectories.

GameNGen answers one of the important questions on the road towards a new paradigm for game engines, one where games are automatically generated, similarly to how images and videos are generated by neural models in recent years. Key questions remain, such as how these neural game engines would be trained and how games would be effectively created in the first place, including how to best leverage human inputs. We are nevertheless extremely excited for the possibilities of this new paradigm.

We're excited too, because this new paradigm is not limited to video games. It has more mundane applications: it opens up entirely new search spaces. Those which are a feature of both antiquity and science fiction. It hints at generative memory palaces. Interactive representations of saved things that can be inhabited by end users. To get a feel for this speculative possibility, let's examine its constituent elements.

First, we have an end user's collection of saved things. An annotated, meaningful assortment of succinct references to artefacts that have been encountered across space and time. These saved things could be consolidated in a single resource or system, or spread across many apps, stores, platforms and hidey-holes.

Second, we have the concept of memory palaces:

...an imaginal technique known to the ancient Greeks and Romans and described by Yates (1966) in her book The Art of Memory as well as by Luria (1969). In this technique the subject memorizes the layout of some building, or the arrangement of shops on a street, or any geographical entity which is composed of a number of discrete loci. When desiring to remember a set of items the subject 'walks' through these loci in their imagination and commits an item to each one by forming an image between the item and any feature of that locus. Retrieval of items is achieved by 'walking' through the loci, allowing the latter to activate the desired items.

A memory palace is an abstract, imagined place that's highly salient to its owner and that is designed to yield information via queries that map to fundamental metaphors of physical being.

Third, we have the realm of spatial computing. This is a domain of human-computer interaction that leans into the embodied cognition of humans and attempts to materialise novel interfaces that are richer, more intuitive and more powerful than conventional graphical user interfaces. Weaker versions of spatial computing are approaches exemplified by organisations like Dynamicland. Stronger versions are those we associate with science fiction—immersive headsets and the like.

Finally, we have the capacity to procedurally generate interactive worlds and agents. This could be the more cutting edge world-gen approaches as shown in the GameNGen paper above, as well as the agent-gen approaches exhibited by the recent Project Sid and the older NeuralMMO. This could also be the more established approaches of current procedural generation operative in video games now.

Saved things; memory palaces; spatial computing; generative worlds and agents—smash them together and what do you get? Here's an imaginary ethnography vignette that fleshes it out...

He ambles to a cozy, cluttered nook carved out of his squished apartment. His workspace. Settling into the familiar chair, he sighs. The trip's over, and work beckons. But not quite yet.

"Aria, let's review some saved items from the trip," he says, reaching for his headset and donning it.

"Of course, sir," Aria responds, a hint of amusement in their voice.

"Procrastinating, are we?"

He chuckles, slipping on the headset.

"Just getting in the right headspace."

"Sure," replies Aria.

"Can you put on the 'new finds' playlist for me?"

A gentle melody trickles into his ears whilst his grand virtual library materialises around him, the pulsing tomes acting as beacons of sense and coherence amidst the aesthetically erratic ordering of books.

He filters for the additions from the weekend and browses idly, summoning snippets of conversations, fleeting thoughts, and other stored context. A faintly glowing volume catches his eye. He looks closer at the accompanying semantics. His heart rate spikes a little.

"What's this one, Aria?"

"Jao sent it to you on Saturday. It's a preprint of a meta-analysis about how infants experience and develop agency in childcare settings. It's going to make quite the splash—to say its conclusions contrast with established consensus is an understatement."

He nods, and picks up the blank book on his tiny desk, along with a stylus. As he opens it, pages fill with text and diagrams. Flipping through, he drags the stylus' point over the meta-analysis's sentences and his idle curiosity transforms into focused interest.

"This... this could change everything for Sarah's lesson planning project," he mutters, an idea forming.

He sends it to her with a note; the morning's meeting that was slated for crisis management? It may have transformed into a wedge for a new opportunity.

On a whim, he shares it with his friend Alex, too. They've always bonded over the obviously important but often absurd nature of his work—even if they always disagree on what to do about it. Alex will appreciate this latest find for the dramatic conclusions, at least.

"Aria, can you connect me with Sarah?" he asks, a new energy in his words.

Sarah's voice echoes around him.

"This early. On a Monday. Really."

He smiles and says nothing whilst Sarah scans the paper. A minute later, her face appears in view.

"Hmm," she says.

"You don't agree?"

She raises an eyebrow.

"You do?"

Her eyebrow raises an increment more.

"Okay, let's meet ten before then. Ciao."

He cuts her connection, imagining the exclamation and exasperation it would provoke and looking forward to the comeback. He sighs again.

"Alright, Aria," he says, preparing himself for more mundane work with renewed purpose, "I guess it's time we get started."

"Indeed, sir," Aria responds, a chirp in her voice. "Shall I pull up your priority tasks?"

"No, let's start with the urgents."

Aria doesn't respond. He turns his head, locates her virtual form. She's grinning and holding up two fingers. He nods. Aria keeps one finger raised and shows him a picture. He nods, intrigued. Aria raises a second finger and shows him another image.

"Jenny's birthday," he says.

"June 26th, sir. A Wednesday. I'd say that's both urgent and a priority. We should probably start there."

"You would say that," he mutters.

His library blurs and a screen appears in front of him. He opens the browser, pulls up the e-store and summons the possible presents he'd saved for this year.

"Aria, some help please..."

In the Institute in Copenhagen, where through those years a number of young physicists from various countries came together for discussions, we used, when in trouble, often to comfort ourselves with jokes, among them the old saying of the two kinds of truth. To the one kind belong statements so simple and clear that the opposite assertion obviously could not be defended. The other kind, the so-called "deep truths," are statements in which the opposite also contains deep truth.

Deep truth or not, the consumer-curator-creator trichotomy is something we use to orientate our thinking and assist with decisions related to Subset. However, we're willing to admit that it is a little contrived.

The boundaries between the three types and their default orientations are a lot fuzzier than the simplistic sorting of all people online into one of three buckets assumes. In reality, people slide between these modes at different moments during the day and at different periods during their life. Our advocacy for consumer-curator-creator profiling is based on the sensed presence of an organic setpoint. A position that individuals naturally exhibit and are drawn back to in proportion to their distance from it (and the duration that distance is sustained for). It's said that haters gon' hate; well, creators gon' create, curators gon' curate, and consumers gon' insert appropriate rhyming verb.

Muddying up this trichotomy further is a more recent observation.

Prior to this post, we've described the tendency of curators to maximise surprise:

[Curators] distribute digests of the things that have interested them; they assemble Dunbar goods for the benefit of small groups; they save succinct references to the things they've found and share them with people they know will be interested in them. The perceptions and actions of curators are designed to perturb their surrounding environment and elicit surprise, again and again, and across every domain. They introduce sources of novelty and divergence and interact in ways that catalyse new relationships and insight.

We've elicited the essence of Subset and its ideal user:

These people whom we're building Subset for roam the labyrinthian corridors of the web, following their many interests to many different places. Along the way—again and again in the midst of their wandering—they will make connections between this thing and that person and attempt to form a bridge between the two.

Subset is for those bridge builders who see things and think of people. Its essence is the simple explanation given when something saved is shared with someone who'll care: "I saw this and thought of you."

And we've charted the richness of sharing patterns:

First, most sharing patterns cross platforms. They involve the transition of something from one walled garden to another. From Instagram to WhatsApp; from Facebook to email; from LinkedIn to Discord.

Second, what's shared is limited in depth. Typically, it's a link plus a little contextualising information. Often—if the shared thing is super consumable—it's just a raw link.

Third, the sharing is usually organic rather than systematic and formal. It's driven by emergent feelings and connections and is usually undertaken in a non-self-conscious, informal, peer-to-peer manner. More deliberate, reasoned patterns exist but they're the minority.

Finally, most sharing patterns are just one thread of a larger communication nexus. The sharer sending memes to their friend via WhatsApp is also sending shortform videos via Snapchat and tagging their friend in Instagram comments.

Quite accidentally, we've stumbled upon three distinct sub-modes that people move between as curators, as part of the longer pace layer undulations between consuming, curating and creating. Curators tend to:

1. Act selfishly

2. Invoke others

3. Construct systems

Self-ish curation isn't harmful or malicious. It's a simple noticing of the cognitive cascades triggered when something interesting emerges within one's environment. It's an orientation towards surprise. It's the acknowledgement of glimmers...

Today I learned about a term called a "glimmer" which is the opposite of a trigger. Glimmers are those moments that make you feel joy, happiness, peace, or gratitude. Once you train your brain to be on the lookout for glimmers, these beautiful moments will appear more and more.

...and the capture of some representative artefact in a sociotechnical system—for example, likes, bookmarks or just opening a new tab.

Other-ish curation builds atop this orientation towards glimmers and surprise. It evolves to incorporate someone else to the discovery of glimmers and the experience of surprise. It's a rapidly executed surprise-save-share sequence with a low time-to-save and time-to-share that builds relationships and, ultimately, generates civilisational progress.

As this sequence reoccurs and becomes ingrained within a curator's behavioural patterns, systems emerge to reduce the non-value add friction within these processes and enhance the ability to pipe the surprising things one saves to others for whom they're likely to prove relevant. This is how one ends up producing email digests or link feeds and/or running dedicated tooling and infrastructure; it's systematised curation of varying sophistication.

The consumer-curator-creator trichotomy is kinda contrived. So is the distinction between self-ish, other-ish and system-ish curation. However, like many of the best contrivances, they are grounded in reality. In observable instances of behaviour and interaction. But the most important impact of these useful fictions is that they help us to see differently and to pose new questions. So let us end with one: which curatorial mode—self-ish, other-ish, system-ish—holds the most latent value for our culture-at-large? If they were adequately served and supported, which mode of curation would promote the greatest progress?

As far as we can tell these are valid and reliable differentiators. They are mostly characteristics, however. Characteristics are observable attributes. Things like clothes, gender, possessions, hair colour. Character, on the other hand, is a thing revealed by the choices one makes. Especially in moments of high consequence. So, what is it that distinguishes the character of curators versus consumers and creators? The answer, we think, can be found in a theory called active inference.

Active inference proposes that agents perceive the world and act within it in a proactive way designed to minimise the surprises they encounter. In Active Inference: The Free Energy Principle in Mind, Brain, and Behavior, Thomas Parr, Giovanni Pezzulo and Karl Friston describe it as such:

Active Inference is a normative framework to characterise Bayes-optimal behavior and cognition in living organisms. Its normative character is evinced in the idea that all facets of behavior and cognition in living organisms follow a unique imperative: minimising the surprise of their sensory observations. Surprise has to be interpreted in a technical sense: it measures how much an agent’s current sensory observations differ from its preferred sensory observations—that is, those that preserve its integrity (e.g., for a fish, being in the water). Importantly, minimising surprise is not something that can be done by passively observing the environment: rather, agents must adaptively control their action-perception loops to solicit desired sensory observations.

Think about a typical consumer, scrolling through a social media feed. Now, think about a creator pushing information into that same feed. Both are agents minimising their respective surprise. Parr, Pezzulo and Friston go on to say:

Under Active Inference, perception and action are two complementary ways to fulfill the same imperative: minimisation of free energy. Perception minimises free energy (and surprise) by (Bayesian) belief updating or changing your mind, thus making your beliefs compatible with sensory observations. Instead, action minimises free energy (and surprise) by changing the world to make it more compatible with your beliefs and goals.

Consumers minimise surprise by adapting their information environment to mirror their desired perceptions. Creators, in contrast, minimise surprise by taking actions that either reinforce the state of their reality or shift it closer to the state they wish it would resemble. What about curators? Well, they're built a little differently—to be clear, not better; just different.

They distribute digests of the things that have interested them; they assemble Dunbar goods for the benefit of small groups; they save succinct references to the things they've found and share them with people they know will be interested in them. The perceptions and actions of curators are designed to perturb their surrounding environment and elicit surprise, again and again, and across every domain. They introduce sources of novelty and divergence and interact in ways that catalyse new relationships and insight.

That is the fundamental character of a curator; they choose to maximise surprise.

A team's Fingerspitzengefühl—their intuitive grasp of a problem space or domain—informs the creation of a simple, loveable, complete product. This product deliberately targets a Schwerpunkt—a strategic focal point yielding outsized results in comparison to expended resources—and triggers "aha" moments in users. These moments of user enlightenment catalyse further engagement and adoption, generating market pull signals that validate the existence and/or intensity of an essence and lay the foundation for additional value-adding layers in the future.

So, obvious question: what's the essence of Subset? The answer: we think it's encapsulated in a phrase: "I saw this and thought of you."

It's no secret that our primary users are the nine percent of people online with a distinct attentional profile. Those who engage more deeply than regular consumers but less intensely than creators with an extrinsic motive. Those ostracised by the paradigms and playbooks that refute curation as a default way of being online and seeking novel cultural infrastructure.

These people whom we're building Subset for roam the labyrinthian corridors of the web, following their many interests to many different places. Along the way—again and again in the midst of their wandering—they will make connections between this thing and that person and attempt to form a bridge between the two.

Subset is for those bridge builders who see things and think of people. Its essence is the simple explanation given when something saved is shared with someone who'll care: "I saw this and thought of you."

The general wisdom of our bistable, consume-or-create system with respect to friction is simple: minimise it. Friction slows, impedes, degrades, and obstructs. It is bad. It must be defeated via any means.

For consumers, this means every touchpoint in the user journey should be slick and smooth and never, ever invoke end user thought. For creators, the friction that must be minimised is the friction that is demonstrably non-value-generating, which equates to the majority of creator workflows and processes. However, the friction minimisation playbook doesn't work for curators—that pesky 9% of people on the web with a distinct attentional profile that doesn't quite fit in either of the acceptable boxes offered to them within the current paradigm.

Curators engage more deeply than consumers but not as deeply as creators. To curate is to deliberately interact over time, so the consumer experience of designed-for-non-thought doesn't work. In contrast, curation doesn't map to instrumental objectives in the way that creation does. Creation has, for the most part, a point—an implicit or explicit downstream outcome or higher-order effect that is both desirable and conceivably connected to the activities of creatorhood.

Curators are thus a unique group. Their attention is sharp and bright and deployed for its own sake. To a curator, friction is the coarse material that strikes a match to flame. For them, no friction means no fire. But too much friction means no fire, as well. The right amount of friction for curators is not zero. But it's not much more than zero, either.

Consider what happens when a curator saves a succinct reference to an interesting thing. What gets saved is, broadly:

• Basic info: core identifying elements of the thing

• Metadata: descriptive data about the thing's creation and capture

• Contextualising info: general and user-specific data that adds meaning

• Technical data: system-level info, item properties and constraints

The non-but-not-much-more-than-zero friction tolerated demands that time and attention and energy shouldn't be spent on trivial elements. The metadata of a saved thing (timestamp, source, user, device) and the technical details (provenance, versioning, attributes, constraints) should be de-frictioned and done invisibly. What matters to the curator is the most basic of info—the thing's name and description—and the context surrounding the item:

• Ontology: annotations that describe what the thing fundamentally is

• Semantics: annotations that allocate meaning to the thing

• Relationships: connections to other things (e.g. items, people or projects)

• Actions: intended uses or next steps associated with the thing

Unfortunately, the key piece—the context surrounding things that evoke the interest of curators—is ineffable to systems outside the origin of that interest. Outside the curator themselves. This presents a problem for people and organisations (like us) that take those with curator-like profiles as their primary users. Specifically: one has a non-but-not-much-more-than-zero amount of friction to allocate across a curator's user journey; where should it go?

The simple answer is that it should go to places that help a curator connect to, enliven and diversify the ineffable context of their deliberate interactions with the surrounding world. The more complex answer? Well, that's what we're trying to figure out.

There needs to be structure for curators, but not too much and not delivered in a way that limits possible emergent outcomes—be they unfolding interests or retrospective patterns. There needs to be a meshing of single player elements that interop intelligently with affordances for multiplayer games. Any representation of context should have enough fidelity to encourage affirmative interaction but should also be permissive enough to allow novel stimuli and creative destruction and to prevent collapse towards dogmatic echo chambers.

It is not the easiest of problems, nor is it the hardest but it is, to us, a particularly interesting one that will have significant consequences when solved.

The distraction of looking down to jot down notes or tapping away at the keyboard can make it hard to stay engaged in the conversation, as it forces us to make quick decisions about what details are important, and there’s always the risk of missing important details while trying to capture previous ones. Not to mention, when faced with back-to-back-to-back meetings, the challenge of summarizing and extracting important details from pages of notes is compounding – and when considered at a group level, there is significant individual and group time waste in modern business with these types of administrative overhead.

Faced with these problems on a daily basis, my team – a small tiger team I like to call OCTO (Office of the CTO) – saw an opportunity to use AI to augment our team meetings. They have developed a simple, and straightforward proof of concept for ourselves, that uses AWS services like Lambda, Transcribe, and Bedrock to transcribe and summarize our virtual team meetings. It allows us to gather notes from our meetings, but stay focused on the conversation itself, as the granular details of the discussion are automatically captured (it even creates a list of to-dos).

In its simplest form, Distill takes an audio file, transcribes it, passes it to an AI model for summarisation and returns the output. In more detail (as described by Vogel but itemised here for readability):

• First, we upload an audio file of our meeting to an S3 bucket.

• Then an S3 trigger notifies a Lambda function, which initiates the transcription process.

• An Event Bridge rule is used to automatically invoke a second Lambda function when any Transcribe job beginning with summarizer- has a newly updated status of COMPLETED.

• Once the transcription is complete, this Lambda function takes the transcript and sends it with an instruction prompt to Bedrock to create a summary.

• In our case, we’re using Claude 3 Sonnet for inference, but you can adapt the code to use any model available to you in Bedrock.

• When inference is complete, the summary of our meeting — including high-level takeaways and any to-dos — is stored back in our S3 bucket.

This evoked a speculative capability for Subset we've been considering. Something that eliminates the need to:

• Look up referenced items, mid-communication, during a call

• Send a follow-up containing links to discussed items after a session

It has two variants. The first provides a basic post-meeting recommendation. The second provides live recommendations via real-time audio. The requirements for this speculative Subset capability are straightforward.

1. At least two people in a digital meeting with live audio

2. The meeting is social and peer-to-peer, not the usual white collar meeting theatre

3. Subset has advanced enough to enable automatic querying of the things a person has saved and to facilitate pattern-based sharing with contacts

4. The participants in the call have curated and saved a rich set of items related to the themes of the conversation that is likely to unfold

With all that in place, let's sketch out our speculative capacities in practice.

The first variant is close to what Distill does. It takes a complete audio transcript of a meeting and transcribes it and eventually returns an output. In our case, however, the output is not a summary. Instead, the output is a collection of saved items. It includes any items explicitly referenced during the conversation. It also includes any items from a user's collection of saved things that are meaningfully close to the topics of the discussion. Essentially, the audio is a token stream used to query a collection of saved items for similarly meaningful things.

The second variant builds on the first. But instead of a single capture-query-response-compile-surface loop, there's many. The audio is converted to a query in real-time, as it's being captured. The results are immediately surfaced to the user who leverages them however they please—perhaps sharing them, perhaps using them to inform part of their own dialogue or parse the ideas of another, perhaps marking them for deeper investigation later.

The first variant can probably be done right now by snapping together existing components from current cloud and AI ecosystems. The second variant is a little more far-fetched. The responsiveness requires local storage, AI capabilities and processing, as well as some smart manoeuvring to interop successfully with the current state of different platforms, browsers and devices.

Both are eminently feasible examples of tools for third places that would reduce the toil associated with curation. Let's dub them MARC—an acronym for My Archivist, as well as shorthand for machine-readable cataloguing.

Yes, we have phenomenal infrastructure for querying a globally distributed, zetta-scale repository of interconnected information. Yes, we know how to do efficient search at the smallest scales—locally on devices and within closed networks. But that isn't enough. Something's still up with contemporary search.

The results aren't reliable and consistent and the experience differs based on the domain of the query, as well as the technical sophistication of the user. It could be better—and it should be—but there's no guarantee it will be. Of course, a bunch of people and organisations are trying to secure improvements. Their efforts include:

• Building an "answer engine" instead of a search engine (see Perplexity)

• Swapping search abstractions and UIs for LLM interactions (see SearchGPT)

• Making search an explicit private good (see Kagi)

• Making traditional, ad-subsidised search less icky (see DuckDuckGo)

• Upgrading traditional search with newer technologies (see Exa)

There are other approaches, too—some theoretical and some in-progress:

• Enhancing traditional search with first-party user data

• Advocating for search as a public good

• Inserting humans into the query-result loop in different ways

• Creating domain-specific maps of the cozyweb.

Yet to us, all of the above seem to be slightly off. They all run a playbook involving a massive central repository and/or a privileged third-party that mediates access to a civilisational-scale distributed system and its connected data stores.

So, let's ask a question: "How do you do search from first principles?" As Shane Parrish puts it: "A first principle is a foundational proposition or assumption that stands alone. We cannot deduce first principles from any other proposition or assumption." Decompose search to its core: when people search, they are attempting to find something. That's fair and hard to refute. But here's where we diverge.

Most of the alternatives above make a fundamental error. They assume that the atomic unit of search is the website. And that such a search must be routed through a hyperobject that straddles society. People search the web. This is only partially true.

People are searching the web because the web is a proxy for other people. People search people. The correct atomic unit of search is not the website but a person. The traces of engagement certain people leave as they traverse the web. However, our current paradigm of search is a dual class landscape that excludes these people.

Our existing bistable system only recognises creators and consumers. But there is a third class of people called curators. There's approximately half a billion of them. They engage more deeply with what they find online. They generate Dunbar goods. They catalyse patterns of sharing. They create ripples of progress and connection. And they do it with no expectation of return or consequent asset. Yet they're overlooked, underserved, and invisible to current search regimes.

Search from first principles doesn't necessarily put them first. But it does allocate curators the same equity as creators and consumers in the great civilisational game of information retrieval. And it does leverage their contributions in a way that vastly improves the efficacy of our existing approaches to search.

Every curator saves a subset of the things they find, and they share those found things in a peer-to-peer fashion via patterns. Right now, that saving and that sharing are illegible and immune to search. We're changing that.

The terrain is shifting and new maps are being drawn. Technological capacities are diffusing at frenetic rates. Everyone is scrambling to orientate to the new reality knocking at our doors. Search is a massive piece of this new reality and it requires a fundamentally different approach.

Our question was, "How do you do search from first principles?" The answer: enable people to search people, not the web.

If it comes from a peer, signals an outlay of effort, builds on an existing relationship, aligns with the recipient's interests, respects the recipient's preferences, is easily evaluable, is part of a larger, high quality sequence, and doesn't demand reciprocation? Likelihood of resonance goes up.

If it comes from an unfamiliar source, lacks personal relevance or prior connection, shows no effort, arrives through unwelcome channels, is difficult to digest, follows a pattern of misses, and imposes unwanted obligations on the recipient? Likelihood of relevance goes down.

We explored this in a previous blog that began with three scenarios:

• The reception of an email digest

• Group chat activity orientated around a live motorsport race

• An in-real-life catchup at a cosy cafe

These acts of curation were undertaken by human actors. But what about curation via a machine agent? How does the machinic nature of the curator impact the likelihood that an act of curation will resonate, and the extent of that resonance?

Consider three machine curation scenarios:

• Using Perplexity to research a task outside your domain of expertise

• Scrolling through your X home feed

• Adding your requirements to a comparison site to receive a custom quote

Now apply an analogous litmus test to the acts of human curation. Imagine that, during the act of curation, the machine agent in each of the three scenarios cited can...

• Signal sufficiently accurate emulation of your worldview

• Demonstrate computational expenditure on your behalf

• Propagate context from prior you-machine interactions

• Show relevance to your past, present or future personal interests

• Engage you in a succinct format over a preferred channel

• Cite an established record of previous high quality curation acts

• Avoid any expectation or request for further engagement

Would the machine's act of curation resonate?

This is a rhetorical question; we don't have an answer for you. But we do think that the answer hinges on two things:

• The equitability of one's stance towards machine agents

• The equivalence of one's evaluation of a machine agent's performance

Equitability is simple to determine.

The quick-and-dirty assessment is to examine whether you perceive machine agents as butlers or as centaurs. A butler-ish perception sees machine agents as entities that slavishly accomplish arbitrary tasks. A centaur-ish perception sees machine agents as entities that extend and augment one's capacities. "Doing something for me" versus "allowing me to do even more", as Matt Webb puts it. A longer assessment is to read Hannes Bajohr's On Artificial and Post-Artificial Texts and note whether you end up in gloom-mode or bloom-mode as a result.

Approximately, if you see machine agents as butlers and end up gloomy after Bajohr's exploration of textual origin and provenance, then you'll set a higher bar for an act of curation via a machine agent in comparison to a human one. Whereas, if you see machine agents as centaurs and end up bloomy after the shifting Overton window of textual origins is assayed by Bajohr, then you'll set a similar bar for a machine agent's act of curation versus a human agent's act.

Speaking of bars; equivalence in evaluation is the other determining factor of whether a machine's act of curation resonates.

There's a tendency when evaluating the sophistication of machine agents and non-human intelligences to either escalate the standard once it's been surpassed and/or to advocate for the winning of a different game once the current game has been mastered by a machine agent. This is dubbed the AI effect:

[It] occurs when onlookers discount the behavior of an artificial intelligence program as not "real" intelligence.

The author Pamela McCorduck writes: "It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." Researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'

Equivalence of evaluation, in the context of curation, means judging a curatorial act by a machine agent with the same benchmarks as one would use for a human.

Imagine that, at the same moment, your friend sends you a link to a new EP from your favourite band whilst the new EP is surfaced front and centre of your Spotify home page. Which act of curation evokes the strongest response, and why? Do you see your friend's understanding of your taste in music as equitable to Spotify's personalised approximation of your taste? And do you judge their efficacy and impact in an equivalent manner?

Let us be clear. This is not to say that:

• Human and machine agents are equal in status, capacity, required rights etc.

• Human and machine agents can, will or should be judged by shared benchmarks

This is not to assert the priority of one agent type over another. But to is to say that where one sits on the equitability and equivalence spectra will fundamentally influence how one responds as a recipient of machine curation. And that matters because we're on the precipice of an even more machinic era of curation. Investigating, now and in advance, where one stands will help us to mitigate some of the inevitable risks of such a transition.

This is not a new problem, by any means, but we have been thinking about models and frameworks and concepts to help us work through it. Below is one such thing.

Traditional trust models lean on a centralised party to either:

• Supply a network from which it is possible to infer node-to-node connectivity and allocate trust and capabilities accordingly (e.g. LinkedIn's 1st, 2nd, 3rd degree connections, Facebook's friends of friends)

• Act as an entity that assesses and authorises the claimed fidelity of nodes in the network (e.g. enterprise authentication solutions) based on some provided evidence

Subset, like other newer initiatives, can't lean on those prior methods because incumbent platforms are deemed at best indifferent and at worst hostile to the new initiative's mandated mission. So, what's the alternative?

One we've been thinking about is focused on the volume of channels that two parties share. A channel is simply a means for one user to send or receive a message from another. It's a transient rail for information exchange. A channel instance is counted when both users assert a positive identification of the other in that channel within some constrained window.

For example, you have a friend that you communicate with in multiple ways. You send WhatsApps, exchange emails and trade direct messages on X. You also have an an acquaintance who you only message on LinkedIn. You are connected to your friend via three channels and your acquaintance via one. There are also people of whom you are aware but with whom you do not communicate over any channel.

The trust and capabilities allocated to your friend thus exceed those allocated to your acquaintance which in turn exceed those of the people with whom you don't communicate. The allocation of trust and capabilities within such a network looks like this:

• Adversarial: hostile sentiment between two users is detected—no trust or capabilities allocated

• None: no channels between two users are established—minimal trust or capabilities allocated

• One: one channel between two users is established—some trust or capabilities allocated

• Many: more than one channel between two users is established—full trust or capabilities allocated

For Subset, this has two primary applications: sharing and search.

One of Subset's core capacities is humane routing—the ability to share something with someone via patterns. Rating within the ANOM model sketched above could determine the power dynamics of the sharing activity. For example, a sender that has just a single trusted channel with a recipient is rate-limited as to the volume and tempo of their sharing with that recipient, whereas a sender-recipient pair that share many channels may be able to override each other's preferences in defined situations.

It could also inform the results retrieved from a P2P network search query. If a searcher has many trusted channels with a peer, the search query could provide higher fidelity results from that peer's pool of relevant results. If they have no trusted channels with a peer then the results would possess more inherent obscurity.

The literal mechanics of something like this are still a work in progress but they're likely to take cues from existing prior art—such as public key cryptography—as well as the newer wave of cryptographic techniques and infrastructure. But one thing is clear: outsourcing user-user trust and verification to incumbent platforms and/or misaligned enterprise vendors is no longer the only option.

Previously, we've conceptualised its use as information bending—the intent-based annotation and distribution of found things—and as an ability to leverage custom information couriers—entities that distribute saved things to high relevance targets in a trustworthy, autonomous, performant way. Now it's time for a third metaphor: routers.

A router is a device that directs the flow of information packets across networks and between devices. It examines inbound traffic, analyses its characteristics and destination, and propels it forward in a way that satisfies the constraints of the sender, the recipient and the network infrastructure being utilised without compromising the integrity of the transmitted message.

As an information and communication networking technology, routers have been operational for over half a century. As an art and science, the routing of information is a practice as old as civilisation itself. We've progressed from the bipedal messengers of antiquity—mythologised by Hermes—to post riders and right on through to the modern approaches, technologies and concepts of information routing.

Machine agents conducting global information symphonies at incomprehensible scales and tempos is an accepted part of modern reality. Unfortunately, so is the dysfunctional delivery of saved things to specific destinations. Consider this scenario: you're online—somewhere, somehow. You see something and think of someone. Now, how do you get it to them?

Typically, the answer involves copy-pasting, cross-platform context-switching, and excessive cognitive friction. Even if you're sophisticated and use a dedicated app for saving and storing found things or use sharing functions integrated to your device's OS, the trafficking of thing X to person Y will be slow, sketchy and not something that fills you with joy.

In comparison to modern routing, it's like keying in a number to a rotary phone, waiting to be connected to an operator, requesting a specific person to speak to, tapping your foot impatiently while they come to the phone, and finally getting them in conversation to deliver a short message. A message which is similar to one you'll probably want to get to them in a couple days time and to one you shared with them last week, too.

It doesn't have to be like this. Just set up a simple pattern— when I engage with something about basic income send it to Joanne—and be an interested human on the internet who finds cool stuff. Then, when something meeting the criteria from your simple pattern is found, produce an output and route it to Joanne. This is humane routing driven by resonant curation. This is one tiny sliver of what Subset enables.

We are all subsisting amongst endless streams of information; we're already managing the flow of packets around us. We're just doing it in a way that is inefficient and detrimental to our relationships. Now it's time to route that flow in a humane way.

In On Exactitude in Science, Jorge Luis Borges describes the absurdity of perfect, one-for-one, representation. The full fiction:

…In that Empire, the Art of Cartography attained such Perfection that the map of a single Province occupied the entirety of a City, and the map of the Empire, the entirety of a Province. In time, those Unconscionable Maps no longer satisfied, and the Cartographers Guilds struck a Map of the Empire whose size was that of the Empire, and which coincided point for point with it. The following Generations, who were not so fond of the Study of Cartography as their Forebears had been, saw that that vast Map was Useless, and not without some Pitilessness was it, that they delivered it up to the Inclemencies of Sun and Winters. In the Deserts of the West, still today, there are Tattered Ruins of that Map, inhabited by Animals and Beggars; in all the Land there is no other Relic of the Disciplines of Geography.

Imagine trying to share such an infinitely precise map; it wouldn't work. What does work: saving succinct references to actual things. The consequence? We promiscuously share and search. But what happens when we save something we discover online today? What exactly is it that gets saved?

This is an important question for us—saving is the first part of the save-share-search loop that we're in the process of modernising. So we've been formulating an answer. What gets saved can be broken down into four parent categories:

• Basic info: core identifying elements of the thing

• Metadata: descriptive data about the thing's creation and capture

• Contextualising info: general and user-specific data that adds meaning

• Technical data: system-level info, item properties and constraints

Within each of these four parent categories are sub-categories of information that make up the body of a succinct reference.

Within the basic info category, the following is saved:

• URL / CID: a mostly machine readable identifier for the thing

• Title / name: a short string of natural language

• Description: a brief summary of the thing and/or its purpose

• Language: the primary language in which the thing is written or presented

Within the metadata category, the following is saved:

• Timestamp: the datetime of when the thing was saved

• Source: where the thing was found (e.g. a website or app)

• User: the entity or account that saved the thing

• Device: the hardware or platform used to save the thing

Within the contextualising info category, the following is saved:

• Ontology: annotations that describe what the thing fundamentally is

• Semantics: annotations that allocate meaning to the thing

• Relationships: connections to other things (e.g. items, people or projects)

• Actions: intended uses or next steps associated with the thing

Within the technical data category, the following is saved:

• Provenance: the thing's origin and history

• Versioning: the current state plus past usage and modifications made

• Attributes: observable specifications and properties derived via analysis

• Constraints: limitations governing the thing's use or access

Enumerated like this, it seems as if a lot of information gets captured when we save a thing. But it isn't that much, really:

• An arbitrary, multi-thousand word essay on the web—ignoring image content, formatting, or any of the enveloping elements of the browser that deliver it—equates to 10-20,000 characters, or approximately 10-20 KB of text

• A relatively minimal set of information about a saved thing using the schema above, in contrast, is going to equate to 1-2,000 characters, or 1-2 KB of text

Of course, both the essay and the saved info themselves contain succinct references to other things (e.g. prior ideas for the essay, compliance frameworks for the saved thing data). They're part of a greater, deeper, civilisation-wide web. But I suspect you see the point.

We save succinct references to found things, and the elements of those succinct references are what enables us to share things with our peers and search across distributed networks for both new and old things.

When sharing, we can quickly identify relevant content using the basic info and metadata, while the contextualising info provides rich background for why the content matters. For instance, when sharing a research paper, a user can easily explain its relevance using ontology and semantics and note any intended actions.

When searching, these elements enable powerful, context-aware queries. Users can search not just by keywords, but by concepts (using ontology and semantics), intended use (via stated actions), or technical specifications. This allows for precise retrieval of saved items, such as finding a specific algorithm based on its conceptual relevance, intended application, or implementation characteristics.

Ultimately, the saving of succinct references transforms sharing and searching from simple content-matching tasks to rich, context-driven processes. Yet, despite its fundamental role in our digital lives, something as simple as saving is not a solved problem. There's little consensus on what gets saved, how it's stored, or the actions that are available to different parties downstream.

We're aiming to change this. Asking "What gets saved?" is just a small step in a longer journey. One where more contacts that leave traces are saved and where more things that are saved are shareable and searchable.

Scenario one: AM peace. It's Friday morning. The beginning of the end of a hellish, chaotic work week. The kids are off to school and you have a half an hour of solitude. Sat on the sofa, you pull out your phone, open your emails and browse the latest issue of your favourite email digest.

Scenario two: motorsport live chat It's a light and fresh midsummer afternoon. In a couple hours, you're meeting friends for a barbecue. But right now, you're all watching the race. The WhatsApp group chat is ding-ding-dinging. One of your friends shares a link to a segment of alternative commentary narrating the crash that just happened and analysing the driver's manoeuvres that led up to it.

Scenario three: coffee and a catchup at a cosy cafe. It's winter. It's wet. It's cold. It's just not nice. But you're inside, warm and content, sat opposite a colleague-turned-friend in a local cafe you've recently taken a liking too. You pause as your friend reaches down to the bag at their feet. They place a CD on the table and slides it towards you.

These are all acts of curation. But what makes them resonate? What is it about such deliberate sharing that makes it work? We have some tentative answers...

• It comes from a peer: a "peer" is someone whose worldview has a non-trivial overlap with your own. That could be a friend or family member, it could be someone with whom you have weak or strong ties, or it could be someone you don't actually know but whose perspective is trusted and assumed to be aligned with your own.

• It's downstream of a prior relationship or engagement: the likelihood for resonance increases exponentially when it comes in the wake of a previous engagement and/or when it connects to the context of an existing relationship.

• It's relevant to past, present or potential future interests: the act of curation aligns explicitly with the interests of the recipient in some way—be it something from the past, because of a current priority, or due to a future situation.

• There's proof of work: bundled within the act of curation is some signal that implies an expenditure of energy, effort or expense above a certain threshold. There's proof that the act involved deliberation and care.

• It's received in a preferred manner: the act of curation is received via a channel and in a format that the recipient is comfortable and competent with.

• It's succinct and easy to rapidly evaluate: the output of the act of curation is short, easy to parse, and includes enough information to make the call between engaging deeper and archiving or not engaging at all.

• It's part of a majority-signal sequence: not every act of curation can be one hundred percent effective but the tendency of the sequence overall has been to be more signal than spam.

• There's no request for or expectation of response: the act of curation is asymmetrical—no burden is placed, implicitly or explicitly, upon the recipient to acknowledge receipt, engage deeply and reply, or reciprocate with an equivalent act.

If all of the above are present, then it's a near certainty that an act of curation will resonate with the recipient. If we invert the elements above, we can also understand when curation will not resonate:

• It comes from an unknown / someone without perspective-overlap

• There's no prior connection or relation with the curator

• It's not relevant to your past, present or future interests

• There's no indication that effort was expended

• It's received via a channel and in a format you dislike / don't use

• It's long and hard to comprehend

• It's part of a sequence of curation acts that haven't landed

• It demands an acknowledgement, response or reply

Think back to the scenarios from the beginning of this post. The pre-work peace, the motorsport chat amongst friends, the cafe meetup. In each case, the act of curation will either resonate or it won't.

What makes it most likely to resonate is that it comes from a peer, signals an outlay of effort, builds on an existing relationship, aligns with the recipient's interests, respects the recipient's preferences, is easily evaluable, is part of a larger, high quality sequence, and doesn't demand reciprocation.

What makes it likely not to resonate is the inverse: it comes from an unfamiliar source, lacks personal relevance or prior connection, shows no effort, arrives through unwelcome channels, is difficult to digest, follows a pattern of misses, and imposes unwanted obligations on the recipient.

Do you agree? Are these criteria the difference between an act of curation hitting or missing? We think so. But the best way for you to find out is to try.

Engage a little more deeply with the world and share what sticks. See things; think of people; share the things with those people. Repeat for time and enjoy the resultant transformation in your life.

If "everything is content" then one's value is coupled to creation. To one's ability to generate and package up remarkable moments. Unfortunately, most people neither want to nor are able to create at a high level and sustainably over time. Thus, we're compelled to think: a minority creates and is superior whilst the majority consumes and is worth less.

For most, this bistable, consume-or-create paradigm is okay. For others—a good 10% of people online—it's not. They want a new default. They want to opt out of this particular game. So they look for an alternative. An alternative best described by Bo Burnham:

They say it’s like the ‘me’ generation. It’s not. The arrogance is taught, or it was cultivated. It’s self-conscious. That’s what it is. It’s conscious of self. Social media - it’s just the market’s answer to a generation that demanded to perform so the market said, 'here - perform. Perform everything to each other, all the time for no reason.' It’s prison - its horrific. It’s performer and audience melded together. What do we want more than to lie in our bed at the end of the day and just watch our life as a satisfied audience member. I know very little about anything. But what I do know is that if you can live your life without an audience, you should do it.

This segment of people online that don't buy in to the "consume-or-create" choice attempt to "live life without an audience". But that's difficult when tools for third places don't really exist and when every act in the digital sphere is architected to further pressurise the gradient between consumption and creation.

Even enlightened consumption is subtly appropriated by the consume-or-create paradigm. Most tools-for-thought, most indie publishing apps, most personal knowledge management systems, most curation-focused products make an assumption: surely, ultimately, the aim must be to ascend to creatorhood. No one would be foolish enough to expend time and energy and tangible resources saving, sharing and searching without an eventual endgame that involved the conversion of that expenditure into some form of audience from which value can be siphoned.

In the current paradigm, if you're not assembling an audience then you're accepting a role as a part of one. Or, more likely, as one piece of many different audiences. Right now, "to live your life without an audience" is still to accede the existence of audiences as supra-organising features in the digital landscape. In the coming paradigm, however, that acceptance is being challenged. An anti-audience stance is emerging.

To be anti-audience is to deny that audiences are a first principle of digital culture at all. To be anti-audience is to refute the idea that the monopolisation of the many's attention by the few is a precondition of "winning". It is to acknowledge an organic upper bound on one's reach and influence. It is to hold to the conviction that one can exhibit curiosity and engage more deeply without an accompanying pressure to build an asset.

There's nothing wrong with audiences—either building them or being amongst them. They are a core part of social dynamics, after all. Yet their prevalence and salience in our culture has intensified to the point where everything is audience-orientated, and where this orientation is pushed as an immutable law of contemporary being. It isn't. Audiences, as we know them, are novel social constructs, not fundamental elements of existence.

To be anti-audience is not to refute their impact but to put the concept of an audience back in its place—which is somewhere less than the most important thing in digital culture.

An email digest—like Matt's Magnificent Seven—represented as a simple pattern looks like this: send seven things from across to web to all subscribers via email every Sunday at 0700. Additional patterns can be constructed using combinations of sources, tempos, frequencies, content characteristics, distribution channels, and contact segments. This makes for a massive space of possible sharing patterns for every individual, group and organisation.

Below, we've presented a small selection and identified some of the higher level themes we've noticed whilst exploring and facilitating sharing patterns. The samples are drawn from our own sharing practices, as well as those we've discovered from others—be they friends, family, peers, colleagues, early Subset users, or participants in our product discovery initiatives. We've generalised and anonymised them to make them a little easier to comprehend. So let's check them out...

1. Share ten things re: the creative life via a weekly email to a niche audience

2. Post generative AI long reads to a personal LinkedIn twice a week

3. Drop golf memes and highlights from Instagram into a WhatsApp group chat

4. Share grappling techniques and seminars from YouTube to a Warpcast channel

5. Continually output health and wellness links as X posts

6. Send potential places or activities to a spouse via email in advance of a trip

7. Notify a friend of studies that provide evidence for a mutually-held belief

8. Post themed essays to a specific channel in a company Slack

9. Follow up a chat with a one-off Discord DM of products that were referenced

10. Bounce home renovation TikToks to one another from within the app

11. Send food pics and good local restaurants to a Facebook Messenger group chat

What did you notice about the above? What commonalities come to the fore when you think about sharing patterns? We've found four.

First, most sharing patterns cross platforms. They involve the transition of something from one walled garden to another. From Instagram to WhatsApp; from Facebook to email; from LinkedIn to Discord.

Second, what's shared is limited in depth. Typically, it's a link plus a little contextualising information. Often—if the shared thing is super consumable—it's just a raw link.

Third, the sharing is usually organic rather than systematic and formal. It's driven by emergent feelings and connections and is usually undertaken in a non-self-conscious, informal, peer-to-peer manner. More deliberate, reasoned patterns exist but they're the minority.

Finally, most sharing patterns are just one thread of a larger communication nexus. The sharer sending memes to their friend via WhatsApp is also sending shortform videos via Snapchat and tagging their friend in Instagram comments.

Sharing is a foundational behaviour in modern culture, with hidden patterns and subtle dynamics that are only now revealing themselves after a generation of being online. Contemporary sharing tends to be cross-platform, depth-limited, informal and P2P in nature, and multi-threaded.

There are likely other fundamental elements of sharing still waiting to be discovered and leveraged. And as we work to bring the existing cultural technology stack up to speed, we intend to lay the foundations for those discoveries. To enable faster, flexible, more sophisticated patterns of sharing without the excess of energy and expense that's currently required

In the meantime, consider your own sharing patterns: what do you share, with whom, and why? How do you go from, "Hmmm, interesting," to, "Hey, I saw this and thought of you"?

Consider the main desktop operating systems: Windows, Mac and Linux. Think about the dominant mobile OSs: Android and iOS. And ponder some niche operating systems: QubesOS, ethOS, Playbit. The end user capabilities common to all these operating systems?

• User interfaces for interacting with the system

• Creating, copying, moving and deleting files and folders

• Installing and managing applications

• Connecting to and browsing the web

• Engaging with audiovisual media

• Editing text and documents

• Managing system settings and customising preferences

• Configuring user accounts

• Backing up data and enabling restoration and recovery

• Searching user data or apps (locally or via cloud)

• Supporting peripheral devices

• Providing and upkeeping basic security mechanisms

• Updating software and maintaining the system

Notice how saving, sharing and search features are either entirely absent, fundamentally constrained, or only available via third-party applications.

As a class, curators (those for whom saving, sharing and search is a default behaviour) make up approximately 10% of these OSs user base. That's significant but not enough for OS providers to optimise for. Especially when the curator class's main products are Dunbar goods, when the curators themselves eschew both the intensive production labour and the neurotic consumption behaviours that have proven so profitable, and when their alternative form of third-place-focused activity is so resistant to traditional monetisation strategies.

There are good reasons for this, but we're reaching the point where the stance is approaching stupid and reckless. "Stupid" because it is preventing the unlock of a tonne of latent value. "Reckless" because the higher order effects are ossifying cultures at scale.

To see what we mean, consider the following taxonomy of things. These are all types of things that most of us have, at one point or other, wanted to save, share or search for:

• text, image, audio, video, game, news, meme, document, person, organisation, community, location, project, product, service, software, object, event, opportunity, activity, protocol

Now pick one of them and imagine what it takes to:

• Save it somewhere/how on one of your devices

• Share it with someone whom would appreciate or benefit from it

• Search for it at some point in the future

Chances are that your imagining of those processes involved a non-trivial amount of manual effort and a chaining of multiple applications. Now consider the effect, the impact of that saving, that sharing, that searching. Chances are you can envision the upside of a store of things you've found interesting, of sharing deliberately with specific people, groups and communities, of being able to rapidly surface interesting things you and your peers have saved and shared.

That upside is what's impeded when saving, sharing and search are not "baked in" to the operating systems of our modern devices. The alternative—the thing we're pushing for—is to convert saving, sharing and search into core, foundational capabilities that are accessible to all. Saving should be simple. Sharing should be protocolised and pattern-based. Searching should be specific to the end user's context. Saving, sharing and search should, by now, be device OS primitives.

They're not because it's not in the interests of the incumbents to make it so. The bistable consume-or-create system that is the status quo is just too comfy, and elevating curators to first-class citizens and making Dunbar goods accessible is just too risky a gambit.

Better to keep saving, sharing and search full of friction. Better to impede the flow of artefacts, lest they flow undesirably and in a way which cannot be maximally siphoned. Better to make core behaviours that many would embody, given the affordance, seem like endeavours of great effort and expense when in fact they are some of the most natural, organic activities of all.

• A reference to the thing

• Information concerning what the thing is

• Insight into what the thing means to the saver

• Some metadata surrounding the thing and its saving

Unsurprisingly, better saving makes for better sharing and search.

Consider two people that have saved a song they like. One has a functioning reference to a primary listing of the track, has annotated it as a jazz audio recording by a specific band, has labelled it as a "song to cry to", and has captured the datetime of the capture. The other has saved a raw link to a social media post that contains a snippet of the same recording posted by one of the band's unofficial fan groups.

The former is much more likely to share the song appropriately with others—perhaps using a group chat with friends focused on music, or to a local community for music events. The latter will likely either never share the saved thing, or coarsely broadcast it to their main social media feed. The former is much more likely to be able to find the saved thing in the future because they can use a range of hooks to fish it out—from the metadata to the song genre or the representation of meaning it evoked. The latter is unlikely to be able to find it in the future—and if they do it'll based on a combination of remembered vibes plus unstructured scrolling and arbitrary keyword searches on what may or may not be the right platform.

This dynamic plays out with all types of things, from the songs we save to the people we meet, from the places we visit to the various textual forms we interact with online. And the trend in recent years has been, in response to this dearth of information available online, for more people to use and more systems to facilitate good saving practices. Two of the biggest enablers of this trend have been tagging and zooming.

Tagging is the assignment of keywords or labels to things in order to improve their accessibility, organisation and management. It's distinct from classification—which usually focuses on the literal ontology of the thing, the essence of what it is—because tagging is tailored to the context of the end user. Netflix's home screen will present you with rows of thematically grouped shows. This is tagging in action. Netflix's search functions, in contrast, will enumerate the types of shows available—movies, TV series, documentaries, et cetera..

Zooming is a complementary function that controls increases and decreases in the resolution with which something appears. The progression from a show's poster art to a short description to a trailer to the first episode to an entire series is a progressive zoom from low resolution to high. An alternative example: going from a book's cover to its blurb to its contents and front matter to chapter skimming to the actual prose. Another example: moving from a continent to a country to a region to a town to a street to an address and back again in Google Maps.

This is all well and good in theory. In practice, the current state of tagging and zooming is far from optimal.

First, configuring systems that provide high quality tagging and zooming experiences is resource-intensive for providers. The bulk of the work is done upfront and then rolled out to end users, often to mixed receptions and limited effect.

Second, it's labour-intensive for end users. If they use what's provided to them, it means an excessive amount of cognitive load to navigate an essentially impersonal, foreign tagging taxonomy and operate clunky zooming mechanics. If they decide to roll their own system and processes, it requires janky coordination of numerous third party tools and a constant vigilance which eventually overwhelms and causes abandonment.

Fortunately, the coming state of tagging and zooming is a big upgrade. It combines:

• Local-first applications

• On-device machine learning and AI

• Retrieval augmented generation architectures

• Continual interaction with past and present end user context

The result is a fast and tailored tagging and zooming experience. And the two things driving it: RAG-tagging and semantic zooming.

RAG-tagging is retrieval augmented generation plus traditional tagging. It uses large language models to generate, shape and refine the tags allocated to saved items based on the both the user's historical context and the requirements of their interactions in the present moment, without recourse to a distant, centralised third party service and global data store.

Traditional tagging is, for the most part, a manual, one-and-done exercise. Something's tagged at point of capture and a monolothic, fragmented schema of tags emerges over time. RAG-tagging, in contrast, continually regenerates the tags applied to items in a schema that accounts for historical tagging and the needs of the present moment.

Semantic zooming (sem-zooming) complements RAG-tagging. Humans have a supposed memory span of seven things (plus or minus two). Sem-zooming is the reformulation of meaning based on a combination of the resolution of the items that are viewed and a human's inherent memory span.

Imagine you have a list of 10,000 items saved. At low resolutions, seeing thousands of pieces of distinct information isn't useful at all. But as you look for a specific type of thing, and move from lower to higher resolutions, from 10,000 things to 1,000 things to 1 thing, the meaning attached to those different view points needs to dynamically adjust. With sem-zooming, the meaning shifts alongside the resolution.

One domain to see this zooming in practice is network visualisation. Standard systems will provide like-for-like displays of networks and their properties. But the best systems will shift the meaning that can be inferred from the visualisation as one zooms in and out, from node to cluster to clusters to the full network.

Better saving practices means better sharing experiences and greater search outcomes. The state of the web right now is compelling innovations in sharing and search, and these activities are getting a lot of attention. But for many people in many domains, good saving is the foundation that good sharing and search relies upon. And it's time for us to appreciate that and innovate on how we save and manage the things we find online.

RAG-tagging and sem-zooming are just two pieces—important ones, nonetheless—of this. The traditional approaches to tagging and zooming don't work for end users that expect both speed of interaction and maximal personal relevance, that demand complex capabilities with simple interfaces and intuitive mediating abstractions. Meanwhile, their utility for providers is also declining—more resource in exchange for increasingly marginal gains in effectiveness and UX isn't the best bargain and indicates a saturation point for current tagging and zooming methodologies.

RAG-tagging and sem-zooming may not be the entire answer but they're a fundamental part of the new and emerging practices for how we save, share and search online. And we're in the process of figuring them out.

Nikita Bier has a simple benchmark for such a devilishly tough question:

If you can't explain your app in fewer than 5 words, do you think your users will ever bother telling anyone about the app?

Before you have product-market fit, product development is not about "adding features"—but distilling a product to a simple marketable sentence. Adding features more often impairs growth.

We're on our way to meeting that standard, but we're not there yet. We haven't yet compressed Subset down to its densest, most communicable kernel. But we have done a little bit of street fighting mathematics and learned, when faced with tough problems, to relax the constraints in the name of progress. So we've answered a weaker form of the question above:

• What type of thing is Subset?

Here's the negative answer, the type of thing Subset is not: a bookmarking tool, a read-later app, a note-taking or personal knowledge management system, an AI summariser, a task management widget, a content discovery platform, a sales and marketing suite, an enterprise automation solution, a news aggregator, an old school or new wave social media, an augment for creators, a community and collaboration entity.

Now here's the positive form, the type of thing Subset is:

• A tool for third places

The "tool for" prefix comes from "tools for thought". As Maggie Appleton describes them:

[Their core function] appears to involve writing notes, connecting them to one another, exploring dynamic views, and then experiencing a kind of emergent wisdom. An enchanting promise.

"Third places" are environments distinct from home ("first place") and work ("second place"). Allie Conti says:

The term, which was coined by the sociologist Ray Oldenburg in the 1980s, essentially refers to a physical location other than work or home where there’s little to no financial barrier to entry and where conversation is the primary activity. The historical examples that Oldenburg cites in his book The Great Good Place include French cafés, German American beer gardens, and English pubs, all of which appeal to people from various walks of life.

Combining the two, we can define "tools for third places" as things that deepen and enrich both synchronous and asynchronous peer-to-peer interactions. Their core operations are the saving, sharing and searching of things we find, for a whole range of reasons, most interesting and arresting.

Here's an example: you're at a cafe, meeting a friend. Imagine the conversation you have. It will probably require the plucking of a phone from a pocket or purse in order to reference or find something. It will likely include an urge to pull out a magic glass rectangle and summon a sliver of the world's information to flavour the unfolding dialogue.

That's the norm. Relationships (new or old) are sequences of interactions that reference other things, from experiences and stories to places and products. This is mostly verbal but it often escalates to the exchange of digital representations of discussed or discussion-adjacent things.

The technology we use makes this possible and adds some value, but it also adds unnecessary friction, interrupts flows and inputs irrelevance to an interaction. Even the liveliest yarn is murdered by a fumbling search for that picture, by an increasingly despairing mini-quest to find the name of that place, space, person or thing, and the dead time that accompanies it.

Not for much longer. Tools for third places, such as Subset, are coming. They'll eliminate friction, deepen connection and provide richer context in these scenarios. And our relationships and our cultures will be better for it.

We've since realised that an important piece was missing. What we didn't do was articulate why completing this mission and delivering a new way actually matters. Below, we correct that error and provide a short overview of the little why and the big why, respectively: relationships and progress.

Relationships are a fundamental building block of our lives, and they're sustained by sequences of interactions that often reference other things, from experiences and stories to places and products. Nowadays, these things are represented digitally, and how we save, share and search amongst them is far from optimal given our current technological capabilities. Our current saving, sharing and search operations are:

• Cumbersome, requiring a context-switch to a distinct state of mind

• Low-relevance, likely to be unaligned with the context of an interaction

• Slow, often interrupting the flow and pace of an interaction

The new way we're working on changes these operations and their impact. It makes them:

• Frictionless, preserving the state of mind occupied during an interaction

• Contextualised, likely to enrich the context of an interaction

• Fast, smoothing the flow and pace of an interaction

Imagine sharing a relevant article with a friend in seconds, rather than taking minutes to search, assess and copy an extract and a link. Imagine recalling a key piece of information from a past conversation without losing the thread of a current discussion. These improvements have profound effects on how we connect and communicate with one another.

Ultimately, better saving, sharing and search means better relationships.

Better saving, sharing and search operations also mean a greater rate of civilisational progress.

There are many factors that contribute to progress, of course: these range from institutional structures and cultural heterogeneity to the rates of construction, maintenance, development and decay of infrastructure. Another big progress indicator is capital creation and allocation: the ability to alchemise labour and technology into liquid capital and get it to the right people, in the right place, at the right time, in the right way.

Fortunately—or unfortunately, depending on one's perspective—many civilisational progress indicators lag information allocation. They dance to the tune of the underlying information symphony. Best case: the right information is generated and it's delivered to the right people, in the right place, at the right time, in the right way.

We don't live in the best case world. Current saving, sharing and search operations have got us closer to it but not as close as we should and could be. We'd like to fix that and, as a higher order effect, unlock several more units of civilisational progress.

"A new way to save, share and search" is about more than mere technological innovation. It's about the base fabric of our relationships and the texture of the emerging reality that unfolds from them over time. Like the civilisational progress indicators that lag behind information allocation practices, our current approaches to saving, sharing and search are no match for the breakneck pace our contemporary society and its environment requires.

We're changing that. We're making saving simple, sharing effortless, and search joyous, and doing it in a way that scales. Why? To enhance relationships and accelerate progress.