The loop has a specific shape. People building productivity tools for the people building productivity tools. Tool marketplaces for the builders of tool marketplaces. Optimization layers optimizing the layers that optimize them. Lately it seems that every feed I open serves a refraction of the same image: AI tools used to make AI tools for the people making AI tools, the dead internet thesis articulated inside the dead internet, the criticism of the loop performed in the loop's own register, on the loop's own platforms, calibrated to the loop's engagement metrics.

The internal coherence of this is what makes it hard to see. When the tool, the user, the producer, and the critic all sit inside the same frame, every position from which the frame might be visible is itself inside the frame. The conversation registers as loud, the participants as serious, the activity as significant. This is a decidedly closed loop.

The loop produces its own artifacts of self-reflection, and the artifacts remain inside the loop because the loop is what produced them.

The most articulate voices in the field describe this condition while standing inside it. In December 2025, Andrej Karpathy posted a thread on X that received fourteen million views. He listed the new programmable layer he was failing to keep up with: "agents, subagents, their prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations." Two months later he was naming his new posture "agentic engineering" and describing himself as having gone from writing eighty percent of his code to delegating eighty percent of it. By March he was saying the bottleneck on his own productivity was himself. The clearest voice in the field was, in real time, narrating his disappearance into the loop's logic.

The slop factory's primary product is the discourse about its own existence.

Where the conversation actually lives

The field thinks it is debating model capabilities, agent orchestration, prompt engineering, and the speed of productivity gains. It is also debating, more loudly, whether Dario Amodei is right that AI will eliminate remedial engineering labor. In his January 2026 essay The Adolescence of Technology, Amodei restates the prediction directly: AI could "displace half of all entry-level white collar jobs in the next 1–5 years," even as economic growth accelerates. He projects sustained annual GDP growth in the 10–20% range as plausible.

On its face, the prediction is arresting. Read in context, it is also revealing. The most prominent voice from inside the AI frontier labs, sounding the most prominent alarm about AI's effects, can articulate his optimistic case as a productivity-shaped ceiling. The post-AGI imagination available to the loop reaches "more output, fewer humans" and stops there.

This is the altitude of the visible debate. The Overton window of contemporary AI discourse sits several clicks below where the field's deepest problems live. The visionary horizons people see from inside the loop, the fully automated workflow, the autonomous agent stack, the AGI labor forecast, are functions of the loop's compression. They appear visionary because the frame is small. From a higher altitude, they are local maxima of an exhausted optimization.

To see the actual horizon, the entire window has to shift up. The conversation's center frame has to relocate three clicks above where it currently is. Once that move happens, the things presented as horizons within the current frame become the new periphery. The field's current visionary stance becomes a footnote. The new center becomes the question of what the work after productivity actually is, who is doing it, and how the value of that work circulates.

The synthesis confusion

The work after productivity rests on a different operation than the work the loop performs.

The current AI-product paradigm treats synthesis as extraction at speed. Take a large quantity of information, distill it down, hand the user the answer, free the user to proceed to the next thing that needs distilling. This is the operation every frontier lab is presently optimizing for. Longer context windows. Memory features. Multi-agent workflows. Reasoning models. The shape of the work is the same in each case: ingestion, compression, output, with the speed of the compression as the figure of merit.

This is one kind of synthesis. There is another.

The other kind treats integration as cultivation across time, multiplicity, disagreement, and slow accretion. It preserves the multiple. It holds disagreement open. It surfaces the convoluted substrate from which an answer might eventually emerge, attends to the contributors who hold their disagreement open against demand, and produces an artifact whose value is its irreducibility. The first kind produces information at scale and is what the entire industry is presently selling. The second produces wisdom over time and constitutes the open category I'm presently focused on.

MIT's Project NANDA reported in July 2025 that ninety-five percent of enterprise AI spending produced no measurable return. Roughly thirty to forty billion dollars of investment, no clear results. The usual readings of this number describe an execution problem, an integration problem, or a "the models need more time" problem. The failure is categorical. The enterprises bought the first kind of synthesis and discovered they needed the second.

The expectation that AI tools will take a large amount of information, immediately identify the opportunities within it, and hand back a clean recommendation is the expectation that synthesis can be performed as extraction at speed. For tasks where the right answer is structurally compressible, retrieval, transcription, first drafts, the bookends of routine knowledge work, extractive synthesis is real and valuable. For decisions where what matters is the relationship among many partial perspectives, the integration of contradiction into something the contradiction does not destroy, or the slow recognition of a pattern the participants themselves did not yet know they were producing, extractive synthesis returns confident smoothness. The smoothness is the failure.

Wisdom is what remains when the synthesis preserves what the speed would have flattened. It is harder to consume. It requires recalibration of the reader's psychology. The work of reading it is the work of recognizing the substrate.

The aperture

Recognizing the substrate requires an aperture wide enough to admit it.

Most product-market-fit thinking solves a problem the user already knows they have. The user knows the problem. The user is searching for the solution. The product anticipates the user's search. The user encounters the product, recognizes themselves in it, and pays for the solution because the solution speaks their problem back to them. This is a coherent form of product work. Its entire success condition is that the user already had the right frame.

Category creation does something else. A new category opens the user's aperture. The user becomes able to see a problem whose existence had been obscured by the framing the previous category imposed. Once the new problem is visible, the relationship to the old problems reorganizes. The old problems become the wrong problems to have been focusing on, artifacts of the previous framing's compression. The new frame is the ground on which the right problems can finally be posed.

A precision worth making, because the term has been worn thin by marketing usage. Category creation operates on the conditions of legibility: what the user can perceive, what the user can name, what the user can pay for. The category-creating object reorganizes those conditions by demonstrating, in its own existence, that a different organization of the conditions is possible. The user encounters the object, finds the available frame too small to hold it, and either dismisses the object or expands the frame. Most dismiss. Some expand.

The near horizon

I suspect the field will move through three observable phases over the next few months. Each is already partly visible.

• Phase one: everyone has their own autonomous agent.

• Phase two: everyone has their orchestrated team of agents, marketed under whichever term wins the news cycle. Agentic engineering and vibe productivity seem to be the current contenders.

• Phase three: everyone has their fully automated workflow, where the agents communicate with other agents, the artifacts are produced by agents, and the human role becomes the pure supervision and verification of agent-to-agent traffic.

At each phase, the slop factory becomes more visible. The empirical signs are already in the data. METR's randomized controlled trial of experienced open-source developers found that experienced developers using AI tools were nineteen percent slower than developers without them, while believing they had been twenty percent faster. Sixty-six percent of developers report spending more time fixing AI-generated code that is almost right. Developer trust in AI accuracy has fallen from forty percent to twenty-nine percent in a single year, while usage continues to rise. The industry is hooked on something it distrusts. The slop accumulates. The participants remain blind to it, because the metrics they use to evaluate their own work are themselves products of the loop.

The slop factory is a present empirical reality that the field is reluctant to name structurally. The structural form has a specific shape: humans have been designed out of a process whose outputs still need humans to mean anything. The factory keeps producing artifacts. The artifacts arrive at recipients who have themselves been delegated to agents. Volume rises. Meaning collapses.

At some inflection point closer than consensus timelines locate it, the field will begin to ask what the agents are for. The default answer ("for productivity") will fail. Productivity, in the loop's working definition, has become "the thing the agents do." The recursion is the failure.

What replaces "productivity" could be many things: leisure, convenience, abundance, some naive form of freedom, accelerated consumerism of empty luxury. What needs to replace "productivity" is wisdom.

Above the loop

The category that comes after productivity is wisdom capital infrastructure.

Wisdom requires the cultivation of contemplative awareness, an increasingly scarce quality of human attention. It has never had infrastructure. The field's failure to perceive this, the loop's particular blindness to the qualities its own logic exhausts, is the structural fact that makes the post-productivity move available now.

A project is being built specifically for the post-productivity moment. Wisdom Cultivators are its primary participants. Wisdom Funders are its primary benefactors (and also potential beneficiaries) - the two roles are not mutually exclusive. The labor it organizes is immaterial in a particular sense: it feeds back into human value generation. The economics are designed so that contribution to the commons is the optimal strategy and conviction-weighted attribution is the mechanism of return. The work it supports is contemplative: slow integration of thinking across time, patient accretion of disagreement, production of artifacts whose value is irreducibility.

What comes next

What comes next is of Existential import/ance. I am, by temperament, a doomer. Much of what is being built right now will collapse within the decade. Most of the loop's output is a kind of potent hallucinatory exhaust, and the participants are getting high on their own supply. Few are building infrastructure that will survive.

What comes after productivity is the practice of cultivating wisdom. Wisdom changes our relationship to our mental loops, our habits of mind and patterns of behavior. It adjusts the cerebral labyrinth. It modifies conduits of attention. The work it asks for has a different shape than the work the loop has trained us to perform.

Travis Wyche is the co-founder of Existential. existential.systems

AI is getting faster; our judgment isn’t.

McKinsey built Lilli in 2023, a generative AI platform trained on nearly a hundred years of the firm's proprietary insights. Over 70% of McKinsey's employees use it. Lilli synthesizes 100,000+ documents, interview transcripts, and frameworks into client-ready output, with a McKinsey "tone of voice" agent that fine-tunes responses to sound McKinsey-esque. Forty percent of McKinsey's revenue now comes from AI and tech advisory, and McKinsey's revenue keeps climbing.

MIT's NANDA project tracked enterprise AI pilots in 2025 and found that 95% produced zero measurable P&L impact. The 2026 answer to that finding has been a wave of memory infrastructure. Mem0, Letta, Zep, Cognee. Persistent memory in flagship models is the latest norm. $16.5M to capture expert knowledge as permanent agent memory. Foundation Capital is calling context graphs a trillion-dollar opportunity. Deloitte is now reporting that 66% of organizations are seeing real efficiency gains. What 2026 calls memory is bigger context windows, persistent sessions, and vector retrieval at longer time scales. The truth is that it is all retrieval. None of it metabolizes anything.

The problem is the category, or the lack of one.

What AI actually does well

AI is excellent at retrieval, transcription, and first drafts. Find a detail buried in a thousand pages, it will. Turn forty minutes of speech into clean text, done. Generate a first draft of a memo, an email, or a business plan, no problem.

These are real strengths with real economic value. $0.06 per million tokens buys an individual real productivity. It buys an organization almost nothing.

What AI doesn't buy you is judgment.

Judgment happens when someone with years in a field, who has made hard calls and lived with the consequences, turns information into understanding. It is where experience meets information. It is what people mean when they say wisdom.

Why do you think the consulting industry exists? Why is it bigger now than it has been in twenty years? Consultants exercise judgment for you. People pay for it. People call it wisdom. McKinsey spent a LOT of money getting fired for learning about trash cans in NYC. Was it worth it?

The actual story in 2026 is even messier. McKinsey has deployed twelve thousand internal AI agents. Forty percent of their client engagements now include an AI component. Lilli, their consultant-facing assistant, accelerates the work their consultants ship. The deliverable still arrives bound, with a partner's signature. A model now does the retrieval inside, and the buyer is still paying for a similar judgment. What the buyer is increasingly getting is a well-formatted Google search, with the firm's blessing. Sometimes that search recommends bigger trash cans, and a city pays $4 million for them.

Real judgment is rarer than the engagement letter suggests. The position taken in real judgment costs something to be wrong about. The experience underneath it came at a specific cost: a deal that went sideways, or a strategy that failed and had to be reckoned with. We’ve all been on a call that aged badly and taught the caller what to look for next time, something that spoke to what was learned.

The foundation model companies are aware of this gap. They are trying to close it through endless update competition, longer context windows, memory features, multi-agent workflows, and reasoning models. None of it is moving the needle on what enterprises were actually trying to buy. Ask your model right now to give you the learnings of this last week, and you’ll understand what we mean.

We believe the fix is architectural, not algorithmic.

What memory doesn't fix

A foundation model trains on a static corpus and serves billions of users. The memory layer above it stores what you tell it across sessions. Neither is the same thing as having been the senior partner who got the consequences wrong on a deal in 2019 and now prices a similar deal differently in 2026. That knowing lives in a specific person, attached to specific consequences. Better memory and better retrieval get the model closer to a useful assistant. None of it turns the model into a body that has lived through what you are asking about.

The memory wave is solving the wrong half of the problem. It makes AI better at calling on what has already been thought. The thinking itself stays invisible, unattributable, and uncompensated when it shapes a decision the thinker will never see.

The structure

Wisdom is information a person has absorbed over time, through real experience, with real consequences. It is metabolized.

"We can know more than we can tell." Michael Polanyi, The Tacit Dimension, 1966

This is why your AI cannot know you, and why no algorithmic improvement will change that. What you want when you want AI to "know you" is something a language model, by construction, cannot do. No amount of RAG retrieval, long-context prompting, or memory graphs will turn a model into a body.

What has not existed until now is the infrastructure to recognize that absorption as economically valuable. To commission work from someone's lived experience. To compensate, attribute, and make visible the people whose thinking actually produced the result.

Knowledge management stores what you already know. Expert networks broker access to other people's knowledge by the hour. Bonfires.ai, built by the DeSciWorld team, capture community conversations across Telegram, Discord, Notion, and email and turn them into a queryable graph that the community can sell. The broader context graph wave, including the team productivity tools, ingests organizational discussion and turns the patterns into platform IP. Each takes a real piece of the problem. None of them recognizes the lived experience itself as the asset, and none of them routes value back to the person whose thinking shaped a synthesis they will never see.

That is the market enterprise AI has been reaching for and missing.

What happens next

There are two possible ways to close this gap.

1. Foundation model companies keep iterating on context and memory, enterprises keep hunting for the right prompt, and the numbers slowly improve as buyers get better at using the existing tools. This is the consensus path. It assumes the problem is user error.

2. A new category of infrastructure gets built, specifically for the things AI cannot do. Not a smarter model. A different shape entirely.

We are building the second one. We call it Existential.

Wisdom is not located in any one place. A person absorbs it through years of real experience and real consequences. Local infrastructure makes that experience legible without ever taking it from them. The network lets many people's structured thinking meet, refine each other, and compound into something none of them could see alone. Wisdom is what those three layers produce together. AI alone has never been able to produce it, because AI alone is only ever one piece of the picture.

In practice, that means three movements:

• A person's thinking, captured locally, passed through many specialized steps, structured into something legible without ever leaving their machine.

• That structured thinking meets other people's structured thinking across the network, where different angles, domains, and histories surface what each would have missed on their own.

• And when someone with a real question and real stakes needs answers, the network surfaces the consequences, the ideas, and the holes from across the contributors whose lived experience is actually relevant to what they are asking.

The people whose thinking produced that synthesis get paid for it via a smart contract, with 70% of revenue routed to them and each contributor's share determined by how their thinking shaped the result.

That is the actual category. Wisdom is the real market, and it has never had infrastructure.

Waitlist at existential.systems. DMs are open.