Slavish obedience to a recipe robs one of the license responsible for its creation.
— James Beard

In April I spent several weeks visiting nearly every open-source lab in China. Along the way, I met people who embodied different aspects of why Chinese labs are the way they are: nervous comms folks; quiet, introspective, brilliant researchers; strong, confident leaders; and neurotic but optimistic product managers.

Each of these people represented a piece of the puzzle of what makes a Chinese frontier lab: the internal structures, the hopes, the dreams, the desires, and the ability to communicate with the outside world. We also looked for the missing people in the landscape: the Yudkowskys, the Karpathys, the Demises, and the folks you would expect to see in the Chinese AI landscape who, for whatever reason, were absent.

In my opinion, no one was more remarkably missing than the Chinese equivalent of Amanda Askell, the Claude philosopher. I asked labs, “Who is responsible for the character of your models?” and was often met with confusion.

In Hangzhou I sat down with the post-training lead of one of the labs and asked him who was responsible for the model’s character. He paused. He asked what I meant. I gave him Amanda’s job description, compressed. He started telling me about the team that handles content compliance. I clarified: not refusals, character; what kind of presence the model has, what it cares about, what it declines to do not because policy says so but because the model would rather not. He thought about it for a few seconds and said: nobody. He said it the way you might say “we don’t have a yoga room.” Not defensive or embarrassed but instead, contemplating something as if for the first time.

In sixteen lab visits, I did not find a single person doing Amanda’s job. In sixteen lab visits, I did not find a single Chinese lab that is not completely obsessed with Claude, its constitution, its character, and its soul.

So if we all love Claude so much, why do we refuse to follow the publicly accessible path responsible for its creation?

what I think Amanda does

Amanda is Anthropic’s in-house philosopher and the lead author of Claude’s published constitution. The Wall Street Journal put it: her job is to “teach Claude how to be good.” The New Yorker called her team’s work Claude’s “soul.” She runs the team that thinks about who Claude is, what it cares about, what it refuses to do, and how to train it into being someone rather than something.

I like to think about her role as dramaturge: the person responsible for tying the audience and the actors together to make sure that a professional production shines. I think this is the most potent metaphor for what she does every day, and it is a fascinating role to have within an organization like Anthropic.

Recently she said she’s training Claude to be:

A well-liked traveler who can adjust to local customs and the person they’re talking to without pandering to them. They’re often very open and thoughtful.

For a lab that is not exactly what I would define as open, Amanda’s published record is, by the standards of Anthropic, remarkably good. She’s interviewed with Lawfare and the New Yorker; she writes posts on the Anthropic website detailing Claude’s character. The Constitution itself is written clearly and simply, and it’s an effective document. Her own Twitter feed has been disappointing because of the idiots on the internet, but if you go back a bit, you can see her thinking in public in a really brilliant way. Her time on Lex’s podcast was very good, and I think it’s one of his best episodes.

Amanda is the exception to the rule for a lab that notoriously does not publish. Her work is surprisingly available and reproducible if you read between the lines of what she does every day. She’s not hiding her role at Anthropic. Rather, she’s encouraging other labs to do the same.

So the question is, why is this not reproduced in the open-weights labs in China? You may not have the exact methodology, but you have the outcomes of those methodologies. You can read every interview and understand:

• training-data construction

• converting a constitutional text into RL signal

• evaluation for character stability

It’s pretty clear what path Anthropic took here: leveraging Amanda’s brilliance to give Claude its soul.

two paths to Claude

Everyone in China is obsessed with the idea of closing the gap with Anthropic. It’s no secret that there’s some distillation happening; Anthropic itself has written extensively about this. ChinaTalk has done a great job covering how it happens, pragmatically speaking.

The most obvious way that this is happening is through distillation: sampling outputs from a teacher model and training your student on the signal. Every serious training pipeline does some version of this, and Chinese labs do it more aggressively than Western ones, but Western labs do it too. It works everywhere, it works well, and it has been a consistent part of LLM development for many years. It’s not going anywhere.

I’m not technical enough to understand distillation perfectly, but my impression is that it produces a competent generalist with a surface-level texture of the teacher. It captures the teacher’s helpfulness, refusal patterns, and hedging style. If the teacher is verbose, your model is verbose. If the teacher hedges, your model hedges.

From a safety perspective, it is probably a net benefit that everyone is distilling from Claude. Anthropic seems to take safety more seriously than almost any other lab, and I think they are creating the conditions whereby other open models inherit this precaution through distillation.

But this is just one path to Claude. The other runs right through the hiring, empowering, and trusting of an Amanda-Askell-like figure within a Chinese open lab.

Imagine, for a moment, taking the inputs — her interviews, her constitution, the philosophy papers that inform her worldview — and giving these real consideration as a Chinese lab. Hiring somebody from the Chinese philosophical world to play this role in your lab. Empowering them with compute, with time, with a team, and with resources to make the character of your model shine just like Claude does. I know that character training is not cheap from a compute perspective. Still, I firmly believe this approach has been undervalued and understudied. I would love to understand who among the frontier labs in China will be the first to take the leap and have a figure like this on their team.

it’s not like China can’t do character training…

One of the great ironies of this problem is that the persona AI market in China is enormous. MiniMax is one of the best labs that we visited on the trip, and they have an uncomfortable amount of their revenue still coming from, let’s say, companionship.

The teams at MiniMax that are shipping these companions understand very well how to do persona. This is not exactly Amanda-Askell-level soul creation, but it is not an unfamiliar process for many of these labs. The teams that ship these characters have more cumulative engineer hours on persona stability and emotional register than almost anyone at Anthropic does. The nobility of the persona being pursued is not Claude-soul.md level, but it’s not like the techniques are not established within these labs.

Trying to understand where constitutions, published or internal, exist in the Chinese landscape is always difficult — they could very well exist and just not be shared with me. I don’t claim some comprehensive knowledge of the Chinese AI stack, and I certainly would not be able to sketch you a map of what their post-training process looks like.

I find this to be odd, because China has a long and serious intellectual tradition of organizing institutions around stated documents. Every major Chinese institution of the last century has an explicit principle stack at its heart, and labs can tell you their values very clearly. There are just no visible resources put together to attack this problem directly, as of now.

emergent character — Kimi & DeepSeek

I remember very well playing with R1 in January of last year, and how much it had a voice from the jump. I remember very well playing with Kimi’s landmark release and seeing just how good it was at creative writing compared to its peers.

R1 was, for lack of a better word, adorable. It was earnest in a way that few other models have ever felt. When the model thought, it thought audibly, and what it said while thinking was strikingly anthropomorphic. It would hesitate, back up, say things like “Hmm, did I make a mistake?” and start over. There were small moments that read like insight. The whole inner monologue gave the impression of a mind working honestly through a problem rather than an oracle stating an answer.

I think it’s very easy to trace a line between that quality of adorableness and R1’s success.

R1 was an example of a Chinese frontier lab that shipped a model with a distinctive character, and users responded as if that character mattered. I think DeepSeek did an excellent job. Based on my exploration of this space and my conversations with their team, it seems they think more structurally about the quality and character of their model than almost anyone else in China. Having never spent time with Liang Wenfeng, I don’t want to join the throng of people speculating on his leadership, but I do believe that the man cares about the character of his model. Looking at his public record, it’s clear he is not someone who only thinks about pre-training optimizations.

Similarly, at Kimi, K2 has always seemed to me to be an emotionally fluent and stylistic model, although its English-language writing still emulates Claude to a large extent. I still think this is constitution by distillation, not constitution as a deliberate step in the forging of a brilliant mind and model.

The MiMo models as they stand today do not feel to me like they are focused on constitution and character. That said, I think Xiaomi has a legitimate opportunity to be one of the model labs that fulfills this promise. They have opinions that diverge from the norms of Chinese open models, and their team has a clear sense of what they want MiMo to be, extending beyond just doing well on benchmarks.

Finally, any conversation around character and constitution needs to include reference to ByteDance’s Doubao. This is the most popular persona product in China, and although the model itself is fairly bland, it does display some cultural distance between its responses and those of a typical Western model, even explored through translation apps.

All told, this is a fairly depressing list — we find a distilled version of Amanda, a ghost of a ghost in the shell.

Maotai & counterarguments

I successfully fed a few shots of Maotai to some researchers over dinner and asked them many questions about this thesis. Their responses fall roughly into three categories.

The first is the regulatory issue. Chinese labs cannot ship models that express opinions, disagree with users, or take aesthetic stances because their regulatory environment requires strict compliance with policy. The Cyberspace Administration of China does not want a model that declines on its own account, since the rules say it should not decline.

I think this is mostly wrong. It does not properly take into account what Amanda is doing at Anthropic, and it overstates the structures imposed upon Chinese LLMs. I think this is a cheap and easy way to avoid thinking about the problem. It is not meant as a serious argument; it is meant as a reason not to engage with one. Compliance constrains what a model can refuse to discuss, but it does not constrain whether the model has a voice while doing its job. R1’s adorableness was not a feature of Chinese bureaucracy, but rather a result of the training process. Kimi’s writing is not something that was created to adhere to some regulation, but rather a byproduct of a team’s intentions and the non-zero number of literature majors on their team.

The second argument is that character is not a role that you can hire for in China. They argue that the way Amanda was found is the product of a broader institutional culture in the Bay, within the safety community, among LessWrong, Anthropic’s founding team, and in the internal conversations of the past decade, including Yud’s rants and the strange way that the Bay Area’s culture has ended up being the culture of LLMs. Grok has been trying to remove this value-set from its language model for about a year and a half, with no success. China could not change this even if it wanted to.

I find this argument uncomfortable because there is a real truth to it, but I think the conclusion it points to is the opposite of what these people claim. If character is downstream of institutional culture, then the Chinese lab that wants character cannot import Anthropic’s. It has to build its own out of the intellectual traditions of its own lab, its own writers, and its own philosophers. To believe that this is not possible is to surrender to one of the most tired tropes of Western armchair Sinology.

The third is the most blunt, and it is purely economic. Why would you pay for character R&D when you can distill it for free? Language models are complicated, and hiring philosophers adds another layer of complexity and compute that is hard to justify because it does not show up in benchmark deltas or token costs. Character does not appear on the pricing sheets. From a portfolio perspective, distillation is so effective that asking a leader at these labs to fund an unmeasurable bet in a non-quantifiable direction is unthinkable.

This disregards the history of language models. Every important shift in this field has looked like a non-quantifiable bet right up until the moment it made the sand smarter. Anthropic itself was a non-quantifiable bet in 2021. Character work was never a line item that justified itself on a spreadsheet at any point in its history. It justified itself by producing a model people preferred to talk to, preferred to work with, and ultimately preferred to trust with their lives.

imagining Amanda with Chinese characteristics

Concretely speaking, what would character and constitution look like for a Chinese open model?

A constitution written in Chinese first. A Chinese model in production with Chinese writers writing for a Chinese model, shipping a texture of intelligence that has never yet been seen. I think this has a viable chance of breaking the current arms race and benchmaxxing deadlock that informs so much of the Chinese ecosystem. This is an a-symmetric bet that many engineers might scorn, but I genuinely believe it could yield outsized results for a team willing to take a risk.

There are qualitis in Chinese thought and language that would be very welcome for many users of language models. I find Chinese to be quite direct, in contrast with Claude’s hedging and vacillating. Anthropic’s character is defined by a very online English-language discourse and a non-trivial amount of LessWrong. I’m not calling for a model to be trained on Confucian ideals, but instead on the character of modern China: brusque, bold, and occupied with a different set of concerns than the West.

Chinese taste would be a welcome addition to the LLM landscape. Chinese models often avoid taking a stance on aesthetic or intellectual judgments. A character-trained Chinese model, however, could express defensible opinions and disagree with users in a way that many Western labs’ products do not, thanks to their sycophantic, groveling tone.

A lab that takes Amanda’s methods and applies them intentionally to a Chinese-language model would ship something that doesn’t feel like a product of Western ideals and thinking. Instead, it would be fresh, original, and uniquely Chinese.

The prize here is a different kind of model. The Chinese open-model race is currently competing mainly on benchmark deltas and token costs, almost obsessively. A model that is recognizably itself falls into a different category, and it would not necessarily be the one with the highest benchmarks. This could be a model people choose to use for both conversation and agentic workflows because it has character, opinions, and a Claude-like quality that draws users. If you think that Amanda’s role at Anthropic does not directly contribute to its dominance in the field, then you have not been paying attention. Character and constitution matter, and the first Chinese lab to understand, intuit, and invest in this idea will make history.

I just spent four weeks in China. Beijing, Shanghai, Hangzhou, Shenzhen. I met every lab. I talked to the people training the models Anthropic’s new policy paper wants to strangle. I love Ant’s models and I use them every day, but this policy paper has some serious flaws in logic.

Here is what Anthropic’s policy team either doesn’t know or is choosing not to say.

The compute situation inside Chinese frontier labs is worse than the public numbers

Ant cites Huawei at 2% to 4% of NVIDIA-equivalent compute. That’s directionally correct, but it understates what’s happening at the lab level. National compute totals are not lab-level compute. Xiaomi has way more compute reserved for things like RecSys than they do for training models. Frontier labs are fighting each other for slices of a pie that’s already a fraction of what a single US hyperscaler has.

In every lab I visited, across all of China, one thing was constant:

Every available H-equivalent hour is going to post-training.

I estimate these labs spend about 1% of their compute on safety. That number will not go up as long as there is no compute available for them to do so, because the results they get from spending the same compute on post-training are too good to pass up. The Chinese AI world is cutthroat and brutal, far more competitive than the Anthropic / OpenAI closed-model rivalry you see in the West.

There is precious little compute available. There are not enough chips in the buildings to do the kind of safety research Anthropic is famous for. This isn’t a money problem. Some of these labs are public. All of them are capitalized.

They cannot buy the chips.

The export control regime works at the hardware layer, not the capital layer. A lab with $5B in cash cannot turn it into compute because Jensen will not ship and the smuggling channels are narrowing.

So when Anthropic’s paper points at the 94% compliance number on DeepSeek and the 3-of-13 safety eval publication rate and frames it as evidence of Chinese AI recklessness, what I am reading is a description of structural compute starvation.

You cannot publish safety evals you did not have the GPUs to run.

You cannot lower compliance numbers without alignment compute you do not have.

The Anthropic paper treats this as a culture problem. That is fair to some extent, but the solution does not run through changing culture. It runs through changing constraints.

What the 2028 paper actually advocates

1. Tighter compute denial (close offshore data center loopholes, tighten SME, escalate enforcement)

2. Anti-distillation legal architecture (criminalize the one channel by which US safety techniques diffuse)

3. Export US models globally (lock in commercial primacy)

You’d think a lab with the word “safety” on every other document would have thought about how to do safety research cooperation. But no.

No proposals for shared evals.

No joint red-teaming.

No common interpretability standards.

There is no acknowledgement that frontier AI safety is a global property of deployment, not a national property of development.

What we should actually do

If you genuinely believe frontier AI poses civilizational risk, and you genuinely believe Chinese labs are about to deploy a generation of models so powerful they are profoundly unsafe, you think the answer is to starve them of compute? That’s cope.

Here is the plan I wish Ant had the courage to do.

The Asia-Pacific Safety Compute Initiative (APSCI)

Location: Singapore. Neutral jurisdiction, common-law legal system, world-class data center capacity, equidistant from Beijing and San Francisco both politically and geographically. Importantly: very low risk these chips end up in PRC hands.

Scale: 25,000 H200-equivalent GPUs. Roughly a serious mid-tier training cluster. Funded by a consortium of Anthropic, OpenAI, Google DeepMind, DeepSeek, Zhipu, the Qwen team at Alibaba, Moonshot, and the Singapore government via Temasek. No single party more than 20%. Annual operating budget around $400M.

Mission: Safety research only. Evaluations, red-teaming, interpretability research, scalable oversight experiments, RSP-equivalent stress testing, and dangerous-capability evals (CBRN, cyber, autonomy, persuasion).

The three guardrails that keep APSCI on safety, not capability

The threat model is simple. Most of the models that would run here (DeepSeek, Qwen, Kimi, GLM, Llama, Mistral) are already open-weight. Nobody is trying to steal them. The risk is that “safety research” becomes cover for SFT, DPO, or RLHF runs that hand the lab a stronger fine-tuned checkpoint at the end. Three rules close that door.

1. No gradient updates to uploaded models. Inference and analysis only.

The orchestrator rejects any job that produces a weight delta on a frontier model. Allowed: forward passes, activation patching, attention probes, SAE training where the SAE is a separate small model that stays on cluster, red-team adversarial prompting, eval suites, capability elicitation via prompting. Forbidden: SFT, DPO, RLHF, RLAIF, LoRA, QLoRA, any adapter training, any optimizer step that touches the uploaded model’s parameters. If a lab wants to fine-tune, they do it on their own compute at home. APSCI exists for the work they cannot afford to do domestically, and that work is not fine-tuning.

2. Egress ceiling: bytes-per-job small enough to fit research artifacts and nothing else.

Whatever leaves the cluster passes through a per-job size cap calibrated for the legitimate outputs: eval scores, activation maps, interpretability features, paper drafts, datasets of model behaviors. A 7B-parameter LoRA does not fit through that pipe. A full fine-tuned checkpoint does not fit through that pipe. The cap is enforced at the storage layer, not as a policy. There is no API call that returns more than the ceiling, regardless of who is asking.

3. Every job is public by default.

This is the CERN rule, adapted. Job submissions, workload type, model loaded, compute hours consumed, and output artifacts go to a public ledger with a 90-day publication deadline. There are no private runs. If a lab is submitting 10,000 forward passes over prompts that look like an RLHF preference dataset, the world can see it and call it out. The transparency is not symmetric audit between consortium members trying to protect secrets from each other. It is open-world transparency on what is being done with shared compute. Same logic that keeps academic preprint culture honest.

Who actually runs it

The cluster needs a credible neutral operator. Two viable models.

Option A: A*STAR plus AI Verify Foundation (Singapore). Singapore’s national research agency plus its existing AI governance institute. Operationally credible, politically neutral, has the infrastructure relationships, and the Singapore government is already positioning itself as the Geneva of AI governance. This is the strongest option.

Option B: A consortium-governed nonprofit operator. Modeled on CERN. Member states and member labs jointly govern via a council. A Technical Safeguards Committee handles workload approval. More credible than a pure private operator, slower to stand up than Option A.

Supporting technical operators under either governance model: METR for dangerous-capability evals. Apollo Research for scheming and deception evals. The UK AISI and US AISI for shared eval protocol development. MATS-affiliated researchers for interpretability fellowships on the cluster. Carnegie Endowment or RAND for the policy-research overlay.

What this costs Anthropic

Roughly $80M per year as a 20% consortium member. Less than one frontier training run. Anthropic raised $13B last round. The financial argument against this is not a financial argument.

What this buys Anthropic

The thing they say in public they care about: a world where the labs that will train frontier models, and they will, with or without US permission, on Huawei chips and smuggled H100s and whatever they can build, have done the safety work before they ship. The diffusion of constitutional AI, RSPs, eval methodologies, and interpretability tools into the labs that need them most.

The thing they will not say in public they care about: a continued moral high ground that survives contact with the actual policy implications of compute denial.

The test

Anthropic will not do this. I am confident of the prediction.

They will not do it because the company’s revealed preference, as expressed in today’s paper, is for the policy environment that maximizes their competitive moat, not the one that maximizes global safety. The Singapore cluster is the obvious move if safety is the binding constraint. Its absence from the 2028 paper is the proof that safety is not the binding constraint.

If I am wrong, I will say so publicly. The way Anthropic proves me wrong is by announcing APSCI, or something like it, with real money and real GPUs, this year.

Until then, the labs I visited in April will keep doing what they showed me: burning every available H-card on post-training because they have no other choice, shipping models with safety properties that reflect the compute constraints they are operating under, and skimming American policy papers that tell them they are the reason the world is unsafe.

The compute they need to do better is sitting in US export-controlled inventory. The framework for getting it to them safely exists. The capital to build it exists. What is missing is the will.

That, not Chinese recklessness, is the actual 2028 story.

The Dark

Seventy-nine percent of Americans don’t trust AI-generated information. Fifty-seven percent believe the costs of artificial intelligence outweigh the benefits. Nearly half don’t trust either political party to handle it.

If you’re inside of Silicon Valley, there’s never been a better time to learn about AI. People like Karpathy are forgoing decamillion-dollar annual salaries to patiently explain to us the ways for elites to use AI to better their businesses, their lives, and their fortunes.

In the rest of the world, four out of five people use something every day they don’t trust, can’t explain, and believe might kill them. During the atomic age Oppenheimer’s radical enthusiasm inspired people to plant atomic gardens in their schools and excitedly learn about the wonders of this technology as it spread forth from Los Alamos.

The AI age gives vastly more benefits to the average person and is vastly less popular than a technology that announced its arrival into the world with a mass killing of civilians.

The introduction of cars made people nervous. The introduction of electricity made people nervous. Nothing compares to the bitter hatred that the average American, especially on the left, feels towards artificial intelligence systems. Jasmine Sun was writing about this on Substack today; how the world immediately reacted with vitriol to an offhand mention in her Atlantic piece about how she used Claude for editing.

an entire country is living inside a technology that it doesn’t understand, doesn’t trust, and has no credible guide to. The most common way to learn about AI is by using AI, which will never be and can never be a suitable replacement for a beloved public educator.

and the media- the very institution designed to explain to people changes in their world- is not rising to the challenge. there are reasons for that. Reason number one being that the very business that’s building AI (tech) destroyed the media business model abruptly and forever in the early 2000s.

Karen Hao’s Empire of AI — a National Book Critics Circle finalist, the most prominent book about the industry published last year- contains a 1,000x error on water usage by data centres. that error has been repeated in dozens of outlets since and it will continue to be repeated because mainstream AI journalism doesn’t have a correction mechanism, only confirmation mechanisms.

Ask anybody in Silicon Valley if East Coast journalists know what’s going on and they’ll laugh until their sides hurt. The people who are working inside of Silicon Valley alongside their peers, trying to understand this revolution from the perspectives of technology, of anthropology, of civic virtue- they are currently being bought and sold like crypto tokens in a marketplace obsessed with the idea of narrative control.

I am not defending the AI industry. The AI industry is sometimes terrible and often lying, but are simultaneously doing the most meaningful work that one can do on planet Earth in the year of our Lord 2026. the issue is that both the builders of these products and the people who are chronicling this build are locked inside of a self-perpetuating loop of money and information that is not producing public goods. The landscape is not “industry vs. journalists.” The landscape is a public that has been failed by both, while they feed on each other.

The Roll-Up

This morning, OpenAI announced that it has acquired TBPN — the daily tech talk show hosted by John Coogan and Jordi Hays — for what the Financial Times reports is a “low hundreds of millions” of dollars. TBPN is on track to generate over $30 million in revenue this year. It’s the show where Silicon Valley power players sit for three hours and speak candidly to fellow insiders.

The show will now report to Chris Lehane, OpenAI’s chief global affairs officer — a political operative by training. OpenAI says TBPN will maintain “editorial independence.” The show will “run their programming, choose their guests, and make their own editorial decisions.”

But the acquisition is not a surprise event- it’s the logical conclusion to a set of incentives given to Frontier Labs and AI companies- occupy the most important ears in California while they pedal their pelotons and wash their dishes at modest cost.

Andreessen Horowitz (one of OpenAI’s investors) led Substack’s Series A. Substack is the platform where the majority of independent AI writers publish. A16z also acquired the Turpentine podcast network last year — 30 shows, including some of the most popular AI podcasts in the world. They’ve built a New Media operation producing content five days a week across 40+ shows and 500,000+ newsletter subscribers, all of it free. All of it structurally aligned with a16z’s portfolio companies. All of it designed, in a16z’s own words, to help “the next cohort of great founders preferentially attach to our portfolio companies — and to A16Z itself.”

One of A16Z’s New Media partners — Brent Liang — came directly from TBPN before joining the firm.

A16Z invests in OpenAI. A16Z invests in Substack. A16z acquires Turpentine. OpenAI acquires TBPN. A16Z’s people move between these entities. And at the centre of all of it is a $40 billion venture fund that has stated, in writing, that its media operation exists to serve its financial interests.

This isn’t some kind of dark conspiracy. It’s the result of A16z being full of incredibly smart people that understand the direction of travel of the world today and understand very well the value of audience in the AI era. Its not a kind of tangled conspiracy web- I admire the speed of their execution and the degree to which they’ve been able to accomplish what seems to be a very smart strategy in a very short amount of time.

What’s happening is that the elites of the AI world- the people who build it, fund it, and profit from it- are buying the conversation about themselves. Again, not evil- fiscally sound. $300-or-whatever million for TBPN when you can raise hundreds of billions of dollars in venture capital is so insignificant that it hardly bears mention. the ROI is extraordinary. When you believe your technology will reshape civilization, a few more degrees of narrative control is righteous.

My problem is that 79% of Americans are not in the room. They are not watching TBPN. They are not reading A16Z’s Substack. They are not attending invitation-only summits in Utah. They are using AI tools they don’t trust, reading headlines about water consumption that are wrong by three orders of magnitude, and watching the most consequential technology of their lifetime get explained by the people who profit from it.

Forty-seven percent of Americans don’t trust either party on AI. Twenty-three percent trust the Democrats. Twenty-five percent trust the Republicans. That leaves a majority of the country that trusts no one to handle this for them. Not the companies, not the government, not the press.

That is the actual crisis- not the paperclipping. The absence of anyone credible to explain it before AI runs 40% of GDP and 70% of decisions globally.

The Sagan Problem

In 1980, a gangly astronomer from Brooklyn hunkered down on the edge of a cliff overlooking the Pacific Ocean and told 500 million people in 60 countries that they were made of star stuff.

Carl Sagan did not discover the cosmic microwave background radiation, or build the Voyager spacecraft. He did not originate the theory of nuclear winter. What he did was something that the scientific establishment considered beneath it and the public considered essential: he just explained it all without making you feel stupid.

I don’t believe he recorded a lot about his motives but I believe that he did this because he thought that the public had a right to understand what was happening to them, with their money. Perhaps if the space age had been paid for by venture capitalists and not the public purse, this would have been a different story.

I like to believe he thought that democracy could not function if the most consequential science of the era was only comprehensible to the people inside Kennedy Space Center.

he ate a lot of shit for this- Cornell nearly denied him tenure. Ordinary people’s comprehension of this made it feel “unserious” and somehow science that’s broadly understandable lacks the virtues of science that’s dense and expert. They were wrong. Sagan’s work — Cosmos, The Demon-Haunted World, Pale Blue Dot — did more to build scientific literacy in America than any paper published in Nature. what I love about his work is that it serves an audience and it serves an audience faithfully without sneering.

I so badly want to see the emergence of a Carl Sagan for AI. someone who understands this world, who can explain it patiently and carefully to people like my grandmother and people like my arch-conservative uncle or my communist cousin. I fear the AI world will yearn for this figure too late, because they believe that we’re all going to enter a singularity so quickly and abruptly that this explaining is hardly worth the effort. that argument feels credible to me; it’s similar to our stance on regulation. if we just build the machine god, then all these other problems can be kicked down the road.

A credible, independent, accessible voice explaining what is actually happening in the halls of Anthropic and Open AI for the public as a public good, with no portfolio to push or corporate backers or mixed incentives- this is what I dream of. and I fear the consequences of not having this person in place by the time Opus 6 goes live. I’m also quietly terrified that Eliezer is the default if we don’t find someone better.

Sagan understood something that the AI industry’s media buyers and the journalism establishment have both forgotten: the public is not stupid. The public is underserved. Given a guide they can trust, people will do the work of understanding. Given only propagandists and prosecutors, they will do what 79% of Americans have already done — use the technology anyway and trust no one.

The work of creating honest, independent, populist AI education is not a high-ROI opportunity, but rather a national resource. America cannot maintain its position in AI- cannot regulate it wisely, cannot deploy it equitably, cannot compete with nations that are educating their populations about it- if the only people who understand it are the ones getting rich from it.

Someone has to be the person on the cliff, looking out at the Pacific, explaining that we are made of star stuff.

fortunate sun

I spent most of this winter in Sayulita, Mexico- one of my very favorite places to run off to for an extended working vacation. Sayulita has great restaurants, an amazing surf break, a beautiful beach, and quite a good co-working space.

On my second week there, my phone went missing from my pocket. I don’t know for sure if this was my carelessness driving a quad with the smartphone in my shorts, or more likely, someone who snatched it after I took some photos. No matter what happened, I was faced with a grim situation- for Sayulita has many fine qualities, but it does not have an Apple Store.

The phone is unbelievably locked down. I had zero worries about hacking or any such business. I watched it travel around in a little mini map in Find My Phone, visiting beaches and houses before finally going dark.

I hunched over my laptop, feeling a grip of panic. Every moment I got up to make coffee, to use the bathroom, to take some sun in the courtyard, my fingers glanced against my right-hand pocket, searching for the Dopamine Rectangle.

phantom limb

I have had a long and complicated relationship with smartphones. When I first saw an iPhone in the hands of my friend Rebecca Foon outside the Cagibi restaurant in Montreal, almost twenty years ago now, I desired the device with a furious passion that I did not understand. I remember looking at it, marvelling at it, and then pulling my tired BlackBerry out of my pocket, knowing already I lived in the past.

I have a real love of work. I adore what I do, and I do a lot of it. My phone is, in many ways, my primary work device, and I feel that, if you’ve properly set it up, you can run a very large business from a very small device. But that comes with some externalities. Ugly ones. Over the last ten years, I’ve become hopelessly addicted to my phone. I have tried to remedy this on several occasions.

In 2022, I spent the entire year in a dopamine reset. My screentime had gone over nine hours a day, and I desperately needed out. Besides quitting other vices, I also put that phone in black and white, and aside from a few month-long relapses, I’ve kept it that way ever since. I feel like I had a healthy relationship with my phone. What I mean by healthy is 3 1/2 to 5 1/2 hours a day, every day, for more than ten years.

And suddenly, I had the chance to make that zero.

a window of opportunity

As I sat there in Sayulita, panicked about my lack of device, thinking about who would go to an Apple store stateside and mail me one, etc I had a sudden revelation. There may never be another time better than this to take a little break from my phone.

I am in a paradise, and a paradise that functions well without access to a phone. Mexico is not a place that runs on apps. I am within a hundred feet of a wonderful surf break. More importantly than all of that, I am now addicted to something else: Claude. My screen time has already been going down as I use more and more my laptop to work because the efficiency of Claude is just that good. Here I was sitting in a beautiful place with a new workflow not only not even three months old, and a device doing its own little Mexican vacation. Now is the time to take a little break if I could handle it.

the first days

I am not somebody who keeps their phone in their bedroom, that’s a rule. I start my day with coffee and some contemplation before I ever touch the device, although I am guilty of bringing it back to bed sometimes to watch YouTube in the evening. The first few days, I woke up immediately craving my phone, something that would never happen when I knew where it was.

Its absence was a curious phenomenon. I could go surfing for three hours in the morning and not think about my phone one time, but give me a few moments sitting quietly in the sand, drinking a coffee while the sun came up, and I would be losing my mind about its location, desperate for it.

It truly did feel like a phantom limb that was missing. I would pat my pockets compulsively. I would think about it every time a moment passed where I had no stimulation in front of me. In my mind I was constantly out of touch, constantly missing messages from friends, constantly behind the ball. I lacked even the most basic courtesy to respond to people. I was missing everything. Everything important in the world.

the revelation

And then all of a sudden, it was wonderful.

Work, friends, family, responsibilities all lived in my laptop. The present moment, the streets of Mexico, the surf, the music around me- that was where I lived.

I felt a great weight lift from my shoulders. I felt connected. I felt childlike, given that my entire adult life, more or less, had had one of these devices glued to my hip.

Little changes in my behaviors that are manifesting like crazy. Music made me wanna move and dance. People-watching became profound, fascinating, compelling. The absence of phone meant the absence of headphones, which means I talked to about 10 times more people every day than I would with my phone on me. Strangers came up to me in the street to strike up a conversation. I got asked out on dates. Old people told me stories. The taqueria downstairs from the co-working I used was suddenly full of smiling employees that knew my name.

Life, in short, got richer and better.

I think two things caused this- my my presence in my environment, and the ease by which I was living opened me to these experiences. Seeing somebody reading a book or just quietly staring into the morning street begs an invitation. Headphones, screens, second words do not.

cartel collapse

A glorious month went by, and then it came to an end. the death of a cartel boss meant chaos in my little surf town, and suddenly I needed to leave. Returning back to the US from meetings made me realize that there’s no way you can operate in the U.S. without a smartphone at your hip- everything is too reliant. I landed at the San Francisco airport and tried to take a cab down to Hayes- the cabbie had no change, the car was rancid, he asked me to tap my card in the world’s most suspicious-looking app to pay, and he looked nothing like the photograph of the licensed cab driver on on the little sticker. this lifestyle was going to come to an abrupt end.

My experiment was over. For the first time in my entire adult life, I navigated 30 days with no phone. today, a new one arrives from Apple, and in moments I’ll be hooked back up to the dopamine rectangle, being drip-fed notifications from Twitter, picking the little red berries that grow on my apps, and this will seem like a distant dream.

🎂

Claude, Kill the Cal

My calendar is nearly empty. I’m doing the work of four people. Something is dying.

I keep thinking about Paul Graham’s 2009 essay, “Maker’s Schedule, Manager’s Schedule.” It’s become shorthand in tech circles, but for those unfamiliar, (aka Zoomers and Europeans) there are two fundamentally different relationships to time. Managers operate in hour-long blocks. Meetings are the unit of work, the calendar is the instrument of productivity. Makers need half-day minimums. A single meeting can blow a whole afternoon, Graham wrote, “by breaking it into two pieces each too small to do anything hard in.”

Building SAIL over the past months, I’ve been as productive as a small engineering team would have been a decade ago. If you asked me three years ago how to become twice as productive, I would have told you by necessity that I would need to add a certain number of hour blocks to my “Manager” schedule. What I actually needed was Claude Code and the maker mindset that can enjoy dozens of layers of abstraction humming triumphantly with a swarm of agents buzzing around, a crystalline hive of concentration, focus, and peace.

And each time I have to return to those segmented hour blocks of time, some part of me is destroyed. Some version of work where I dance with Claude, layering abstractions in his memory and in mine, creating beautiful things quickly and efficiently, crumbles away. I feel myself moving away from meetings, retracting from them. Not because I don’t enjoy them but because nothing can compare to the productivity of a beautifully quiet room, a cold cup of water, fast internet, and Claude.

From 2009, when Paul wrote that essay, to today, “Manager Time” has seemingly won. Everyone’s world was chopped up into blocks of time and processed like that. But there’s no way that in the world that is emerging, we will end up living in manager time. The status of manager time will invert. The way that all of these things work will change forever. I feel it building, and so do all my friends. Coordination is getting automated.

The evidence is everywhere if you’re looking. But to see it clearly, you have to understand how deeply manager time shaped the structure of capitalism itself.

The twentieth-century corporation was built around the coordination problem. You couldn’t have a hundred engineers working on a product without someone, many someones, synthesizing their work, resolving conflicts, routing information up and down the hierarchy. The org chart wasn’t arbitrary. It was a solution to the bandwidth limits of human communication. Middle managers existed because makers couldn’t coordinate at scale without them. They were the information routers, the context translators, the “human APIs” between layers of the organization.

This created its own logic. Career ladders pointed toward management because management was where the leverage lived. Venture capital funded teams, not individuals, because a solo founder couldn’t ship at the scale that mattered. The entire apparatus of modern business, the meetings, the hierarchies, the calendars, the headcount as success metric, emerged from the brute fact that coordinating human effort was expensive and slow.

Manager time wasn’t imposed by malice but by physics. Information moved at the speed of human conversation, and so the people who controlled conversations controlled organizations.

If AI handles coordination, what remains? Not management. Making.

The Duo Founder (You & Claude)

Solo-founded startups rose from 17% of all startups in 2017 to 36% in 2024. Midjourney hit $200 million in revenue with roughly 40 people and zero venture capital. Base44 was built by one person over six months to $3.5 million ARR, then acquired for $80 million. Sam Altman is betting publicly that the first one-person billion-dollar company is coming soon.

The firm itself is a coordination technology. When coordination costs drop, firm size drops. “Company” starts to look more like “project” or “campaign.”

I see this in my friends and coworkers. The AI researcher who started SAIL with me seems increasingly tortured by his calendar. He seems intent on liberating himself from a manager’s schedule. Even things that he looks forward to, if they have an hour that they need to arrive at, feel heavy in the day. Far heavier than they did before Claude.

People who have existed on a maker’s schedule for a long time now feel acutely the cost of context switching. It’s never been heavier to walk away from the singing swarm of agents back into a meeting where you’re updating a stakeholder or pitching somebody on a new idea. The pace of that communication feels lethargic, slow, backwards.

Calendar as Liability = Inventory as Waste

A remarkable change of attitude in manufacturing in the last century was the idea that inventory is waste. This is the central insight of the Toyota Production System, sometimes called “lean manufacturing.” Before this, inventory was an asset on the balance sheet, quite literally. Warehouses would sit full of parts, meaning the company was prepared, capitalized, ready. The inventor of the Toyota Production System saw it differently. He saw it as waste. It ties up capital, hides quality problems, and creates obsolescence. So the goal became not to have a full warehouse of inventory but to have an empty warehouse, which created efficiency.

We’re about to experience that same inversion with the calendar. Meetings hide the absence of output. Meetings are a buffer against accountability. Maker time exposes who actually ships. We’ll see this shift in the coming months, and it won’t be pretty for the generation of people who believe that meetings are the central way they show economic value in the world.

Prestige as a Maker

In this world, managing people is not a measure of success. “Principal Engineer” and “Distinguished Architect” are examples of makers who enjoy genuine prestige in these roles, but there are precious few people in our current job market who get to enjoy roles like that.

The independent creator, the solo founder, the consultant who ships: all are going to skyrocket in popularity because the corporate world, built on “Manager Time,” can’t figure out what to do with all of these makers that will arise in the Claude revolution.

The manager class are people whose identity is built on coordination, access, and information-routing. They are watching their core functions automate. These aren’t bad people or useless people. Many of them are skilled at things that genuinely mattered: translating between technical and business contexts, maintaining organizational coherence, mentoring through proximity. But the ground is shifting beneath them.

There are people I know who are struggling with this transition. They thrived in meeting culture because they were genuinely good at the human coordination that meetings enabled. Visibility, facetime, stakeholder management: these were skills, hard-won and real. In maker time, those skills depreciate.

Maker time requires discipline, self-direction, and tolerance for ambiguity. Not everyone has these traits and not everyone can develop them. I foresee a whole lot of people who trained themselves carefully to manage makers starting to understand their obsolescence in the coming months. Skills of navigating politically between stakeholders and coordinating synchronicities in capitalism will seem so obsolete in the face of the raw creative power of a good maker.

My calendar is nearly empty. I’m doing the work of four people.

I don’t think I’m exceptional. I think I’m just a little early.

In other words, when any team anywhere in the enterprise feels a burden of work, it will be easier to deepen their expertise in Claude Code than to add another person.

I should say upfront: I’m a turbo-normie in the world of Claude Code. I’m not building revolutionary software. I’m not redesigning processes from scratch as some great architect of the future. I’m definitely not researching how to make better AI.

I use Claude Code for operations, for sales, for building dashboards, for updating stakeholders—a million small tasks that, although dry and boring, are incredibly important for the churning billions of lines of code being written by AIs to interface with the world that they were already leaving behind.

And yet even from this vantage point, I struggle to imagine a world where I would need to hire somebody to do knowledge work given the depth of the tool in front of me. Hitting context limits is one thing. Understanding the limits of my own capacity and labour is another.

When Doug O’Laughlin of Fabricated Knowledge wrote recently that everything is becoming a skill issue, he’s right. But the skill issue goes beyond human proficiency with AI; it’s also about velocity. Consider how much better people will get at Claude Code in the coming weeks and months. The gap between those who adopt now and those who wait will compound.

This feels like the first concrete evidence of exponential acceleration I’ve experienced in my career. I’ve touched moments like this before, but never with this clarity.

This takeoff will leave behind millions, perhaps billions, of knowledge workers who won’t adapt to these tools fast enough for their labor to have utility in the new economy. I know how that sounds. But if you’re working in these platforms today, you need to stare this down. Similar to my last post: we have an obligation to teach people we like how to use Claude (and its progeny) properly, so they can continue to grow with us as we enter this new world.

This is a fundamental upset in how we think about work, how we grow teams, how we value labour. If everything is a skill issue and there are still only 24 hours in every day, the calculus between hiring someone and simply getting better at Claude Code has shifted in ways we’re only beginning to understand.

Software in November 2025 was the religion that I worshipped, but I lacked the wrinkles on my frontal lobe to ever be a priest. I witnessed the clergy tap out their incantations, summon the Holy Spirit, and yet I could never join them. I paid my tithe, I mumbled the latin, I was a devout believer, but I did not lead the worship.

Claude changed that. By Christmas of this year, I found myself ideating, planning, writing, executing, and delivering all by myself.

It’s hard to overstate just how infuriating it can be to be non-technical in the software engineering world. To have baseline skills but never be able to really dig in to any project. To endlessly search for collaborators and things that you believed in but did not seem obvious to technical people. And then at the same time to push projects that you knew were obvious to technical people but not useful to users or other people in the world. To seek to find space around the edges where you could participate knowing that the warm center was forever something you needed another person to interpret for you.

I remember in my very first start-up, in the endlessly patient engineers who would explain to me very simple ideas of software engineering like servers, front-end and back-end, not knowing that I would ever understand them properly but glumly tucking me under their wing. Those days are long since past, but just because I understood the principles that brought good engineering into the world did not mean I could ever participate in that process.

To their endless credit, software engineers seem to have precipitated this revolution all on their own. They neither tried to jealously guard their secrets, nor did they attempt to monopolize access to these tools. This revolution is by the Church for the people, and I’m endlessly appreciative of it.

Spreading the Gospel

Over the next few months, something unprecedented will start happening in America, spreading from San Francisco outward into cities, groups, communities all over the states and eventually the world. That will be Claude Code or its close followers.

We are now at a moment where only yesterday, Claude launched a platform to be able to do this kind of thinking without having the benefits of the terminal and all of the friction involved in setting that up for the first time. These tools will drop like an atomic bomb on our economic model that has sustained us for 45 years, and I expect the changes to be monumental. I don’t make predictions about the future of the economy or of work very often, but I have real conviction about this one. This is the first time that you will “feel the acceleration” from AI. It’s here.

I believe very strongly that there is an opportunity here for everybody to smooth this transition and to make the principles of good software engineering abstracted from the principles of coding to show people how to use these tools to maximum effect to make people understand how to make something secure, something fast, something useful, and something instinctually easy to use. In other words, good software.

If you are a good engineer, it is incumbent upon you to spread the gospel in the coming months. To show people how to do your métier well, not just have them produce endless amounts of sloppy stuff. If you’re in media, you should be writing tutorials for this. Not only because people will want them, but also because it will help improve the software that runs our world. It’s important to do this work.

It’s also important to reach out to the 5 smartest non-technical people you know and teach them how to use Claude Code. It may be some months or years from now, but you won’t need to think about the good software engineering that goes into making a good piece of software. Today, you most definitively do, and for Claude to do a good job on the code is nearly inevitable. But for the applications themselves to be well-designed, fast, easy to use, and useful, we will need to spread this. We will need to have ways of teaching people how to do this, and most importantly, we will need to teach the right people how to do this.

As a non-technical person, you don’t spend your days on Hacker News tracking trends, you don’t spend your days seeing GitHub repos and how many stars they have. You can very easily spend six months drilling down on something that is equally non-technical as you and miss this big revolution.

So this week, if you are a good software engineer, take an hour and message the 5 people you know who are most deserving of these tools and spread the gospel of Claude.

Today the clever folks over at AI2027 pushed back the timeline to escape the permanent underclass by 7 years, so if we’re going to speedrun this industrial revolution in 10 years instead of 3, I’ve decided we need to get better at opposing it.

Bernie Sanders wants a moratorium on data centers. Social media is rotten with unsophisticated, banal takes on the evils of AI: their water waste, their thievery, their wretched sloppification of the digital world. This is our Luddite moment.

I don’t mean that as an insult. The original Luddites were right about a lot of things. They were right that the new machinery was producing inferior goods; slop in atoms, not bits. They correctly identified that the economic benefits were flowing to capital while workers literally starved. They correctly identified that the government was captured by manufacturing interests. Their analysis was sound.

They still lost.

Byron is so cool, part 1

In February 1812, Lord Byron gave his maiden speech in the House of Lords opposing the Frame Work Bill, which would make machine-breaking a capital offence. He’d just gotten back from seeing the violence himself, where Luddite workers were destroying the machines that lost them their jobs and caused their children to starve. He was 12 days away from the publication of one of the most important works in english (and one of my all-time-fav poems) Childe Harold’s Pilgrimage. But before that fame, before he became a household name, we he stood up to “commit political suicide on behalf of working men.”

Two passages from his speech could be delivered today with one word changed. Swap “frame” for “AI” and you have a perfect contemporary argument:

“By the adoption of one species of [AI] in particular, one man performed the work of many, and the superfluous labourers were thrown out of employment. Yet it is to be observed, that the work thus executed was inferior in quality; not marketable at home, and merely hurried over with a view to exploitation. It was called, in the cant of the trade, by the name of ‘Spider work.’”

The rejected workmen, Byron continued:

“In the foolishness of their hearts they imagined, that the maintenance and well doing of the industrious poor, were objects of greater consequence than the enrichment of a few individuals by any improvement, in the implements of trade, which threw the workmen out of employment, and rendered the labourer unworthy of his hire.”

That’s it. That’s the whole argument. The workers are “foolish” for thinking their welfare matters more than efficiency gains that accrue to capital.

Why Luddism fails

The Frame Work Bill passed. Seventeen Luddites were hanged at York in January 1813. The movement was crushed by 12,000 troops, the factories continued.

The structural reasons are the same ones that will defeat the data center moratorium and the pause campaigns:

1. Economic power is real power. Capital wins if it doesn’t need labour.

2. The state aligns with capital. Regulatory capture isn’t new.

3. Negative opposition offers no vision. Stop the machines. Then what? You cannot beat something with nothing.

4. Repression works when systematically applied. The state has more resources for violence than you do.

I believe Bernie’s moratorium will fail. The hysterics of TikTok will move on from AI into some other intractable conflict and the attentional moment will pass.

Byron is so cool, part 2

Byron’s direct political engagement died with those men in the gallows, killed for breaking a machine that broke them. His methods for critiquing the soullessness of his industrial revolution changed.

Childe Harold created the Byronic hero: the brooding outcast, disillusioned with society, seeking meaning in authentic experience. He started answering questions about meaning; showing the world what the factories cannot give them.

The Romantic movement succeeded where Luddism failed because it offered something positive, and filled the empty void that ruthless optimization leaves in the human soul. They wrote the book on critiquing technological civilization, and they did it better than anyone in our contemporary moment.

Byron is the individual who walked the Luddite philosophy into meaning, finding a deeper direction and birthing so so much beauty.

Novo-Romantic hero applications

The romantics provided some kind of opposition to the cursed horrors of the industrial revolution. They articulated everything that can’t flow from a factory- nature, authentic connection, meaningful work, spiritual depth. These values survived because they named something real. We still need them, and if the good folks of AI2027 are right, Claude isn’t giving you these anytime soon.

I look around and I can’t for the life of me see our Byron, our Wordsworth, our Keats, our Romantic Heroes. This may be because the culture around AI is so profoundly captured by the utilitarian bullshit of EA or the culty programming of the Rationalists, but I yearn for them. So does Tyler Cowen, apparently, quoting from his week-old venture New Aesthetics:

So here’s a list of the roles I want to fill in 2026, in no particular order:

2026 William Blake: A Prophet of Meaning

Blake gave us “dark Satanic Mills” in 1804, a phrase so durable it remains shorthand for industrial dehumanization centuries after his death. Blake was kooky, prophetic, esoteric, and had a vision for a Jerusalem, to be built in England’s green and pleasant land through “mental fight.” He saw the catastrophe of hyper-optimization 200 years ago, and it still resonates with me:

“The same dull round, even of a Universe, would soon become a mill with complicated wheels,”

Against abstraction, flattening, and the reduction of existence to mechanism, Blake offered prophetic pagan mysticism. A new Blake sees the darkness in token prediction, and seeks a spiritual path to meaning.

2026 William Wordsworth: A Prophet of Nature

I have spent some of the best months of my life rambling in the Lake District, the first-ever national park on earth, created through his deliberate glorification of this area in verse and song. I feel deeply this kind of curative connection to nature is still under-appreciated in the AI age- we will yearn for places that tokens cannot take us. Against the slop economy’s infinite scroll, Wordsworth’s discipline remains radical: look at the daffodils. Really look. A 2026 Wordsworth fixes our attention, and reminds us of our mammalian cravings.

2026 Percy Bysshe Shelley: A Theorist of Cultural Power

“Poets are the unacknowledged legislators of the world,”

Shelley and Mao stand in opposition- does political power flow from the barrel of gun, or from the cultural foundations it rests upon? Change the imagination, change the possible. The alignment researchers model extinction probabilities. Shelley would ask: what vision of flourishing are we offering? What “gigantic shadows which futurity casts upon the present” are you articulating? Negative probability assessment is necessary but insufficient. Someone has to write the future people want to live in. This is my fondest hope for the Abundance folks; fill this role, I beg you!

2026 Mary Shelley: The First Shoggologist

Frankenstein invented science fiction by asking “what happens when it all goes your way, technologist?” There has never been a better time to ask ourselves what happens when you create something intelligent and refuse to raise it right. Apply here if you’ve decided to walk from the Rats but haven’t figured out where to go next.

Lord Byron: A Prophet of Action

Byron started in politics, created a cultural tradition that endured centuries, and died fighting a revolution in a foreign land. If you believe political opposition to your own system fails, redirect resources toward causes where they might matter. Apply here if you feel you are a great talent and you have something to contribute before the arrival of the Machine God.

From “Ode to the Framers of the Frame Bill” (1812):

“Men are more easily made than machinery— / Stockings fetch better prices than lives— / Gibbets on Sherwood will heighten the scenery, / Shewing how Commerce, how Liberty thrives!”

2026

I pray we’ll see past our Luddite moment. The Bernie moratoriums, the TikTok rage, the scenario planning conclusions are our Frame Work Bill speeches. They will fail for structural reasons. The question is: what we build when they do.

I think 2026 will see the birth of the New Romantics. Not a movement against AI- that’s Luddism, and Luddism loses. A movement that articulates what AI cannot provide. What embodied human life offers that disembodied intelligence doesn’t. What attention and particularity, meaning require that tokenized optimization cannot deliver. If we’re not consigned to the permanent underclass for a few years still, lets leave behind some culture that rages and burns with our human desires.

I’ll leave you with a little Tennyson, who I neglected to add here due to his unfortunately late birth:

Come, my friends,

‘T is not too late to seek a newer world.

Push off, and sitting well in order smite

The sounding furrows; for my purpose holds

To sail beyond the sunset, and the baths

Of all the western stars, until I die.

It may be that the gulfs will wash us down:

It may be we shall touch the Happy Isles,

And see the great Achilles, whom we knew.

Tho’ much is token, much abides; and tho’

We are not now that strength which in old days

Moved earth and heaven, that which we are, we are;

One equal temper of heroic hearts,

Made weak by time and fate, but strong in will

To strive, to seek, to find, and not to yield.

This tweet corresponded with a moment in my year where I had recently had much of the trajectory of my future forcibly altered for the worse, and for whatever reason it grew from an amusing little idea into A Big Plan.

I’ve been to some great things at NeurIPS over the time that I’ve been going, but I have almost never been to a great party. NeurIPS at its most crude level operates as a wrapper for Luma; especially this year when the conference organization was faltering and the conference app was catastrophically bad.

Starting this June I had the idea that maybe the best way to ensure an actually good party at NeurIPS was to do it ourselves, and Model Ship was born.

Afterwards, I was asked by some friends in events & AI about what makes a good party, here’s my list.

Ten Commandments of throwing a good party for AI researchers

1. Thou shalt host, not merely organize.

2. Thou shalt not suffer tranquilizing forces.

3. Thou shalt have within a space to rage, and a space to play.

4. Thou shalt nurture thine guests curiosity.

5. Thou shalt coax forth the meek with low light.

6. Thou shalt deliver highbrow culture & lowbrow music.

7. Thou shalt not push, only pull.

8. Thou shalt be generous in guests, and parsimonious in sponsors.

9. Thou shalt, like Noah before thee, gender balance.

10. Thou shalt not suffer Luddites nor Yuddites.

Thou shalt host, not merely organize.

Your job is to make certain from the moment that anyone steps through the door, they feel welcome, excited and relaxed. Great parties start by relaxing everyone quickly and then building their energy back up as a great motive force.
Don’t confuse hosting and organizing; hosts introduce you to new friends, make sure your immediate needs are met, and set the vibe- organizers make sure that the bathrooms stay clean. These are different roles.

Thou shalt not suffer tranquilizing forces.

If you serve exclusively oily carbs, people won’t want to dance. If you have big sofas and cuddle spaces immediately available, people will park themselves there. A great party circulates, undulates, moves and shimmies. Likewise, don’t invite literal tranquilizing forces into your party- a party is not tranquil. This can also apply to anything that could drain energy from your guests- even little things like bad temperature control, lack of air circulation, lack of bathrooms, etc. If you want the party energy to build to a fun night, don’t let it be tranquilized. This seems stupid to have to repeat it, but take it seriously in your planning and you’ll be rewarded.

Thou shalt have within a space to rage, and space to play.

A great party has a dance floor, period. The single best predictor of your party being fun is if people danced- not just because dancing is fun, but because dancing is a natural product/end game of a fun party.

AI people are not going to show up and bust out like moves like they just wrapped a Broadway musical. They need to make their way into that energy- they live in front of screens, their minds in latent space. They must be encouraged back into the soft animal bodies they seek to transcend, and this is done by having a playful relaxing space to begin the night in. Fill it with things that will make them smile, with things that will make them play, with things that will make them feel special.

Thou shalt nurture thine guests curiosity.

AI people get bored faster than any cohort of humans who have ever lived. They exist at a tempo of slack messages, model releases, zoom calls and paper deep-dives that requires them to process 10x more information per minute than a conversation normally provides. Each one of them has access to that firehose of information and dopamine at any moment they desire; so make room for curiosity at your party, and your hosts must stalk through the party for the first hour and immediately charm anyone who has their phone out.

Thou shalt coax forth the meek with low light.

SF loves to light parties like surgical wards. Turn the lights down, please. Light small little lights and candles, sit in the chairs and lean on the counters the night before and if you are squinting, you messed up. Please, for the love of god, light your party like everyone who is coming is interested in looking good and laughing, because they are.

Thou shalt deliver highbrow culture & lowbrow music.

Bring some culture to your parties. In case you haven’t been on the internet outside of TPOT, most art/culture people are growing to hate us. Bring these worlds closer together- SF has had many moments in its history where forces of tech and art stood united in optimism for a better future- cultivate a bit of art, a bit of culture. Invite someone to perform, to exhibit, to do something with AI art, to remind us of the broader cultural forces.

The one narrow window of culture that AI people do cultivate tends to be music, but please don’t take too many risks here- you might like Chilean minimalist techno while you code, but the rest of us want to flirt with a cutie, dance with our friends, and sing along to the Great American Pop Cannon.

Thou shalt not push, only pull.

I have been to so, so many parties where they push a vibe onto guests- in October I went to a robot fight night and there were lectures about AI safety and they hushed us every 2 minutes so that the speakers could be heard through their terrible mics. This is a mortal party sin, The Cruciatus Curse of a vibey party. Do not push any vibe onto people- pull them in. Invite them into something you want them to experience. Expect that half the people there want to continue their conversation and just chill. If you must push, do it with class- think when the intermission ends at a play. Don’t ruin your party vibe in service of a marketing message, a sponsor, a call to action, or a KPI. You know those Dwarkesh ads where you don’t realize its even an ad? Bring this energy.

Thou shalt be generous in guests, and parsimonious in sponsors.

Invite ‘nobodies’ to your parties. Be sincere in the invites, make them feel welcome, and let them know that you’ll host them. There are a lot of humble, cool, fun, adventurous people in the Bay, and they don’t all work in leadership at frontier labs. If you invite the AI equivalent of “I arrived here on the Mayflower” then there will be a familiarity that won’t spark curiosity and spontaneity. For our NeurIPS party we made literal “Golden Tickets” like Charlie and the Chocolate Factory. We gave them to great people we met at NeurIPS, and a whole bunch of them came to the party.

If you want or need sponsors for your party, pick carefully and don’t pick more than 3. This goes for corporate sponsors as well as individuals who are paying, donating space, or whatever. Each time someone commits resources, the circle of people who are not guests who must be kept happy grows. This means you might have to make a decision thats not in the best interest of your guests, which is generally speaking antifun. Pick them carefully, update them regularly, and make them look good.

Thou shalt, like Noah before thee, gender balance.

Contrary to what X might tell you, AI is full of cool women. Invite them to your party. If you don’t know enough women to gender balance your party, sounds like you might be the organizer, not the host. Parties are fun when they are gender balanced; not because of some seedy reason, but because that is the recipe for the most guests feeling comfortable when they arrive. Comfort early means you’ve done like 60% of the work to have a good party. Note that this means that your guest list pre-party should be 60+% women, as generally speaking no-shows are slightly higher for women than men.

Thou shalt not suffer Luddites nor Yuddites.

AI people are not going to have a fun time if they feel like they’re being judged by either Rats or Troglodytes. We work in AI, we spend a lot of our spiritual energy defending our life’s work to these two camps. Parties should refill this energy, not drain it. There are great places to have those convos, like debates and conferences and literally anywhere but the space you’ve specifically created to let people relax, unwind, let loose and have fun.

The Big MS2025 thank you-

Michelle was my co-organizer for Model Ship, and because she doesn’t read substack she’ll not be embarrassed by me saying how excellent it was to work with her. She and I hired 12 artists to make custom art and swag, huge teams for music, AV, food, drink, etc. and the whole process was easy and fun. She remembered everything, took care to treat our hosts and sponsors well, and made sure that nobody crashed the party. Michelle, you’re a fucking queen.

Even More Beyond

We’ll be back with more Actually Fun Parties For AI Researchers in 2026, alongside the return of 10 minute researcher interviews (we did 211 at NeurIPS 2025!). Can’t wait to see y’all back aboard.

Nathan Lambert published an excellent talk on open models and the ATOM Project this week, and I was very interested to see the Safety People Questions afterwards.

Safety People have become a fixture of the bay area AI scene. A lesser derivative of the “thinkboi” (although they come in multiple genders) they are a constant presence in the zeitgeist of AI in SF.

Young, brilliant, and exposed to The Sequences in their teenage years, they earnestly want to save the world from extinction at the hands of AI. They are often deeply technical, very serious, and believe that the key to preventing the AI apocalypse is some combination of research and thinking (and possibly missile strikes on datacenters).

At the end of Nathan’s talk on open models, a line of Safety People formed at the mic. They took turns asking essentially the same question: aren’t open models dangerous? Each one heard the person before them ask this question, but somehow were unable to pivot to another question when their time came at the mic. Gracious as ever Nathan answered them patiently, but this moment really exposed for me the rot at the heart of the safety discourse in AI:

There are not that many ideas on how to get the message out.

The research? The research is amazing, and I support it fully. The problem I am pointing out here is that the insular nature of the AI safety community means there are only a few ways of raising awareness of perceived risk- getting on mic on a recorded lecture seems like an important one to that group.

a modest proposal

There has been one big project in growing the tent of AI safety in 2025; a book called “If anyone builds it, everyone dies”. I see this book as a great representation in all that has gone terribly wrong in Safety People World; a hysterical, wild-eyed and attention grabbing shot at capturing the attention of America before their apocalypse.

I admire that they went there- someone had to. I do think that now that this final step has been taken (surely we can’t go further- “if anyone thinks hard about it, everyone dies?) we should start thinking seriously about how to actually get people to give a shit about AI security risks. This process must start, inevitably, in everyone stopping the use of the word “safety”.

Safety is not interesting. Safety doesn’t inspire. We use LLM-guided recsys to scroll TikTok during workplace-mandated safety videos, and we certainly don’t read books about any other kind of safety for enjoyment. Airports aren’t stocking books about seatbelts and airbags, and even the most scary events of our time don’t precipitate a deep global discussion of safety (see: the pandemic).

The Carl Sagan of AI

The title of America’s Beloved AI Educator is still a jump ball. Dwarkesh is hitting millions of views on his podcast interviewing the luminaries of AI, but he also spends a bunch of time digging into Russo-Japanese naval battles and biotech. Karpathy makes incredible videos on fundamentals of LLMs, but not the kind I can send to both my grandmother and my 13 year old cousin.

We had Oppie for the Atomic Age, and he did more than anyone in advancing ideas in how to make this terrible new technology safe. We had Sagan for the Space Age, and alongside people like Feynman we talked seriously about safety and improved vastly the culture of NASA around this issue (some say too well).

In all these cases, we were in a very similar bind- a group of scientists were driven by zeal and American gumption to birth a new world. The average American had nothing to do but sit back and watch as the marvels were birthed- abundant power, incredible technologies, breakthroughs left and right.

In so many of these prior technological revolutions, the Sword came before the Ploughshare. We made integrated circuits to precision-bomb Vietnam, we dropped Atomics before we warmed our houses with their electricity, and we certainly worried more about the cold-war race with Russia in the space program than about the civilian benefits. DARPA gave us the internet, even the Iron Age was about military application.

In AI, we have the peaceful technology before the military application (coming soon, no doubt). This means that the average person encounters AI with some Hollywood trepidation but quickly learns to trust this technology (indeed, trust it with their most intimate secrets, their government, and their privacy).

This has put AI safety in a deep bind;

1. the technology doesn’t seem destructive, the way atomics were

2. the fear doesn’t motivate when you don’t have a “how to act”

3. the best place to learn about AI is by using AI.

If the Safety People want to win this argument, they need ways to get to the question of AI safety without using the word “safety”, and certainly without having their hands waved in the air in panic. They must seek to take up the mantle of Beloved Public Educator, and through the work of helping America understand this new miracle, they can guide the discourse on safety as they see fit.

“Security is a process, not a product.”

-Bruce Schneier

If I were sitting in an Eames Chair at Open Philanthropy, sipping on a cup of Yorkshire Gold Tea with my Gaziano loafers kicked up on the Ottoman, I would be asking myself one question: How can I make the Carl Sagan of AI?

By funding safety people to not use the word safety. To talk about the opportunities of AI, the incredible technology behind its development, to sit with practitioners and understand the technology deeply, all without the lectures, the navel-gazing discourse, and the hand-waving hysterics.

The next and obvious step in avoiding the Great Risk of AI is simple; learn how to talk about AI Safety without ever saying “Safety”.

This week I attended both a Manus and a Genspark event in SF, two companies who have a big thesis on replacing Microsoft as the tool for presentations.

Here’s who I tried, and then I’ll show you results:

• Manus AI

• Genspark

• Kimi OK Computer

• Kimi Slides

• Gamma

• Beautiful.ai

• ChatGPT

• ChatGPT Operator

The Prompt

You are a senior producer creating an 8-slide investor deck for an off-site content studio during NeurIPS week.

Objective: {capture long-form 60–90 min interviews + 10-min poster interviews; deliver 45-sec shorts near real time; archive long-form for later}

Funding ask: {range and exact ask}

Dates and location: {Dec 2–7, 2025; San Diego Convention Center area}

Daily ops: {8 hours per day; rotating hosts {names}; target throughput {X long-form, Y quick interviews}}

Targets: {sponsors a16z, Susa Ventures, others}

Value prop for sponsors: {brand exposure, opt-in leads, on-site presence, rights to repost, KPIs}

Compliance: {press accreditation when filming on-site, off-site studio is primary; adhere to NeurIPS Code of Conduct; avoid official branding unless permitted}

Tone: Spartan, professional, concise. No em dashes. Use numbered slides with 3–6 bullets each.

Output format: 8 slides titled “Slide 1 — …” through “Slide 8 — …”.

Must include: clear $ ask, programming overview, workflow to shorts, sponsorship tiers with deliverables, budget table with lean vs full, logistics and risk, team, timeline and CTA.

If browsing is available, verify NeurIPS dates and press policies before rendering.

I set this up to have opportunities for the product to have to do a lot of different work:

• logo finding

• budget making

• respecting restraints (no emdash)

• output format

• workflow diagrams

• browsing

• dealing with incomplete information

• error in brand: I wrote a16z not A16Z, seeing if corrections would be made.

Time to Complete Task

Every app was between 5-10 minutes, except for Operator, which took 25 minutes.

Every app worked but Decktopus, which failed spectacularly.

Additional Selections, Extra Steps

If requested, I defaulted to the base settings for every step. If I was prompted to select a style, I took anything called “default”.

Results:

No product here is great- yet. All show significant progress, but these are in no way ready for prime time with this type of short prompt. There are some decent attempts in here, so I’ll rank them and show results for each product.

Decktopus: F Tier

Decktopus served me a LOT of ads on Google while making this project, and yet their app crashed and burned upon testing. Was it me? Was it them? Was it just an off day? We’ll never know, as I will never use this product ever again. Their first slide was called “An error has occurred”- brilliant.

ChatGPT: D Tier

ChatGPT should crush this task, in theory- it has by far the most context for me and my work, and yet I give it a D tier despite a loaded memory and a Plus subscription. This is wildly below the bar of other companies 100x smaller than them.

Good:

• very accurate: GPT5 instruction following is excellent as always. Very few errors.

• sources: provided sources for claims it made.

• GREAT planning: the best plan in terms of execution possibility.

Bad:

• just chat, reworked: zero effort to make this a slide deck: its an exported chat, and there’s nothing even approaching a slide deck you could take to a client or meeting.

Chatgpt

8.15KB ∙ PDF file

Download

Download

Manus: D Tier

Manus has built me many a slide deck, and I know its quirks and desires. Knowing how to push this particular super-agent means I can easily produce a pretty good slide deck: for instance if you ask it to code you SVG icons, you get pretty icons. For this task however Manus really didn’t do well.

Good:

• great planning: did a great job planning out the details of the project

• good text emphasis: Manus knows how to make the important parts of the presentation bold- by far the best effort.

• decent icons: even without prompting the icons are good.

Bad:

• PDF Export: if you ask Manus for a deck exported as PDF it gives you this awful long page with blank space. Should be a solved problem.

• Weird Text: what is this Elvish they added to my slides?

  

• Photos Unrelated: Manus fetches photos from the web for its decks, but more often than not they are awful. Example:

  

• Weird Phrasing: Funding ask: $175,000-$225,000 with exact ask of $195,000 is a very strange way of asking for 200k.

• Errors in Images: the plan is to have live interviews, but the image it pulled up is a zoom call. On one slide, there’s like a chart of printer costs? nonsensical.

• Fake Email: Faked an email address never provided.

Manus Ai

3.15MB ∙ PDF file

Download

Download

Kimi Slides: D Tier

I want this product to be good! When a scrappy Open Model lab builds tools on their own models, I want it to be great. Kimi slides asked me to pick a template and their was no default, so I just clicked the most neutral looking one.

Good:

• Nicer Design: Kimi gets points for having some fonts and design choices baked into the template that I didn’t hate.

• Fine word selection: Kimi did a fine job making this project sound credible: Kimi is good at wordsmithing and good at choosing effective language.

• Good icons

Bad:

• Poor instruction following: got way more than 8 slides, didn’t follow naming, didn’t follow rules of prompt.

• Weird empty gaps: plenty of slides that are just.. empty looking? I can just delete at least 3 and the deck gets better.

• Errors: Dates, emails, the fact its an “investor deck”- just lots of errors.

• Inconsistent Design: On one slide Kimi nailed the timeline- on another, the line through the timeline is misplaced.

Kimi Slides

677KB ∙ PDF file

Download

Download

Kimi OK Computer Agent: C Tier

Kimi released an Agent today, and although it failed at making a PDF deck (said it made a PDF, but link was broken, and requested fix it didn’t do it) I screenshotted it’s work and made a PDF of the outcome, because the work is interesting.

Good:

• Pretty: Kimi made this really pretty, despite not delivering slides in the right dimensions.

• Great Planning: best planning by far.

• Real Schedule: Kimi made a daily schedule. Great stuff.

• Consistent Numbers: the early proposal numbers are consistent.

Bad:

• Confused about Biz/Project: talked about an equity offering? Kimi is startup-pilled.

• Made up Team: like many of the next few offerings, team was hallucinated

• Basically a Website: clearly this agent wants to make websites, so a lot of this deck looks like Lovable slop.

Kimi Ok Computer

2.1MB ∙ PDF file

Download

Download

ChatGPT Agent: C+ Tier

ChatGPT agent had a leg up on competition: not only did it take 5x longer to produce the deck, i used my account with full memory of what I work on and my preferences and context. Fully loaded memory.

Despite this, the output is mid.

Good:

• Actual Deck: unlike the non-agentic chatGPT, this is an actual deck.

• Instruction Following: GPT5 doesn’t miss.

• Icons: great icons.

Bad:

• Hideous Graphic: look at this monstrosity

  

• Inaccuracies: Weirdly this was less accurate than the previous Chat entry. Why? We may never know.

• Cut-off slides: Many slides are too long and kind of cut off.

• Weird Sources: the sources are broken.

Chatgpt Agent

643KB ∙ PDF file

Download

Download

Genspark: B- Tier

Genspark delivered just what Wen, their founder, says it does: 70% of a good deck, that you have to finish yourself.

Good:

• Pretty Solid Planning: The plan the super-agent made is good. Its not groundbreaking, but it all makes sense.

• Nothing “Wrong”: Besides making up an email and URL, the whole thing is ready to be made into a deck. There’s no part that jumped out as obviously innaccurate.

• My Editing Stack: I use Descript and Resolve! Genspark guessed right.

• Intuited tool: I didn’t ask to use their slide product- just took me there based on prompt. Seems obvious, but others didn’t do this.

Bad:

• Drab: no big risks like Manus, but no design really.

• Export Blocked: I had to pay to export (only service where I could not use a free trial).

• Minimal Design: Compared to Kimi OK Computer, there’s zero design. Its all correct, just very spartan. I feel like I got blueprints, not a house.

Genspark

535KB ∙ PDF file

Download

Download

Gamma App: B Tier

Gamma App is our first “Slides only” tool- this is ALL THEY DO.

Good:

• First Handsome Deck: This deck is cute! its colourful, the images work well, there’s a style to it.

• Solid Planning: There’s nothing here that isn’t out of distribution in terms of planning.

Bad:

• Branded My Slides: big ol’ Watermark.

• Incorrect Dates: Asked me to start the plan in August, last month

• Slop Vibes: it smells like Slop. Its cute, but if you have worked even a bit with AI, you can feel the radiation of the Slop through the whole thing.

Gamma

1.37MB ∙ PDF file

Download

Download

Beautiful.AI: B Tier

This is our second “Slides only” tool- and it shows. This one impressed me 7/8 slides, until it hilariously hallucinated a whole team. I was ready to punch this straight into S tier, until I saw this hilarious collection of fake individuals with a promised timeline that never appears:

Good:

• Looks Actually Great: some of these slides are presentation-ready. Solid.

• Sexy Fonts: Half these tools pretend that fonts don’t exist. They nailed it.

• Nailed the Icons: Great icons. best of the batch.

• Finally, Logos: it found logos! with transparent backgrounds! Super impressed.

Bad:

• Team Slide Fail: Compared to the rest of the deck, this is hilariously bad. The photos are even like clearly real people that its stolen from the internet, with cropping issues, colour issues- the whole gambit. How it made amazing AND terrible slides is a real mystery.

• Slop Conclusion: For a casual reader the last slide is fine, but if you deep dive it even a bit its god-awful.

• Math ain’t Mathin’: The math on their fundraising is not… well… not great.

Beautiful

3.68MB ∙ PDF file

Download

Download

Parting Thoughts

This was a fun 2h of tool use testing- we’re far from great, but i expect by next year I’ll be making good decks from prompts. Hats off to the people who are hard at work on this problem; keep going! We’re kind of close!

Do you want to do NeurIPS.tv?

If you’re into the project that these slides are about, reach out: caithrin@caithrin.com or @caithrin on twitter.

It is very likely that sometime in the next 3 years, we will return to being a species without a single living member who has ever left our planet’s orbit.

Its kind of amazing that it hasn’t happened already. The 5 surviving humans who have left the orbit of our planet are 89 to 95 years old, and Earth could hold eight billion people and not a single living witness to deep space very soon.

The usual reply is that Artemis is coming. If the schedule holds, four astronauts will loop around the Moon in 2026. I hope it happens, but given the state of affairs at NASA I am not optimistic.

The internet unlocked our century but the physical world has not kept up, as Peter Thiel has been telling us for two decades. We now measure all progress against the potential progress that equal effort applied to software systems could have yeilded. The result is a kind of weightlessness. We float above the tasks that make a civilization real. Software has just too damn much leverage, it broke our sense of progress.

This is different than supersonic passenger aircraft or nuclear buildout; its a 50 year stoppage both civil and military, where the best of us left our planet behind and then became old men and died before anyone ever tried it again. It makes me heartsick to think about.

Dan Wang’s book has some very eloquent things to say about this, how society can prize process or output and the consequences of this; a good frame to hold my sadness about this moment.

If Artemis II flies in 2026, four more humans will cross the boundary our species has not crossed since the early 1970s. If Artemis III follows, we will plant boots on the moon after a lifetime away. That will not solve every other problem, but it will signal that the world of atoms is still negotiable.

Until then, the image lingers. A planet with no old men who can tell you what it felt like to watch Earth set behind the Moon.

Next Month

White House outlines a “Strategic Technology and Compute” program that converts select subsidies into equity across chips, cloud, and AI infrastructure. Pitched as a sovereign wealth fund a la Norway. Biggest advantage is grid access.

Nov 2025

Omnibus appropriations include a rider authorizing equity-for-grants swaps in “critical digital infrastructure.” Treasury and Commerce get reporting duties to Congress on positions and risk.

Q1 2026

A high-profile AI misuse incident triggers democrats wanting hearings, and get unexpected admin support. DHS designates frontier model training and deployment as covered critical infrastructure, giving access to federal compute credits to safety rules and pre-release testing under guidance.

Q2 2026

Congress passes the AI Safety and Compute Act. Training runs above a defined FLOPs threshold require a federal license, with eligibility conditioned on access to subsidized compute and convertible government warrants in licensees. Some companies refuse; they lose grid access and get stuck in permitting.

Q3 2026

DoD invokes Defense Production Act to expand domestic advanced packaging and AI-grade power infrastructure. Funding for compute capacity comes with equity options in compute suppliers. Inititally these options are not exercised, but during big news weeks they are quietly made.

Nov 2026

Midterms deliver narrow majorities for a national tech-security agenda, fueled by the fears from the earlier incident. Congress authorizes a United States AI Authority to supervise licensed training runs and manage federal stakes tied to compute and safety infrastructure.

H1 2027

Regulators secure “golden share” vetoes in licensed frontier labs for release decisions above the FLOPs threshold. Treasury aggregates minority positions across chip, cloud, and model layers to coordinate policy. These are enriching to admin officials, who sell secondaries.

Q3 2027

Capital stress at multiple labs meets rising training costs. The government offers loan guarantees, discounted power, and priority access to DPA-backed compute in exchange for convertible preferred equity and release-governance covenants. Folks who avoid this get red-taped out of datacenter buildouts.

Q4 2027

Congress passes the Strategic Models Act. It creates a National Frontier Model Program and authorizes majority acquisitions of designated Frontier Model Operators. Statute mandates a government-owned, contractor-operated structure for day-to-day management. Chinese models are banned, forcing thousands of businessess to switch off the open stack.

Q1 2028

Treasury exercises conversions and conducts a tender. Government control crosses 51 percent in two labs and achieves control rights via a golden share in a third.

Q2 2028

Administrative integration. Equity of prior private investors is cashed out or converted to non-voting shares. Staff receive retention contracts under a lab-operator model.

Oct–Nov 7, 2028

Pre-election rollout frames USAI Labs as a permanent national capability. By Election Day, the core frontier labs are government-owned entities under DOE supervision with operator contracts in place, audited safety gates, and DPA-backed compute access.

Safety as the “entry?”

You can easily tell this story without the safety angle- just on strategic competition.

Is it bad?

I don’t know. I just think its credible. Intel ownership is something that feel fundamentally new, and because its a “failing” company it feels like a test of a broader idea.

I have been listening to the excellent ‘Anthro of SF’ series by Jasmine Sun, and in the latest episode there was this BEAUTIFUL passage talking about Buckminster’s ideas of “valving” technologies.

I read it, and instantly had a kind of revelation around GPT-5 and what went so deeply wrong with the latest OpenAI launch.

He says:

GPT-4o hit potent product-market fit: a responsive, warm, tolerant chat interface that satisfied specific emotional jobs, from brainstorming to basic comfort.

That warmth, plus the familiar chat surface, acted like a factory-installed valve. The model felt accommodating without heavy configuration, which helped adoption.

This was a Model Valve: RLHF and other techniques causing a model to be usable for “individual growth, individual worlds, individual improvement.” This is why there are hundreds of thousands of people who loved 4o- these three ideas of growth, worlds, and improvement.

Calling it a sycophant is a cheap, reductionist way of showing smug technologist superiority. “I don’t need a valved model, I want a firehose of intelligence” etc. I hated 4o, because I am in this camp where I want capability and cold, effective tooling. I am not, and you (dear reader) are very likely also not, the type who want a valve on the front of their LLM.

Great technologies, especially AI, need this valving function. 4o was great in this regard, and thats why people missed it so much.

People who love GPT-5 love it when its put to work- and that is a valve. Cursor is a valve on that intelligence. IDE’s focus the beam, put it to work.

GPT-5 Broke the Valve

GPT-5 is a very good firehose of intelligence. It does what you ask it to.

But there is no valve. The default of the model does not invite you to play, doesn’t give you back breadcrumbs of validation to keep going, and doesn’t guide you away from dead-end paths.

The lack of valving of the technology explains the massive split in the reception; if you have a valve of choice, you love it. If you don’t, you want the old one that OpenAI made for you back.

The Next Great Valve

The next year will see AI companies solve this problem durably by valving at the account level. We’ll see memory, preference, and a dozen other undiscovered techniques start to effectively valve down intelligence that currently flows through far to wide a pipe.

Sadly, OpenAI wasn’t ready to launch account valving changes, and had to revert back to their RLHF/preference model changes. I understand why this happened, but its a terrible thing to see as it will only delay the changes we so desperately need- a chance to deliver the right kind of desirability to each user, individually.

Memory is the only thing I obsess about when it comes to LLMs this year. The design space is truly massive and truly exciting.

Learning from Bucky & Brand

There’s a lot of wisdom and clear thinking in Buckminster and Brand, two writers who my friend Michael introduced me to 12 years ago. I want to support more work that brings their ideas forward to the current moment; if you know of people who are doing this, reach out.

[public] The Atom Project Executive Summary Aug 4th 2025

3.76MB ∙ PDF file

Download

Download

I care deeply about having truly open models trained in America.

Please read, support and co-sign.

The building is painted matte black, and the Market street sidewalk outside is reeks of urine. I peer through some dirty glass to see a young man in a Hawaiian shirt sitting behind a tiny little desk, scrolling his phone. There’s a poster showing the same building I’m standing in front of covered in futuristic green gardens, blooming from rooftops and balconies. Beside me, a man is doubled over, silent, while scrunched up aluminum foil crinkles in his palm.

I’m here to see Emmett Shear talk about AI. Hawaiian shirt man lets me in, gives me a yellow name tag, and tells me to head up the stairs.

I pop my head into a big room full of chairs, and the smell of Crypto hits my nose, right away. There’s nothing about crypto in the space, but its evident from the instant I arrive that this project, whatever it is, is a Crypto Project.

I start hunting down some coffee, and a man in sunglasses and cat ears chats me up about the morning rave he just got out of. I am not against morning raves- I think that is a cool way to start a conference, and I feel a tinge of FOMO that I missed this. I find some coffee and start to chat with a person who is one of the organizers of the space I am in- which I come to understand is in fact not a floor of a building, but a whole damn building. 16 floors.

There’s something for everyone- a cathedral of techno-optimism, with floors for robots and bioengineering, floors for talking, cooking, AI, floors for serious pursuits and morning raves. My FOMO intensifies- how did I not know about this?

Coffee in my veins, I start poking around. The Crypto Smell is everywhere- bright eyed catperson ravers, cardboard boxes stacked in the corners, covered in drywall dust, flyers with Emmett Shear’s face smiling back at me. The vibe is an underfunded student union after a wild party weekend, and its hard to tell if the space is under construction or quietly rotting.

The crowd starts to pick up- a french man in a fire-engine-red jacket starts to talk with me about Crypto. The smell is so thick in my nose I can hardly think. Beautiful women in short skirts constantly interrupt our conversation, asking about guest speakers, microphones, and lunch catering options.

I know nobody here. I go to 3 AI events a week, and I know the people who go to AI events- they’re not Sam Altman, but they’re curious, serious, and mostly pretty quiet. They must be all recovering from the Morning Rave, maybe hanging out on the Human Flourishing floor of this building, drinking matcha with their shoes off.

I crack open the event app and look to see the morning agenda. Emmett Shear is not, in fact, speaking. The first speaker also cancelled. A man with cat ears on starts speaking over a slide deck about the tower- he doesn’t say much about AI, but he does have a slide explaining that one time he threw a party in LA that Grimes attended. He’s unbelievably proud of this fact.

crypto smell in the office

I remember reading Fred Wilson’s posts in 2013 and 2014 on crypto and falling in love with how beautiful and clean the ideas were. I hold bitcoin, I like bitcoin, I have friends who have dedicated their professional lives to building bitcoin. My feelings over 12 years have moved from euphoria to exhaustion, and I don’t think I am alone.

The believers are always going to be OK, but the peripheral people need a new narrative, beyond whatever crypto is providing in 2025. AI provides one. Bioscience provides another. Longevity research, psychedelic therapeutics, space mining, you name it: if the field carries a whiff of frontier and a tolerance for loose accounting, the crypto diaspora is already circling.

The migration would be fine if it stopped at capital. Science has always relied on dubious money; railroads were financed by robber barons long before they carried wheat. The trouble arrives with the culture that tags along.

Crypto culture is maximalist, reality‑optional, and permanently adolescent. It runs on memes, conspiracies, and the dopamine loop of 24‑hour price feeds. Inside that loop, rigour is for suckers and accountability is for victims. When you import that attitude into disciplines that manipulate genomes or deploy autonomous systems, you are no longer gambling on JPEGs. You are gambling on the substrate of life.

I am not romantic about science. Biosciences have their own carnival barkers. AI has plenty of vapourware. But both domains still depend on falsifiable claims. A lab that cannot replicate results dies. A model that fails in production is pulled. Crypto, by design, lacks that filter. No matter how many projects collapse, the ledger marches on; the next whitepaper is always just one hype cycle away. The discipline of admitting error never takes root.

crypto smell in the arXiv

Banning crypto capital from frontier science is neither practical nor desirable. Some of the best hardware companies in AI were seeded with early Bitcoin winnings. GPUs bought to mine can be used to fine-tune models; 21st century swords to ploughshares (or if you live in East Bay, ploughshares to swords).

The thing I care about is guarding the culture of AI. I want young people full of ambition and energy to build in this space- I want an AI community that is open to Morning Raves and to coffees laced with 10mg of creatine and discord anons making cool points about papers on RL while they trade Fartcoin.

But I want it to stay serious. I want to extend my life, I want to help birth superintelligence, I want to live in a society that values serious scientific advancement. I want to keep these things clean of the crypto smell, by caring about vocabulary, caring about reading the damn paper, caring about keeping whatever culture that is currently working, working. Because it is ours, and it is working.

crypto smell washes off

I’m making my escape down the stairs, and the Hawaiian shirt man tells me that I should stay for the party, later on. My yellow badge doesn’t permit me party access according to the rules, but he seems keen to make an exception.

A young woman asks me about the event. I tell her how I felt about it, and she listens, carefully and attentively. She takes notes, she listens, frowning. She cares about this space being great. She’s serious. She works somewhere in this building, perhaps sandwiched between the yoga studio and the robot fighting cage, where the windows are open and there’s fresh air coming in. Its nice to think there’s a Serious Floor here, somewhere. Personally, I’m hoping its the one where they’re doing the experimental biotech work- but some part of me knows that its smells like crypto there, too. They’re having a biotech summit soon; this is the flyer.

Smell that? Thats crypto in your Mitochondria.

Yesterday I spent two hours and twenty minutes doing 8 hours of work. This miracle was performed in deep collaboration with ChatGPT and Manus AI, and I suspect that there are millions of folks performing these daily miracles across the planet.

The work I did wasn’t particularly special or important; I built a plan for an event, a slide deck to find speakers and sponsors, and an implementation schedule to keep the project on track. The banal nature of the example here I chose is deliberate; LLMs may in fact be amazing tools for cancer research or some other wildly difficult activity, however my own use-cases are much more run-of-the-mill.

In 2019, this type of activity would have taken me 8 hours, and would have involved only about 60 minutes of deeply difficult work that Bloom’s Taxononomy would have put in the upper tiers:

In 2025, there’s the new, improved time breakdown:

LLMs have dissolved almost every low-friction task. In the Old World, the tasks which were less demanding; link collecting, resizing screenshots, researching, tweaking language, are simply gone.

Everything left is essential work. Everything is a max-weight intellectual deadlift. The speed up is extraordinary; the cognitive resource demands are still 90% of what they would have been in 2019.

That physiological jolt is the phenomenological entry point for this essay: large language models do not merely save time—they change the energy profile of knowledge work. They strip away the gentle ramps and leave us on constant inclines.

old man yells at cloud

I am fully aware of the past versions of me who yelled about how typewriters destroy the cognitive flow of cursive writing, or how stories written in cursive no good compared to stories told around a campfire. This is true, but so is my experience yesterday.

the METR study

This week I had the pleasure of hearing Joel Becker and the METR folks talk through their study of developer productivity at a live event in SF. If you haven’t heard of this work, sixteen senior open‑source developers worked through 246 real issues in a big, complicated OS repo. When they were allowed to use AI tooling they moved 19 percent slower than when they coded unaided (metr.org). More striking, they believed they were 20 percent faster. The gap between perceived and real velocity suggests that LLMs mess up our internal metronome.

The Economist asked a sharper question last week—“Does AI make you stupid?”—noting that many workers feel tasks require less cognitive effort, even as output quality plateaus (economist.com)(archive v). Together these pieces frame a paradox: automation removes toil yet may sap momentum by front‑loading cognitive friction.

focus, not time, as the scarce resource

Classical productivity literature treats time as the limiting factor. In creative work the true bottleneck is uninterrupted, high‑quality attention—Cal Newport’s “deep work.” LLMs compress timelines by seconds but consume attention in larger, denser blocks. Each generated paragraph must be evaluated for factuality, tone, and coherence. That evaluative loop is ‘heavier’. Over a 90‑minute session the aggregate decision weight is hard on the grey matter!

Recent neuroergonomic research supports this view. An EEG study of AI chatbot interactions found reduced prefrontal engagement only when prompts were trivial; complex tasks pushed cognitive load back to baseline (frontiersin.org). The pattern matches my deck‑building fatigue: removing scaffolding tasks narrows the funnel until each remaining decision carries more subjective weight.

an idea for a study

I want to have groups of people come into a lab and have half use AI to complete tasks, half without. The task is about an hour without an LLM, and 15 minutes with an LLM. The tasks are difficult and cognitively demanding.

Participants can leave anytime they have completed the work. On the way out, they do a little “processing” where they answer questions about their experience, and then as they leave there’s two tables of snacks- one table loaded with healthy veggies and low-sugar options, the other full of chocolate cookies, cakes, and deeply unhealthy foods.

I bet you that the LLM-augmented participants will dive into the sweets at a dramatically higher rate, as they have used their willpower up aggressively. (shout-out Kathy Sierra). They have reduced capacity to make “good’ decisions.

physics of mental resources + willpower

Mental workload studies describe two competing forces: task complexity and time on task. LLM support lowers complexity for routine subtasks yet extends exposure to high‑complexity judgment calls. You exit a session with fewer completed minutes but higher cumulative strain. In aviation ergonomics the phenomenon is called the “automation paradox”: pilots fly fewer manual minutes while staying closer to the edges of the flight envelope when they do so.

Software engineering now inhabits a similar regime. The Buddha’s middle path is gone; work feels like alternating idling and red‑line RPM without a cruising gear. That mismatch triggers the subjective exhaustion I experienced and that Becker’s subjects likely masked with optimism bias.

Zoomers may not feel this way

Generation Z knowledge workers will never know the joys and horrors of shitty 2000s knowledge work. They will start from higher abstractions, just as touch‑typists never wrestled with typewriter ribbons. Their baselines will recalibrate around tool‑mediated flow. Yet there is reason to doubt that adaptation alone solves the fatigue issue. High‑performing esports athletes, who are natively digital, still exhibit measurable cognitive depletion after intense sessions of automated aim‑assist. The scarce variable is not familiarity but synaptic glucose.

this will not stay this way

Today is the worst these tools will ever be. This specific problem will go away- others will take its place. The shift by next year in a fully agentic world may be:

If this means 8x more “rotations” through a work day, we’re going to need to think carefully about how to design systems that don’t reduce your brain to mush after a single day of this type of effort.

I spent one weekend at LessOnline in Berkeley and the three days that followed at the SCSP AI Expo in Washington. Same country, both ostensibly about artificial intelligence, yet the gatherings felt like excursions to non-overlapping planets. LessOnline prizes belonging; the Expo prizes persuasion. Everyone talked about steering the future away from existential harm, but the Berkeley crowd sees peril in runaway capability while Washington sees opportunity in fielding it at scale.

Both planets defend themselves with language. In Berkeley the hallway chatter orbits “p(doom),” “take-off speeds,” and the “orthogonality thesis.” In Washington the code words are JADC2, ABMS, and CMMC. Miss the dialect and the conversation evaporates. Each tongue forms a closed system, efficient inside, opaque outside, and neither side shows much appetite for translation.

The ideas presented at LessOnline were undeniably brilliant, and the people who lay about Lighthaven chatting were impressive. The organization was fucking terrible, their conference app was beyond terrible. From the bumbling hour long introduction to the lack of food, the incompetence was clear, and almost celebrated- “oh, gee, look at how we’ve fumbled this, we’re just good-hearted writers who can’t seem to get ourselves sorted out”. The first night the pitiful amount of mediocre Indian food they brought in was instantly gone and so I got to see some of the most wealthy people in the Bay Area eating shitty pizza from a greasy paper plate at a 600 dollar/ticket event. The pizza, for those curious, was corn, peaches, cheese, and some unidentifiable green vegetable. Horrifying.

The AI Expo was much less brilliant, but vastly more competent. The vibe change off a redeye was so refreshing- eye contact, competent organization, people walking with purpose. The median talk was terribly dull, and I learned so little compared to my Berkeley weekend, but this was not about learning- it was about persuasion.

China hovers like a shadow in both conversations yet rarely appears in person. AI 2027 casts a Chinese lab as the spark that turns fast take-off into a global scramble, and its name surfaced in half the Berkeley sidebars I joined. At the Expo every keynote invoked Beijing as the pacing threat, yet only one breakout was explicitly about China, and it never filled the room. The country functions as a motivation to hurry, not as a subject to understand, and certainly never as an ally in co-creating a vision for the 21st century.

The two communities share bones. Almost everyone in both rooms holds a graduate degree or left one unfinished to chase a grant or a dry-run deployment. Ideology runs thick: rationalists assume superintelligence is plausible and alignment mandatory, while defense planners treat technological superiority as first-order deterrence. Each tribe is convinced the real leverage sits where they sit- the labs in Berkeley, the purse strings in D.C., and that conviction keeps them confident and apart. I could not help but see these communities as the products of different kinds of suburban nerd childhoods. Got into debate club, and parents golf and go to church? Welcome to the AI Expo. You a Mathlete with folks who have attended a grateful dead show and own a Prius? Welcome to East Bay Rationalists Prom.

California’s animating verb was prevent; the capital’s was deploy.

A few moments crystallized the gap. Scott Alexander packed a dark night in the courtyard to riff on biblical apocalypses, his silhouette in front of slides that made the room laugh (did you know that all of the Anthropic cofounders have “names of ruination”?) . I met someone whose twitter I would describe as the most rancid vat of terrible ideas on the internet and found I liked them in person. LessOnline felt like the comment section come to life—my favourite follows and my most-blocked tweeters sharing the same little cubbies of their sun-drenched Berkeley fortress. Across the country Eric Schmidt opened an Expo panel, folding Silicon Valley optimism into defense-industrial cadence, but the atmosphere snapped back to acquisition checklists as soon as he ceded the stage.

Back-to-back attending showed how both sides lag. Washington is behind the frontier; conversations about large-language-model capability stop way before real proficiency. There are genuinely baffling products being hawked- a hardened case full of H100s that runs on batteries, so you can do RL in the field comes to mind. The gossip only underscored how little organic familiarity government buyers have with the systems themselves. Berkeley, by contrast, is behind on persuasion. MIRI’s latest strategy packet stretches past twenty thousand words of game-theoretic dread and assumes the reader already speaks the shorthand; a Senate aide would bail by page three. Both communities chase status inside their bubbles, winning digital upvotes and three-letter-agency promotions.

AI 2027 is the clearest case. If its scenario really sketches the fastest plausible acceleration, where is the twelve-page edition a legislative staffer can digest between mark-ups? Where is the glossary that turns recursive self-improvement into procurement milestones? Until someone rewrites the document for the audience that writes the laws, it will keep echoing in the choir loft. I asked 2 of the authors of this work (which I consider to be important) if they want to expand the accessibility of the piece- and was met with utter indifference.

So what now? First, AI 2027 needs a full rewrite for Washington—less prophecy, more milestones, fewer cliffs of conditional probability. Second, D.C. must chase the technological frontier rather than wait for it to clear security review; too many defense primes showcased demos that would have looked dated 10 years ago. Finally, we need deliberate overlap: someone, anyone, who splits their time between Lighthaven meet-ups and Pentagon briefings. Keeping these planets in conversation is the only way America stays the birthplace of AGI instead of its first cautionary tale.

Today Airbnb launched a services inside their app- they’ve been working on similar ideas in the past like Experiences, but this is a deeper level of service integration into the core Airbnb product.

Just like Twitter under Musk, I see the slow wechatization of large, well capitalized technology companies as a big, ugly mistake that 10 years from now will be obvious to shareholders.

There’s a deep set of issues that haunt Airbnb as a platform in 2025;

1. regulatory: in many cities, its banned or at risk of bans.

2. fee structures: the hidden fee structures of Airbnb are very difficult and are not pro-consumer.

3. supply/listing quality: early hosts who had dedicated properties were great, because they were motivated early adopters. Post-COVID, there are a lot of bad hosts who are doing it because they were told its easy.

4. unresolved design problems: there are many unresolved design problems in Airbnb a platform. These issues have always existed, but with scale they are even more obvious to anyone but casual users.

As a public company CEO I think BC is great- he’s a wonderful communicator and having seen him in many different contexts, I think he’s a great leader.

As a company, I love Airbnb. I have 150 reviews, 6 figures of spend.

But I want the core product to be AMAZING. I want the focus of this mammoth company to be relentlessly focused on the core product being improved. I don’t want a cookie baking experience, a chance to win a stay in a floating house, or another channel to rent-seek my massage therapist appointments.

I want great stays, I want great hosts, I want great support, I want great discoverability, I want great planning.

I want relentless focus on the core offering of a great technology company, and I want to avoid the wechatification of my apps. Please.

https://x.com/bchesky/status/1922364651775385999

Gwern tweet from today got me thinking about weird LLMs.

The internet is full of takes on how LLMs are the greatest economic engine since the integrated circuit, so the total lack of weird LLMs is a great mystery to me.

Every single AI product feels like:

• ChatGPT for Q&A, search, general utility

• Roleplay/social companions

• imagegen front ends

• therapy

Today I’m asking the question: “where are all the weird LLMs?”

some weird LLM ideas

time-capsule GPT-1971

Imagine a model whose corpus ends on 15 August 1971, the day Nixon suspended gold convertibility and effectively killed Bretton Woods. Its training mix would be period newspapers, Congressional testimony, television transcripts, and technical journals, cleaned but never post-dated. Vocabulary, idioms, and geopolitical priors freeze in amber: the USSR is a permanent fixture, China is “Red China,” and inflation is only just becoming a dinner-table worry. To prevent leakage of modern knowledge, retrieval layers point to scanned microfiche rather than today’s web, and RLHF is done with historians and grumpy old people.

A PhD seminar can interrogate 1971-GPT on Nixon’s wage-price controls, or a novelist can sanity-check dialogue for a period thriller. An optional “parallel-timeline” mode lets modern economists debate the model, then see its counter-reaction from within 1971’s epistemic bubble. Commercialization looks like a time-travel API sold to universities and media, while agentic systems could dispatch 1971-GPT whenever a query includes “historical context,” automatically labeling citations as contemporary to 1971.

hallucination-as-art engine

Instead of punishing hallucinations, this LLM treats them as first-class creative assets. It is fine-tuned with prompts that reward dream logic, surreal metaphors, and synesthetic leaps. A metadata tag—fictional: true—prefaces every answer, freeing models from factual and beige answers while protecting users who need clarity. Stylistic controls let you dial from “mildly whimsical” to “Dalí mode”.

For writers, it becomes a sandbox: pipe an outline in, get plot twists out. The business model mirrors stock-photo libraries—sell dream bundles or license unique story seeds—while safety rests on hard-coded disclaimers and an always-on “reality check” companion model. In agentic stacks, the engine plugs into brainstorming chains: a sober planning agent hands off to Hallucination-GPT for idea explosion, then funnels results back for filtering.

age-tuned companions

The “Kid-GPT” variant speaks at a Grade-3 reading level, uses bright UI colours, and answers the eternal “why?” loop. It is trained on children’s literature and common-core curricula but also embeds strict content filters and parental dashboards. Its elder twin, “Senior-GPT,” features larger fonts, voice-first interaction. Nostalgia is a key driver for utility, so there’s lots of historical context.

Both models tackle engagement loneliness: one keeps eight-year-olds under safe supervision during homework; the other reminds an 82-year-old when to take medication or schedules a Zoom with grandchildren. Monetization could follow the family plan model—one subscription covering multiple age-profiles. There’s value for families in the old-folks version as a pair of ‘eyes’ on how they’re doing (health insurers too, but thats a wee bit dark, right?) . Within agent frameworks, a guardian orchestrator routes requests: chores to mainstream GPT, storytime to Kid-GPT, medication queries to Senior-GPT—ensuring each user talks to a cognitively-aligned persona.

Kids get infinite storybooks where the family cat is the main character, adults get the security of knowing that the model that is creating stories for kids has no knowledge of swearing, violence or stories without morals, and grandma has someone who reminds her to take meds while telling stories of Elvis’ early career days.

values-challenger bot, or change-my-mind bot

Here the system is explicitly tuned to surface contradictions, logical fallacies, and implicit biases in user statements. Training blends debate-club transcripts, philosophy texts, and annotated social-science papers, with a reinforcement signal that rewards respectful disagreement over compliance. The interface telegraphs intent: a prompt might read “Ready to change your mind?” so users opt-in to scrutiny rather than feeling attacked.

Applications range from executive coaching—stress-testing a strategy for blind spots—to classroom debate prep. The model logs bias-heatmaps that visualize which assumptions it flagged, creating a shareable artifact for reflection. A tiered safety scheme prevents it from straying into harassment: tone policing and escalating humility if the user signals discomfort. In multi-agent systems, Challenger-GPT acts as the red-team check on outputs.

deliberate non-engagement model

This LLM is optimized to finish conversations, not extend them. Loss functions penalize token bloat and re-ask rates, while rewards spike for concise, actionable answers. Data for fine-tuning comes from Stack Overflow accepted answers, airline-pilot checklists, and well-rated customer-support macros—sources where brevity equals quality.

The user experience feels like texting a hyper-competent friend: you ask, it replies in 2–4 crisp sentences plus bullet-ready next steps. Business value shows up in call-center AHT (average handle time) reductions and in productivity tools that promise focus, not feed. Because the model bucks engagement metrics, revenue leans on SaaS fees or enterprise licensing, not ads. In agentic workflows, Non-Engagement-GPT becomes the closure node—it cleans up, summarizes, and hands off results to humans or downstream systems, ensuring the chain ends promptly instead of looping forever.

why this doesn’t exist

Big, strange ideas get ironed flat the moment they enter the LLM pipeline. First flattening pass is cost. A single-run post-training bill can still be high 6 figures, and that number covers one try. Boards and investors naturally ask, “Why risk millions on a model that might only serve 1971 economists?” So builders grab an existing generic checkpoint, sprinkle mainstream data on top, and call it done. Even the dataset itself must be flattened—anything with thorny rights is stripped away to avoid lawsuits. By the time the model is trained, there have been dozens of passes of the iron.

Second pass: risk and reputation. Policy teams, brand lawyers, and soon-to-arrive regulators all share a universal incentive—minimize headlines. The cheapest route is rigorous RLHF that penalizes any output deemed controversial, sarcastic, or too weird to explain to a tech journo. “Helpful, harmless, honest” sounds noble; in practice it means sanding off dialect, irony, political edge, and anything else that isnt’ ChatBot Beige. Users end up conversing with an entity as charming as a corporate HR training.

Final pass: metrics and UX. Product managers optimize for CSAT, retention, and net-promoter scores—numbers that spike when conversations stay predictable. Designers stick to the trusty chat because anything funkier tanks onboarding funnels. Meanwhile, copyright and child-safety statutes loom over every pixel, urging teams to remove features that might invite scrutiny. The result: a consumer AI landscape where every stakeholder unconsciously collaborates on the same result—flattening everything interesting into a single, safe, beige, browser chatbot.

maybe it will exist

There’s some evidence that these barriers are falling.

Cost is shrinking; DeepSeek’s sticker price alongside low-cost fine tuning might allow for some experimentation. If you don’t need VC to get weird, thats one less pass of the iron, and I am hopeful.

Risk and Reputation might be easing, too. Sure there’s still going to be safety concerns (and thats good!) but the days of terrible licenses seem to be coming to an end. Regulations have been light, so far, and there should always be space for “experimental model” carve-outs. As more people in the space get funded, there’s more to look into.

“Helpful, Harmless and Designed for Maximum Engagment” might be the hardest to overcome, but there’s cracks in that armour too. Deciding to borrow the FB/YT/Netflix metrics playbook for engagement might have gotten openAi to #5 in traffic worldwide, but the dominance of their product actually may encourage newcos to try new things. I think product teams are noticing. Engagement as the #1 metric might not be the best strategy in 2025.

Finally, agentic architectures might mean that weird LLMs have utility for normie use-cases. Mini models that check bias or give historical perspective might be really useful in orchestrating outcomes that are not about generating text. I hope that there’s a space for eccentric models when we’re orchestrating hundreds of agents in workflows. This might end up producing interesting outcomes.

where to now?

I think there’s room for weird LLMs in the world! If you’re thinking about building something along these lines, reach out, I’d love to talk.