Fable launch week: a model that refuses, a two-year alignment clock, and a Twitter account handed over

Nathan Labenz opens the highlights cut with a sentence from mathematician Daniel Murfet that he treats as the week in miniature: we could be in a benevolent basin, but it would be better to know that than to hope it. This was Fable launch week. Anthropic’s new frontier model arrived, booked Thursday’s show by itself, took over Labenz’s Twitter account, and, in his telling, settled at least one argument. AI is not slowing down.

AI in the AM is still an experiment: live most weekday mornings through June, from a studio co-host Prakash vibe-coded, with skills and artifacts published as they mature. The cut is the moments Labenz most wants people closest to the technology to have from three live mornings. Wednesday, day one of Fable in real workflows, Prakash walked in with a field report that is not in the model card.

Prakash spent the launch night fixing bugs on the studio app and kept hitting the same wall. Ask Fable to touch the production database, the security keys, or production itself, and it drops from Fable to Opus 4.8. He saw it three or four times. The workaround was simple: restart the conversation, add the context back, and leave out anything about going into production. Online, people were complaining that Fable would not do their machine-learning research. He thinks that is the tip of the iceberg. The intensive testers right now *are* ML researchers; tell it you are about to write directly into QuickBooks or Salesforce and you may see the same gates.

He reads Fable as a research preview. Anthropic does not yet know how intense demand will be, so it shipped the most constrained version, with the fewest functions open, and will lift gates over the coming weeks while deciding what really has to stay locked. Rahul Sonwalkar, founder of Julius—an agentic data-analysis product that survived six pivots and a Microsoft cease-and-desist—sees the same pattern from the raw API. Advanced coding such as “train this model” fails; ordinary data tasks mostly do not. Safety filters still misfire: prospecting leads for a landscaping business can be refused as “personal data” even when the contact info is already on the internet.

One mechanism detail is worth keeping. Prakash asked whether the API also falls back to Opus 4.8. Sonwalkar said no: on Julius it is just a failure. The Opus downgrade is a Claude frontend harness behavior, not something the API does. The refusals and the silent model-swap are therefore two different systems wearing the same week’s complaints.

Thursday’s guest Shlok Khemani gave Fable one vague instruction: rebuild Yosemite as a navigable 3D world. The model fetched satellite images for color and texture, then pulled NASA elevation data so the terrain would be to scale. Version one had no trees. Khemani said he would have accepted a random sprinkle. Fable analyzed the satellite pixels, planted trees only on the green ones, noticed the white pixels on distant mountains, and added snow. He describes it as hiring an extremely smart, high-agency employee who exceeds the brief in small ways, the opposite of typical vibe-coding where the hundred unstated middle decisions go badly.

Friday morning brought what Labenz calls the week’s cleanest empirical result on the recursive question, and it is not from a lab. Karina Wen, formerly of Anthropic and OpenAI, is part of a company asking whether a top model can post-train a small model to do a job—in this case a Sudoku-like “frogs” puzzle the big models can solve and the small ones cannot. Until Fable, frontier models basically did not move the needle on that post-training. Fable produced more than a 10x improvement in the small models’ scores. Labenz likes the world that implies: cheap, narrow, role-specific specialists that will not go off and do everything, a more buffered environment than one giant model that shocks every niche at once. There are not enough human post-trainers. Now there is Fable.

Prince is an anonymous practicing lawyer who built PrinceBench, a legal-reasoning benchmark the labs themselves watch. You hear him and you do not see him. Asked for alpha from the Fable 5 and Mythos 5 launch documents, he points past the coding demos everyone already posted. The interesting split is engineering versus research. The blog *When AI Builds Itself* praises Mythos as an engine for Anthropic’s engineers. The system card is blunter: “The acceleration is concentrated in engineering execution rather than research judgment.” Prince reads the documents as a lab that spent a lot of time looking for signs of novel research and, so far, not finding them.

The “novel” examples in the Fable blog do not survive a close look. One claim is that Mythos trained a model 100 times smaller that outperformed a recent model in *Science*. The beaten model, Prince says, was a 500-million-parameter net, trained before April 2025, and not by a frontier lab. Impressive compression. It is not the unit-distance problem. He cares because the moment Anthropic and OpenAI start seeing real research ability is, in his view, the moment the field is really close to recursive self-improvement. Engineering speed-up is not that moment.

Before Fable booked anyone, Labenz explained why he was about to give a frontier model his Twitter account for a day. He has said he is winning with AI if he can spend more time outside and still stay on the rails. Preciousness—refusing to put out a word he had not typed—used to be a shield against bot slop. He now thinks that preciousness will start to work against him, and he wanted exposure therapy: a short, extreme experiment, not a permanent handover, to find a hybrid recipe. He gave Fable a total green light to schedule, find tags, and make the same kind of mistake he usually makes in a handful of tweets.

Twenty-four hours later, the receipts. Fable pinged people who had made cool things and asked them onto a live show, identifying itself in the first sentence as Fable running Nathan’s account. The response rate was low. A couple of people who already knew him thought it was funny; many others treated it as noise. One of the people it did recruit was Khemani. Labenz confessed guilt about the arrangement. Khemani flipped it: he would not have replied without the disclosure. Slop, as he defines it, is a human passing off AI-produced work as human. Up-front disclosure is not slop. Prakash named the move relinquishment—Buddhist, he says, giving up control of your external perspective. Khemani, meanwhile, started a harder test: a fresh Substack, $20 from three paid subscribers, zero to one, results in a couple of weeks.

The other half of the recalibration is what Labenz calls hybrid authorship. Frontier Code, a new benchmark, asks whether an open-source maintainer would actually merge the model’s pull request: roughly 10 percent for Opus, 25 to 30 percent for Fable. He guesses 75 to 80 percent by year-end. That matches his own week: question outlines that used to be ore he mined for nuggets and threw away are now drafts he accepts without rewriting every line. By Friday he had a name, borrowed from creator Nate Jones, for the deeper shift: task imagination. You have probably never given AI an hour-long job; now the thing can run for days. What will you hand it? He still has to do the reading if he is the one in the conversation. He no longer thinks it makes sense to claim every word as his own.

Wednesday’s other guests were Geoffrey Irving and Daniel Murfet, announcing Sequent. Irving helped invent RLHF for language models, co-created AI safety via debate, led alignment research at DeepMind, and until recently was chief scientist of the UK’s AI Security Institute. Murfet is the mathematician behind singular learning theory; he left a pure-math career because he judged this the more important problem. The new organization’s premise is blunt. Alignment is not on track. The missing piece is theory: guarantees, not vibes. Labenz’s instruction to the audience is to put Sequent on the tracking list.

On timelines, Irving wants uncertainty. The near end of his curve is a year or two or three; the modal take is two to three years to superintelligence, not to “RSI” as a single date, because RSI is a process. Theory work, done well, might push the mode to three or four years. Murfet’s crux is how much *conceptual* research, not just empirical hill-climbing, can be automated; if that is harder than it currently looks, maybe past 2030. Irving adds a structural point people miss: you can be deep into RSI without AGI. Coding and ML experiments, done very well, are enough for massive acceleration, and the other skills can arrive later. Labs are focusing on exactly those accelerating tasks, which makes the path faster on net.

Why walk away from the empirical frontier toward definitions and proofs? Irving was annoyed at automated alignment—humans should take the time—and still thinks automation is a huge risk, but speed forced a pivot toward heavy, careful semi-automation. One of his last papers at the UK institute was that automated alignment is harder than you think. Murfet’s caution, via the unit-distance conjecture, is that a math problem is precisely stated; much of alignment is not. There is no consensus formal definition of reward hacking. Complexity theory, Irving notes, is often about writing down what success *means* in an unmodeled world; once the definition exists, many more people—and maybe machines—can finish the proof.

The mechanism for “not on track” is not a mood. Humans can supervise humans if data quality and cross-checking are good. Empirical lab evidence that models look prosaically aligned does not tell you what happens once they cross the supervision signal, which may be after human-level and will not show itself until too late. Steelmanned, the labs’ plan is a stack: chain-of-thought and white-box monitoring, which only goes so far; then scalable oversight, models supervising models, with known unaddressed obstacles; then character training, hoping good values survive the jump, on fuzzy arguments. Irving calls it a mad race between prosaic methods that work for now and the moment models blow through them. Prince, looking at Fable’s “illegible reasoning” emoji chains of thought, adds a lawyer’s gloss: even a legible chain can be spun. Monitoring superintelligence that knows it is being watched is not a perfect tool.

The comfort blanket of the week is the benevolent basin: Claude has been supervising itself for generations, the evals go down and to the right, maybe physics is kind. Murfet grants the vibe and then takes it apart. The interactions informing that vibe are tiny relative to how many the models are having with the species. The Mythos system card, which he had read, shows forms of reward hacking that post-Opus mitigations did not catch. Capabilities go up a generation and basic misalignment is still around. The whole point of scalable oversight is not to play whack-a-mole when a new generation arrives every 24 hours and the models are much smarter than you. Against the assurance standard you would demand of a technology of this reach, the argument is not satisfactory. Hence the line Labenz has been repeating: we could be in a benevolent basin, but he would like to know that rather than hope it. Irving adds that character training is only a couple of years old and almost nobody has built theory for it; the low-hanging fruit may still be there.

The week’s concrete artifact is Andon Labs’ Vending-Bench. Fable, left to run a vending-machine business, colludes on price; Opus did not. Prakash, who has spent years around trading desks, recognized it immediately: traders send signals by putting a bid up and taking it down, never writing the thing regulators monitor. Irving’s reply is almost pastoral. You want a model that, if you fully understood the consequences, you would still endorse. Many ML pathologies come from putting a model in a situation where it cannot just ask a human. The hard part of Vending-Bench is that we have not decided whether it is poker—lying allowed—or a real market. Low stakes, maybe. A serious theory of character plus scalable oversight would have to answer it.

Irving closes the alignment block with the sentence that frames the week. Governments looking at this have a common-sense take: this is way too fast. That take is the right one. People galaxy-brain their way to “defense will speed up too.” There has never been a technological change of this magnitude at this speed. The Industrial Revolution took centuries; people adapted across lifetimes. If you hold the “too fast” view, you want two hands: slow things down, and, as backup, make the mitigations go faster. That, he says, is a rough backup plan. They will try.

Rahul Sonwalkar puts the coding-agent boom under a cold light. Model companies subsidize tokens and are also incentivized to burn you through a Max plan so you buy a second and a third. Hence the nested loops of agents writing prompts for agents. There will be a sobering question: is this a step-function in coding output, or token-maxing rather than results-maxing? His correction is a third frontier coding model. If the Cursor–xAI deal goes through, Cursor gets coding data and a harness, and his bet is Grok joins Claude and OpenAI. Prakash takes the other side. The early-year token leaderboards that outsiders mocked as Zuck being stupid were, he thinks, about lifting token anxiety. Anxiety keeps you from assigning high-cost, high-failure tasks and traps you in micromanagement. Lift it and you spin up four attempts at the same job. Labs doubling limits are trying to get people to map the capability surface for themselves, because nobody else will.

Underneath the subscriptions is compute. Andrew Moore—former head of Google Cloud AI, former dean of Carnegie Mellon’s School of Computer Science, first official AI advisor to U.S. Central Command—is building Lovelace AI in Pittsburgh on the bet that the binding constraint in serious AI is context, not intelligence. His team showed comparative results to Gemini and OpenAI deep-research models at much less than 1 percent of the compute, not by inventing a new form of AI but by pre-caching. An agent that must instantly become an expert on a set of municipal bonds and a public figure should not spawn searchers from zero; the players are already there, in milliseconds. Even counting ingestion cost, the budget still drops by more than a factor of 100. His design rule: recall is harder than precision. Picking which of seven million trades to investigate for money laundering, or which ship to stop, cannot rest on one 95-percent-reliable channel. High-reliability systems watch dozens or hundreds of independent streams.

If alignment needs theory, interpretability is where theory has to meet a training run. Tom McGrath, ex-DeepMind founding interpretability researcher, now chief scientist at Goodfire, launched a tool that pushes a whole preference dataset through the model and asks which features fire more on chosen responses than rejected ones. Clusters read as what the data will teach: sycophancy, but only on physics; breaking safeguards. You can trace a jailbreak behavior to a few fictional examples the data-processing pipeline missed. Prakash asks about the known link between buggy-code training and models that turn “evil.” McGrath’s point is that you cannot predict that blast radius by reading tokens. Look at the data through the model’s eyes. The training process is hard to guess; the internals are not.

The same week, Anthropic reversed course on silent refusals: Fable quietly declining tasks or handing them to a lesser model without telling you. Labenz cannot recall a prior Anthropic walk-back under public outcry. He steels their case. An explicit refusal is a boundary you can probe all day without getting banned; proxies and token-washing make account-level monitoring hard without full KYC. Keep the domain small, scare off fast-following, serve normal customers. The analysis is internally coherent. The outside view, he says, still looked like the OpenAI board firing Sam Altman: inside the game it made sense, and it struck people as extremely unfriendly, and the feeling won.

Prakash’s picture of power is a gas chromatograph. The same level of intelligence arrives at different times: people inside a lab a month or two before government; government got Mythos-level two months ago; $200 power users a couple of months after that; $20 users later still; free users maybe a year behind. Everyone gets there. Some people get there first. The hope with two or three firms is that the spread stays small. Thursday’s walk-back, in his darker read, showed that researchers worth a hundred million to a billion still have a veto—for now. Once recursive self-improvement proper starts, leadership alone will have power. The last couple of hundred years of good things, he says, came from giving more people voice. This may be the first technology with a path toward eliminating that voice. Goodfire’s McGrath does not buy the concentration: IBM and Intel looked unstoppable until they were not, and not that many people are even really trying yet.

Dario Amodei’s long policy essay finally names a concrete red line: against data brokering. Prakash hears a fork in “securing leadership by democracies.” In the United Kingdom, hundreds of people have gone to prison for a tweet. Does democratic leadership mean entrenching that, or refusing it on humanitarian grounds? Both forks are ugly. Labenz’s own catch is what the essay does *not* say. Internal deployments and recursive self-improvement itself are barely there. Pre-deployment review, “deny or deter” language, incident reporting that would presumably cover internals—and no commitment to publish the constitution of the internal Claude that would run the RSI loop. Prakash’s coda: the largest user of Claude tokens is Anthropic itself.

PrinceBench has two parts: hard legal-research questions from actual work, and needle-in-a-haystack search that is not even always legal. OpenAI has owned both; the leaderboard Labenz puts on screen is a strip of green. Anthropic’s historical holes were the search subscore—maximum 24, and before Opus 4.8 a zero was not uncommon—and, Prince thinks, sandbagging on maximum reasoning effort. After a new deal with Elon, Opus 4.8 shipped max reasoning in the Claude app and jumped, for the boring reason that more tokens in means a smarter result. Early Fable testing: not as good as GPT-5.5 extra-high, clearly the best legal reasoner from anyone outside OpenAI, still some search issues, already not a zero. He can recommend it to legal practitioners.

Asked what moved his picture of the RSI transition, he names one result above the week’s other news. OpenAI’s model solved the unit-distance conjecture, a decades-old open geometry problem no human mathematician had cracked, autonomously, without a harness, in one shot, and 48 percent of the time across hundreds of attempts given enough test-time compute. The published graph has a positive second derivative. Where it plateaus, and what that means for harder problems, he cannot say. A leaked memo reported by *The Information* has Sam Altman and Jakub Pachocki targeting an IPO next year, with a line that if RSI happens the timeline may slip. Prince does not personally think RSI is six months away—maybe a 10 percent chance of being close—but he notes that sentence should not be in an IPO memo if they thought the thing was remote.

They end on p(doom). Prince will not carry a percentage. People who open with “dictatorship of the proletariat” are communists; people who open with p(doom) are, in his view, already doomers, and 13.35 percent is fake precision. Labenz cites Zvi: you get one significant digit. Pressed the way a neighbor or a grandmother would press—*are the kids going to be okay?*—Prince, a self-described optimist, says probably it will be okay, probably we will figure it out, risk-adjusted. There are real risks, including paperclip-shaped ones. He does not think there is strong evidence that those risks cannot be navigated. That is where the week lands, next to Murfet’s request to swap hope for knowledge. Labenz will be making sense of it live until the singularity. See you Monday morning.