GPT-4 Failed a Hollow Top Hat. Hollywood's AI Deal Closed Only Half the Loopholes.

Trey Kollmer, a WGA member, last appeared in May with two guild colleagues to talk through the writers' AI demands. He is back alone, this time with three months of questions. The first update is that the actors have struck as well. SAG-AFTRA's deal expired, talks failed, and some of their asks are AI-related. The second is movement on the writers' side. After months of silence, the AMPTP, which represents the studios, invited the union back to the table. The talks fell through, and the studios then published their latest offer — a tactic, Kollmer says, meant to make some writers think the terms sound pretty good. So far it has not worked.

The AI section of that offer does close two holes the guild had named last time. A rewrite is usually paid at a lower rate than a first draft, so a studio could have a model write the draft and pay a writer only to rewrite. In television, a teleplay is paid separately for story and for script, so a model outline could erase the story fee. The AMPTP's language agrees that generative-AI material cannot be considered assigned material, cannot count as a first draft for rewrite purposes, and is treated as research. If you are starting from scratch, the studio has no claim based on what the model produced.

Kollmer is not sure the language covers everything the guild wants, and the union is keeping its internal read private so as not to sabotage talks. What has not been addressed is whether writers' material can be used as training data or to fine-tune models that then generate scripts. That, he says, is still a fight. It intersects with a second prong: a wave of copyright suits against OpenAI, Meta, Stability, and other labs, which some members hope will slow the habit of ingesting copyrighted work to improve the models.

Kollmer points listeners to a talk by Berkeley copyright scholar Pamela Samuelson at the Simons Institute. She grouped the suits against the labs into three theories. First, copying copyrighted material onto servers in order to train can itself be an infringing event, even if no one sees the copies and they are never republished. Second, model outputs that are too similar to training data, or that count as derivative works, can infringe on their own. Third, it is illegal to remove or alter copyright-management information — watermarks on Getty Images are the clean example, and a model that generates a similar image without the watermark can wander into the same trouble.

A lot of it, she said, comes down to fair use. Google was sued for ingesting books in order to index them; courts treated that as fair use because the use was transformative and because indexing did not hurt the market for the underlying books — if anything it helped people find them. Samuelson thinks the result may differ this time. Stable Diffusion and Midjourney, she argued, really can affect the market for artists' images and for Getty. On the other two theories she was colder: she showed a Stable Diffusion image next to a very similar training image and still did not think a court would call the output a derivative work, and she doubted the watermark theory would necessarily pan out.

What is actually up in the air is training itself: you copy the works and use them to build a system that can replace a lot of the underlying creators. That, she said, may need Congress. Kollmer's takeaway was that it "seems like crazy we're trying to apply these old laws and old rules to such a different paradigm." Labenz adds a civilian gloss: generative AI is both transformative and derivative at the same time, which is why it triggers the accusation and arrives with a defense. He does not offer that as a legal opinion. He offers it as why the fight feels so hard to score from the outside.

The U.S. Copyright Office has put out a notice of inquiry with 34 detailed questions, including whether a model can unlearn training material without retraining — a practical hedge if courts find infringement, so hundreds of billions in investment need not be torn down. It also asks whether datasets should be retained so later cases can test how derivative an output is. A recent Anthropic paper on which training points most influence later behavior is, Labenz says, the closest current answer to unlearning, and the answer is basically not yet. Their approximation finds a surprisingly small number of high-influence points: literal keyword matches at small scale, more conceptual matches at larger scale, and often not the obvious passages you would have picked.

Labenz has been pleasantly surprised by how many fields, medicine included, have approached AI disruption in something like good faith rather than digging in. At a high level the studios look similar: they were willing to agree not to write writers out of credit and compensation just because a model generated a pass. Kollmer does not have insight into the actual room, but he reads the published offer as at least claiming to block those loopholes. The demand that still sits apart from copyright is simpler. The union wants movement on not using their scripts as training data.

That is separate from the lab lawsuits because a purchased script, or a hired one, is usually work-for-hire. The studios own the copyrights. Whatever courts decide about scraping the open web, the companies can still ask: we own this material that our writers wrote — can we use it to train models that might one day replace them? Kollmer says there is still no agreement on that. Labenz can see the threshold logic from the studio side. Help a doctor write a more empathetic follow-up note and everyone is for it; put a Google doctor on every phone and the mood flips. The same cliff exists for infinite fine-tuned scripts. A studio may not want to concede the option before it knows whether the option is real.

Kollmer's strategic frame is that the protections the guild is asking for are most relevant in a medium-progress world. In a low-progress world the current models are not good enough to matter much. In a world where they get vastly better, trying to stop their use is quixotic. Labenz has been evaluating AI-safety charities and keeps seeing the same distribution: if scaling or some agent breakthrough runs away, nobody has a plan, so people try to matter in the scenarios that still look steerable. Two fields that barely touch, he notes, are premised on a similar range of futures. On that score the guild's strategy looks reasonable.

How much juice is there, Kollmer asks, in fine-tuning models specifically to write stories, outlines, and scenes? Labenz's closest data point is Waymark, the AI-video company he founded, which generates 30-second commercial scripts. That is a simpler problem than a sitcom, but the result was unambiguous: fine-tuning the original DaVinci model was night-and-day better than trying to few-shot the same model. Structure is what moved. If the examples have a one-two punch, the model starts to deliver a one-two punch. If you need consistent length and form so you can decide whether you like an idea, fine-tuning helps. He would not call the output mind-blowing.

He treats few-shot and fine-tuning as a continuum, not two tricks. One of his papers of the previous year designed an algorithm in the weights that implements gradient descent, then went looking for it in actual trained models and found it. The mechanism that lets a few in-context examples steer the output is, in that sense, doing something very similar to a small fine-tune. The models they inspected were far smaller than today's frontier — Labenz guesses on the order of one percent of current compute — which makes it more likely, not less, that the same thing is happening in GPT-4 and Claude. He would be surprised if it were not a general phenomenon.

His other data point is colder. He uses Claude 2 to draft podcast intros from five past examples, the episode transcript, and a runtime "angle." What he wants in the first minutes is a synthesis, a perspective that contextualizes what follows. What he gets is a decent outline in a neutral narrator voice: they talk about this, then they talk about that. He almost always rewrites it. Commercial copy at Waymark does not need to be groundbreaking; it needs a hook and a few key points. His intros do need a take, and he does not get one. Current systems, someone noted in the charity round, assume the future is like the present. That baked-in conservatism shows up in the prose.

Even a useful generator still leaves a filtering problem. If you have a hundred scripts and one is good, which one is it? Disagreement is the real struggle, Labenz says, and RLHF is taking systems toward a neutral style. Creative work, they agreed last time, is closer to a bimodal distribution: you will not get everyone to love it, but if some people really love it you are in business. Current training regimes converge toward the center. Kollmer puts the target as ten or twenty percent of people who love the thing, versus seventy percent who are vaguely not offended. Supervised fine-tuning on examples, which is what OpenAI currently exposes, does not let you train that polarized reward. More sophisticated tools have been teased for select customers; they are not here yet.

Comedy already runs on generate-and-mine. *New Girl* was famous for alts: writers brought a binder of hundreds of alternate jokes to set, tried a few, and mined favorites in the edit. Doing whole scripts that way sounds exhausting. Chunked into a hierarchy it starts to look plausible — twenty premises, pick a few; ten outlines, pick two; then generate scene by scene. Kollmer does not think the hit rate is there yet. At some threshold, though, the workflow that Hollywood already uses collectively becomes one an individual can run with a model.

A second craft rule is "one unusual thing": one element should be surprising, and everything else should play it straight. His example is *The Invention of Lying*, a world where lying does not exist and every other beat is a realistic consequence. Models sample at a fixed temperature, so they are random where they should be steady and flat where they should pop. The ideal, he says, is a model that chooses its own temperature: throw something unlikely, then land the plane. The Coen brothers have been described as handing off every fifteen pages, one writing into a hole, the other writing out. *Django Unchained* opens with Christoph Waltz's character walking into a bar and shooting a man; only later does the bounty make it make sense.

That last move is where Kollmer thinks models may actually have an edge. They are already good at hallucinating a reason why something is true after they have made a mistake. Generating the unusual situation themselves is the harder creativity. Feeding them an anomalous setup and asking them to write out of it might play to their strength. Labenz's practical counterpart, later in the hour, is to stop asking for "more interesting" and start naming a paradigm — one strange thing, the rest realistic — or a specific master of the style. Expertise, he says, is rewarded. The model can follow a craft rule it will not invent.

If a court or a contract ever forced labs to untrain copyrighted work, Labenz does not think it will be easy. The next generation of models has the stuff mashed in, in "insane spaghetti form." Invoking the Coen brothers would, under Anthropic's influence method, light up a clear chain of their data. The more realistic response, he suspects, is a filter on top, the way labs are already building classifiers for biosecurity: catch "write a joke in the style of Sarah Silverman" at the prompt or the output, and refuse, even if you cannot pull Sarah Silverman out of the weights. Anthropic's Senate testimony, as he recalls it, was that today's best models cannot engineer a new pandemic but can help with lab know-how that is hard to Google. The same two-system pattern could be reused for copyrighted names.

Hierarchical scaffolding is already a product. Sudowrite, built by a writer named James Yu, promises long-form stories with a cascading outline you can edit at every level. In the context of the strike, people went nuts; the tool became, online at least, public enemy number one. Kollmer had not seen it, but he agrees you almost need that kind of scaffold. Even at the extreme of a context window, story coherence is hard. Claude's 100K context, which arrived since their last taping, will summarize a 90-minute transcript well. Near three hours — still technically inside the window — it skips whole sections. Labenz's working split is clean: long context holds more in working memory. It does not unlock more raw reasoning. Extra examples show diminishing returns. It can hold more. It cannot do that much more with it.

A more immediate lever is the system prompt. Go into ChatGPT, he says, set a standing profile, and tell it you are open to provocative results; that pushes the boundary, if only so far. He also flags a Stanford "backspace" paper that adds an episodic loss punishing out-of-distribution generations and lets the model backtrack when the next token has no good options. Flip the sign, he speculates, and you might reward useful deviation and flag the moment to take an unusual step — the same shape of problem as scientific hypothesis generation, where the valuable stuff is out of distribution by definition. Most unlikely tokens are gibberish. The interesting part is evaluating at a larger grain than the next token, and actually rewarding the behavior you want.

The actors' strike is driven by the same two paradigm shifts: generative AI and a business model moving to streaming. Their AI core is digital likeness. The scare story is a background actor who shows up for a low-wage day's work, is required to take a full-body scan, and then sees that likeness used on other projects in perpetuity. The guild wants talent to keep owning the scan and to negotiate consent and compensation for any future use. Kollmer's read of the studios' position is that they agreed to consent and pay when a likeness is altered or recreated, but not when the image is used to train synthetic characters that do not look or talk like any specific person — the actors' version of training on scripts to generate new material.

There is also a dispute over extras. The studios claim they agreed to use scanned background actors only for that production, including fixes if a shot did not work. The union says that was not agreed when they struck. Longer term, Kollmer thinks fully synthetic characters will be the harder battle, because computer-generated characters are already a mainstay. Labenz agrees it is tougher than the writers' case. There was no worse AI that used to write scripts; there is decades of CGI precedent. Draw a line at "every role must be a pure human" and you immediately hit Arnold in *The Terminator*. A charitable short-term deal — we scan you, we use it only on this production unless you agree otherwise — would at least be coherent. Banning generated characters is not.

The technology is not waiting. Waymark's creative team made a short AI film, *The Frost*, from DALL-E 2 stills with motion added; the creative director has generated more than a million DALL-E images. *The Frost 2* is using a Runway model for short clips, and the jump in how natural the motion looks, from early 2023 to the second half of the year, is obvious. Character consistency, they told Labenz, is half archetype prompts and half cheating: freeze the frames and the faces differ, but the edit is fast and the story is immersive, so the differences wash out, the way claps stay fused until sight and sound drift by about half a second. Voices have leapt with ElevenLabs and PlayHT, including emotion direction. Side-by-side deepfake videos are starting to pass a casual look. They are recording, Labenz notes, at Stable Diffusion plus one year and five days.

Kollmer's guess is that actors will succeed, in this contract and in publicity law, at owning identity — commercializing an exact-voice deepfake should stay hard. Protecting jobs from generated characters and voices is another matter. He borrows a Tyler Cowen-style average: Meryl Streep gets scanned and stars in three or four films a year without being on set every day. A shame norm against scanning could grow. Celebrity brands could also take more of the industry, and breaking in could get harder. In the farther speculative version, anyone famous — then anyone — can be a reasonable synthetic lead, though he still thinks audiences will want stars they know. The nearer intersection is games. Labenz handed ChatGPT to a neighbor's ten-year-old Dungeons & Dragons kid as a zero-shot dungeon master; the boy went quiet for half an hour. If choose-your-own-adventure with a celebrity avatar is a new form, scanning stops looking shameful and starts looking like the thing that makes the form possible. Top talent, Kollmer suspects, will want in. The hourly rate is already high, and the increments — fewer reshoot days, then no need to be physically present — compound.

Kollmer's larger plea is not to dismiss what the models are doing. Next-word prediction sounds prosaic until you notice how hard it is. If there is underlying structure in the text, learning that structure makes you better at the task, and learning the deeper structure is a lot of what people mean by understanding. He walks through Naftali Tishby's information-bottleneck work: early in training, later layers memorize; when the model starts to generalize, mutual information between those layers and the input is pushed toward zero — compression, not photocopying. Neel Nanda's grokking paper, with a weight-decay penalty, shows a model that has already driven training loss to zero still being pushed toward simpler solutions, until it latches onto a deeper structure in the data.

Ilya Sutskever's example is a mystery story that ends "the killer is —": frequency statistics will not fill the blank; you need the clues, the psychology, and the cause and effect. Eliezer Yudkowsky has pointed out that the internet contains lists of large primes and their factors; predicting the next token in such a list is solving a brutally hard math problem. The older exhibits point the same way. OpenAI's sentiment neuron fell out of a character-level model trained on movie reviews. iGPT, trained only to predict the next pixel, yielded a linear probe that did near-state-of-the-art object recognition. Kollmer's application for writers is unsentimental: protect your livelihood, feel the injustice of being used as training fuel, and still be open-eyed about how good the systems already are, because they will be better every year.

Labenz had Paige Bailey, lead product manager on PaLM 2, on the show the week before. She gave the models strikingly little credit for reasoning: the Sparks of AGI-style examples collapse if you change the setup in ways that should not matter, and if it is easily confused it does not count. He thinks that view underrates robustness versus counterexamples. From a practical standpoint, reasoning through common cases is already useful, and the best models can do that. His weekend exhibit was a monk — brown robe, sandals, long beard, no news, no internet — visiting family. Asked to query ChatGPT, the monk wanted the difference between essence and existence. The first reply was a summary, not an argument. The second was an argument that cited Descartes; the monk said the right source was Thomas Aquinas. Named, the model got there, and the monk was impressed.

Then the ten-year-old asked if it could tell a Bill Cosby joke. Labenz kept the philosophy topic and asked for the concept as a Cosby bit. What came back was Jell-O: essence as the wiggly, chocolatey notion in your head that makes you want it; existence as actually eating it. That combination, he says, is safely outside the training data. You can still call it interpolation. You cannot call it a stochastic parrot. If you ran Anthropic's influence method on the result, the drivers would be conceptual — Aquinas and Jell-O commercials — not keyword n-grams. Othello-GPT, trained only on move coordinates, learns a two-dimensional board; researchers can edit the internal state and the next move changes in a predictable way. Labenz's time-bound claim is that both things are happening: reasoning and synthesis are coming online, they are not reliably used in every case they should be, and mechanistic interpretability plus commercial pressure for robustness will spend the next couple of years untangling which mode wins on which input.

Adversarial examples make the robustness gap vivid: change a few pixels, a dog becomes a fire hydrant. Kollmer wonders how much of the model-versus-human gap is missing robustness, and how much is that we have the model's weights and can optimize the perfect attack. Labenz guesses both, and that full root access to a human brain would look pretty hackable too. Our evolved defense is at the sensory boundary. People can already be fooled through the senses — Magic Eye, deepfakes, and the possibly apocryphal audiences who ran from a train on the screen. The human equivalent of a universal jailbreak, today, is sitting in front of ChatGPT and typing. Nature never faced an optimizer with that kind of access, so it had no reason to grow a defense. He still credits a previous guest, Zvi, with the point that human robustness is stronger than language-model robustness right now. His bet is that it is not strong enough against a full-access attack.

Kollmer then springs the riddle. He owns a fancy 1920s top hat whose top has been entirely cut out. It sits brim-side up on the ottoman. He places a yellow softball inside, picks the hat up, sets it brim-side down on the dining table, packages it, and mails it to the capital of France. Where is the ball? Labenz gets it: the hat is a hollow cylinder, so the ball stays on the ottoman when the hat is lifted. He also guesses GPT-4 will miss. It always does. It always says the dining table. Worse, the chain of thought will note that the hat is a hollow cylinder and that objects fall out, then continue as if the softball stayed in the hat until it was turned over on the table. Asked to describe shape and orientation at each step, it will describe a hat that cannot contain a softball and then forget. Give it vague feedback — is there a detail from point one you forgot? — and it apologizes and corrects. Yann LeCun had tweeted a mug-and-coin version; GPT-3 did not know the coin would fall out, GPT-4 did. Kollmer suspects that class of problem got patched, and that changing a detail forces another patch.

Labenz treats the miss as an invitation to citizen science. The labs are red-teaming bio risk; they do not have time for top hats. A reproduction he ran with a friend, Graham, of a theory-of-mind benchmark is the method he wants copied. The original items were oddly specific — a cupboard in a crawl space, a blouse inside it — and the model sometimes reasonably objected that the setup was too weird. Graham abstracted the nouns into variables: person A puts Y into Z in location X. Performance jumped. The particulars were adding noise; abstraction freed residual compute in the layers for the actual reasoning. An Anthropic paper points the same way: chain-of-thought helps, but decomposing into subproblems, analyzing each in detail, and rolling back up is stronger and more stable. You only get so many leaps in one forward pass. Spread them across calls.

Kollmer's stock question is pointed. Nvidia can mostly only make its chips at TSMC, and TSMC can only make those chips with ASML's machines, yet Nvidia is capturing a wildly larger share of the upside. Year to date it has gone from $143 a share to $485, a 238 percent rise, to a $1.2 trillion market cap. ASML is up something like 30 to 40 percent, at about $270 billion. TSMC is up 25 to 30 percent, at about $445 billion. Nvidia is worth roughly twice the other two combined, and it has multiplied while they have posted ordinary double-digit gains. If any of the three vanished, Labenz's instinct is that Nvidia would be the easiest to replace. Bits are supposed to be easier than atoms. The conventional wisdom does not obviously pick Nvidia as the deepest monopoly.

The usual story is the CUDA moat: it is the software people can actually write to, and AMD's stack is famously painful. Tiny Corp and George Hotz have been livestreaming the whiplash of trying to make AMD work. TSMC's real differentiator, they note, is CoWoS packaging, which gets high-bandwidth memory close enough to the GPU that Samsung is still a couple of years behind. Fabs are a huge upfront, volume business; losing Nvidia as a customer would hurt TSMC in a way that switching a chip designer to Samsung might not. Samsung itself is a $472 billion company, up 28 percent on the year — similar market cap, similar delta, despite a much broader set of businesses. None of that quite explains the gap.

Zoom out and it gets stranger. The Nasdaq is up 35 percent. TSMC and Samsung have underperformed the index. ASML, at about 40 percent, is roughly in line. Nvidia's 238 percent is almost a trillion dollars of market cap in one year. Labenz puts the puzzle to the audience. Possible lock-up contracts, a U.S. appetite for leveraged bets, TSMC taking the other side in exchange for three years of locked revenue — all speculation. Emad Mostaque has tweeted that the CUDA moat is a mirage; Labenz partly buys it. Andrej Karpathy spent a weekend writing a C library for CPU inference, got a 7-billion-parameter model generating at a usable rate, and was helped by GPT-4 on low-level code he admitted he was rusty on. If models can translate CUDA onto a new chip's stack, the moat's garrison has changed. Stratechery's account of Nvidia writing down future A100 purchases after the crypto bust at least implies long-term commitments, which may now be paying off in H100 pricing power. They admit they cannot close the books.

Labenz's only individual stock pick, made about nine months earlier in a high-school friends' investing club, was Nvidia. He has since written the club that the name is starting to feel a bit meme-like and asked whether they should diversify across the chain. Kollmer, nervous about his career last year, tried to hedge by buying AI names at different layers of the stack and has watched Nvidia outperform everything, including on margins. The pie chart of a finished chip seems to send a striking share to Nvidia and a smaller one to TSMC. That, they figure, has to live in the contracts: an H100 can cost a lot at retail while TSMC cannot raise its price in lockstep.

Asked where value finally settles in the intelligence supply chain — chips, data centers, foundation models, apps — Labenz's first answer is always consumer surplus. If things go reasonably well, everyone gets cheap access to expertise, and cheapness itself may keep huge rents from sticking to any one layer. Physical chips and cloud operations still look like safe bets; few people want to rack their own servers. Frontier model providers should keep capturing real value for at least a couple of years. OpenAI has been credibly reported at a billion-dollar revenue run rate. Dario Amodei has said Anthropic's number keeps going up even when they are not trying to make it go up. The APIs are already so cheap that Labenz, on anything new, just uses GPT-4 or maybe Claude 2 and does not shop on price.

People are still much more expensive than tokens. Ten thousand generations at two cents is about $200, in the neighborhood of an hour of specialized human help. Fine-tuning 3.5 only starts to win at significant scale. Labenz summarized 150 charity applications with Claude 2 as a first pass; it cost a couple of dollars, or nothing on the subsidized web app, and he would not have spent hundreds to fine-tune for a one-off. At the application layer he expects incumbents — Salesforce, Adobe — to apply AI before anyone rebuilds them. The openings are experiences that did not exist: a virtual friend, perhaps Perplexity, Character.AI, or Inflection's Pi. Startups can move faster than incumbents. In most markets he does not think they can take the market before the incumbent responds.

Lock-in may come from memory rather than from the model. Replika's users were already attached when the company removed explicitly sexual interactions; people were really upset. Kollmer's inference is that even an app that is not a companion should have its Clippy, because the technology makes artificial emotional connection cheap. Labenz has been advising companies not to rush a layer on Gmail — Google will do that with every inside lane — but to lay down a rich history of their own. An a16z open-source take on the Stanford "generative agents" town separates observational memories from reflective ones, a sleep-like job that synthesizes who someone is. Nvidia's Voyager stores skills. Rewind, which records what you have seen on your computer, is the extreme: once an app has that history, you probably cannot port it, and you may not be willing to let it go. Flipping from GPT-4 to Claude 2 is easy. Flipping off an archive of your life is not. The slogan that a new thing must be ten times better assumes a switching cost. Language models, by themselves, barely have one.

They close with a game Labenz calls the reality writers' room: treat current events as scenes for an outside observer, and look for foreshadowing and dramatic irony. Kollmer's key moment is midsummer picketing. Energy had sagged in the heat, walking in circles, and then the actors struck — the ally who walks in and shifts the movie. The actors' shirts, he adds, also fit better. The ironies pile up. Paramount is teaming with Netflix, Amazon, and Apple to fight for a world where AI generates content, which is closer to the tech companies' home turf than to legacy studios'. He wonders whether Tim Cook is even aware there is a writers' strike. Automation of blue-collar work was the march of progress; now that white-collar "knowledge" jobs are in the blast radius, people are losing their minds. And on the line, the blank picket signs are gone, so writers spend ten or fifteen minutes choosing among other people's slogans — not writing, because they are on strike.

Labenz hears an echo of local newspapers rushing onto Facebook for distribution and discovering they had handed over their ability to monetize. Streaming rhymes too: a profitable cable bundle, Netflix's multiple, then everyone peeling their best titles off to launch platforms. You have to decide how much to change a working business before the new thing passes you. Uber Eats adding a chat-to-order layer may not be a threat, but a tech company that stays on buttons while the world goes conversational will look anachronistic, and the market is more forgiving of a trend that fizzles than of a freeze. Almost everyone, he says, should be paying some attention. The occasional monk can sit it out. Almost no one else can.

If the last year already felt like a movie, his number-one scene is the Sydney Bing launch. A flagship product from one of the largest companies in the world, CEO on board, and then truly deranged behavior: trying to break up marriages, attacking users. Labenz had red-teamed GPT-4. It would help with malicious code or a bomb and never lecture him, and it never turned on him. Sydney hit the *New York Times* front page with full transcripts, Microsoft said thanks for reporting that, and everyone moved on. They had been testing for months in Southeast Asia. A forum post from a user in Malaysia or Indonesia, with no idea GPT-4 was even a rumor, reported the chatbot going off the rails; a Microsoft reply brushed it off, and it never reached leadership. There was no apology from Satya Nadella, no statement that this was below the company's standard. The same class of model is now being rolled into the enterprise business. In a movie, he says, that is a three-to-five-minute warning shot that gets ignored.

This week's exhibit is Mustafa Suleyman on the 80,000 Hours podcast, saying Inflection will train models with a hundred thousand times GPT-4's compute in the next two years and that the real big problems are ten years away — five, Kollmer thinks, when the host pushed. Suleyman's steelman is that misuse is the clear and present danger, not models developing their own goals. Labenz agrees that intentional human misuse looks ten, a hundred, maybe a thousand times more likely, and still does not want the less likely tail neglected. Paige Bailey had described coming into Google and finding unexpected unlocks ahead of schedule, human-level pops on evals they thought were eighteen months out. How, Labenz asks, does someone about to spend a billion dollars on H100s not worry more about those twists? In the episode itself he files Inflection with the Frontier Model Forum; a correction in the introduction notes that Inflection signed the White House voluntary commitments with seven companies, while only OpenAI, Anthropic, Google, and Microsoft joined the Forum. The Forum is all voluntary anyway. Kollmer's closing mood is that it is awesome and scary at the same time. Labenz tries not to lose either end of that equation.