Tue 22 September 2026

On an hour-long show called Redacted, the guest Zach Vorhees told viewers that Senator Bernie Sanders had proposed twenty years in prison for anyone who so much as fine-tunes an AI model. He said the proposal makes no distinction between building a model from scratch and adjusting an existing one for your own business. He said the threshold is undefined, and offered an example: an older version of ChatGPT could write an email faster than he can, so would that count? He said this was literally what the senator had written. He quoted no text, and the programme showed none.

The bill is real. The Ban Artificial Superintelligence Act, introduced on 4 September 2026 by Sanders with Representative Greg Casar, carries up to twenty years in prison for individuals and what its sponsors call a corporate death penalty for companies — penalties they compare openly to unlawfully developing a nuclear weapon. It would also pause advanced AI development until a new cabinet-level regulator sets safety rules.

His quotation of the definition is close to accurate, and then drops the part that carries it. The bill reaches a system that exhibits, or can easily be modified to exhibit, capabilities matching or exceeding human cognitive performance across a broad range of domains or tasks. Broad range of domains or tasks. The email argument only works if the threshold is any single task; it is not. Reporting on the bill says today's frontier models do not meet the definition.

The fine-tuning claim is unsupported by any source we could find, and the reason is structural. The bill is written around what a system can do, not how it was built — which is precisely why it draws no line between base training and fine-tuning. It does not regulate by training method at all. He read the absence of a distinction as the two being lumped together. It is more likely that neither is the hook.

There is a serious criticism of this bill, and it is not the one the show made. Two provisions are genuinely loose: can easily be modified to exhibit reaches a model that is not superintelligent but could readily be made so, and the temporary pause on advanced AI development is reported as undefined. Gary Marcus — who has spent years arguing that AI is dangerous, so is attacking from the opposite direction — opposes the bill on exactly those grounds, calling it too broad and naive about how hard these capabilities are to measure.

One caveat on our own checking, recorded because it is the kind of thing that usually goes unsaid. The senator's own website refused our requests for both the press release and the bill summary, and we did not locate the text on the congressional record. This rests on secondary reporting — though the definitional wording is quoted identically across several unrelated outlets.

The same programme makes an argument that does not depend on anybody coordinating anything. Vorhees claims DeepSeek has released a model delivering close to frontier performance at roughly 1.4 to 1.5% of frontier pricing. His analogy is an electricity bill falling from $250 a month to $2.50.

He then claims that about 1% of customers provide roughly 80% of the revenue at the large AI companies — big enterprises paying one to two thousand dollars a day for the volume they get through. If that 1% can be served at a fraction of the price, the revenue base goes with it.

The mechanism he describes is distillation, and he explains it plainly. A cheaper laboratory subscribes to an expensive model, asks it enormous volumes of questions across many fields, and trains a much smaller model on the answers. His claim is that the result runs on high-end consumer hardware — that a business paying $100,000 a month for access could instead spend about $10,000 on graphics cards, once.

From which he draws his conclusion, and it is the sharpest line in the hour: the last thing a dying industry does is crawl to Congress for regulation. On this reading the safety push is not alarm but a moat, built by companies who cannot out-innovate a competitor selling the same thing for a fiftieth of the price.

None of these numbers is verified. They come from one person, on one programme, with no documents shown, and the same speaker gets other things demonstrably wrong. But unlike the coordination claim, this one is checkable — and if the pricing and the revenue concentration are anywhere near right, the commercial picture is different from the one usually described.

His one constructive proposal sits here too. He points to South Korea, which he says is moving towards free AI compute for its citizens, treating access the way a country treats water or electricity, and argues for a right to compute as part of a new social contract. He also argues that open models should not come only from China, and calls on American companies to publish theirs.

The premise, stated in the opening seconds: a researcher resigns from an AI company warning the world is at risk, a magazine puts AI on its cover, a documentary appears, and two senior political figures issue warnings — all inside about a week. Therefore it was arranged.

No document is produced. No funding trail, no insider, no mechanism. The entire argument is that several things happened close together. The host says twice that he initially believed the warnings himself before deciding they were staged.

Then, near the end, the guest says that AI theoretically does have a way of killing us, offers a scenario in which a system threatens to destroy itself unless left running, states that there are real doomsday scenarios, and accepts that the most advanced laboratories will and should face government pressure to slow down on the most dangerous work. Which is, in substance, the position he has spent the hour calling manufactured.

He also says flatly that AI is underhyped and there is no bubble, while predicting in the same conversation that the same companies' valuations and share offerings will collapse. Both cannot be the settled reading of the same week.

Two things are worth knowing before this entry is used for anything. The programme is political commentary rather than technology reporting, and its own disclaimer describes the content as for entertainment purposes only. And the transcript is machine-generated and mangles names and products heavily, so no name or figure taken from it should be treated as accurate without going back to the video.

The archive keeps it anyway. A reference that records only the sources it agrees with is not a reference.

Adam Foroughi runs an advertising business inside mobile games — introduced on the All-In podcast as probably the best founder no one has ever heard of. His figures, all his own: over a billion people a day play casual mobile games, mostly adults; his own platform carried around $11 billion of advertising a year when he last disclosed it and has grown roughly 60% a year since, putting it near $20 billion; the whole mobile gaming advertising market he estimates at about $50 billion a year.

His thesis is stated in four words. Advertising is ML 1.0. He argues that advertising systems were the first commercial use of the techniques now driving AI, that recommendation engines and large language models are built differently but follow the same path, that research crosses between them in both directions, and that many of the researchers working on language models started their careers on advertising systems.

The structural advantage he claims for advertising is feedback. When you predict an outcome and then sell against the prediction, you learn immediately what the prediction was worth. He concedes the value of a language model to society is far greater — but says advertising is deep learning's most profitable application.

The recovery of his own company is his evidence. Going public in April 2021 at about $28 billion, the shares fell through 2022 to a low near $3.8 billion — a 92% drawdown — in a year the company says it earned a billion dollars. He stopped speaking to investors entirely on the grounds that they were not buying, and spent roughly $6 billion buying back his own stock. Then in April 2023 the company replaced its old statistical model with a deep learning one, and because everything is priced on performance, better predictions meant advertisers spent more.

The part most relevant to this archive is his answer on moats. Asked why competitors have not undercut his claimed 84% margins, he says complex technology plus distinctive data compounds — and reaches for AI to make the point, unprompted: by that token Anthropic should not be running away with the language model space, but a model that reaches scale and gets adopted by a large community becomes a moat that is hard to overcome.

That is an operator's claim rather than an analysis, made by a man whose own margins have survived a decade of people who could have competed them away. It is evidence of a kind, from an adjacent market.

Foroughi splits advertising in two. At the bottom of the funnel you already know what you want and are researching the purchase. That is Google's search business, and it is what an advertising product built into a chatbot will compete with.

His argument is that this creates very little new economic activity. If you were always going to buy the shoes, closing the loop faster moves the transaction; it does not create one.

Discovery is the other half: showing somebody a thing they did not know existed. That, he says, creates demand that was not there before, and it is what makes Meta's advertising business what it is — and what his own company is moving towards.

On shopping agents he is a sceptic, and openly has a stake in being one. He accepts agents will handle repeat purchases — his example is putting a supplement subscription on autopilot. But he argues the ordinary shopper wants to browse, compare, buy and track the parcel, and that the dopamine hit from going through it is the point. Told afterwards that an agent could have saved them 20% on a $50 purchase, he does not think they would care.

His framing of who is being designed for is the line worth keeping: we over-index on the Twitterverse and forget the average shopper is not that. He describes his audience as the New York Times audience, and notes how many people still use Yahoo's websites every day.

The hosts put the suspicion everyone has to a man who would know: you mention something at lunch and the advertisement appears that afternoon.

His answer is no, with reasons. His company does not track location at all — he calls precise location tracking to serve an advertisement a really heavy concept. And transmitting microphone audio at that scale and parsing it into something sellable is, in his word, not realistic.

What is actually happening, he says, is that you did something trackable and forgot: a search, a browse, a product you looked at. Then you mentioned it. On platforms that hold your relationships, what the people around you looked at can feed into what you are shown.

He is answering for his own company, not for the industry, and he is not a neutral party. But the explanation is the interesting bit: the effect people find eerie is being produced without the microphone, from ordinary traces nobody remembers leaving.

His position on privacy rules follows from it. He wants them clear rather than absent, because once they are clear the technology adapts — and he notes that after Apple restricted targeting, his company heard from users complaining the advertisements had got worse.