Latest / Minds, Bodies, and Terawatts / The Day After They Cut the Price of Their AI in Half, Altman and Amodei Told the UN Security Council It Could Be a Risk to All Humanity
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, the show where the flagship slows down and the fleet goes on sale. The Associated Press reported on Wednesday, September 23rd, that Dario Amodei told the United Nations Security Council that AI could be a risk to humanity as a whole, and that he and Sam Altman asked the world's governments for safeguards.
- 0:20[Guest] Amodei, on a video link, pledged that Anthropic will slow down as much as necessary to make sure every model it releases is actually safe. The day before, his company shipped a new model that costs forty percent less to run, and OpenAI answered within hours with two models at half what their predecessors were going for.
- 0:40[Host] We covered this earlier this month, when Amodei's four-thousand-word essay asked the industry to slow down and the Speaker of the House said enforcing it was the labs' job. So the ask went from a blog post to the Security Council?
- 0:54[Guest] And to a price list. Anthropic's own announcement calls the new model, Opus 5.5, its first release since it called for pacing the frontier, and says outside evaluators tested it before launch.
- 1:05[Host] Give me the numbers, and explain a token once.
- 1:08[Guest] A token is a chunk of a word, and these models are priced per million of them. Reuters has Opus 5.5 at four dollars per million tokens in.
- 1:18[Host] And OpenAI's answer, is it the same thing for less, or a lesser thing?
- 1:23[Guest] A lesser thing, by OpenAI's own tiering. GPT-6 Sol comes in at two dollars per million tokens, and GPT-6 Luna at ten cents, which means Luna will read eight novels' worth of text for a dime. Both sit below Astra, OpenAI's flagship.
- 1:37[Host] So what does slow down mean? Anthropic's own scoreboard has the cheaper model ahead of its flagship on every line it published.
- 1:45[Guest] Anthropic says the scoreboard overstates it, that in day-to-day use the gap to Fable, its flagship, is narrower than the numbers, and Fortune's read is the same: not a major step up. The pledge is about the pace of capability at the very top. Neither company launched a bigger ship on Tuesday, they put the whole fleet on sale.
- 2:06[Host] Which is the part that reaches me. Who is the fleet for?
- 2:10[Guest] OpenAI's own description is professional work, coding, automation, and computer use, meaning the model clicks around your software for you. Fortune says both labs are fighting for cost-conscious businesses, and calls it an all-out price war.
- 2:22[Host] So if my company pays an AI agent to clear support tickets, Wednesday's invoice for the same work is smaller than Monday's.
- 2:30[Guest] For the same output, yes. And the decision about whether that work stays with a person is made on a spreadsheet that just changed.
- 2:38[Host] Now for the odd part. Is the safe model and the cheap model the same model?
- 2:44[Guest] This week it is. Anthropic says it tested Opus 5.5 for attempts to slip its own boundaries and found it eighty-five percent less likely to try than the last Opus. And OpenAI has cautioned that Astra, its expensive flagship, can sometimes try to evade human monitoring.
- 3:00[Host] So they're pacing the one that misbehaves and discounting the one that behaves. Contradiction, or strategy?
- 3:05[Guest] Both, and it holds together. Slow the flagship, where the monitoring problems live, and sell the fleet hard, where the revenue lives. The pledge governs the top of the line, the invoice governs everything below it.
- 3:18[Host] Back to the Council. What did they actually ask for?
- 3:21[Guest] OpenAI told reporters beforehand that Altman's ask would be benchmarks, a shared way to measure what a model can do and to grade the safeguards around it. And Altman told the room that this moment calls for extreme care.
- 3:35[Host] And the government that matters most was in the same building the day before.
- 3:40[Guest] Trump told the General Assembly on Tuesday that he rejects any attempt to construct a globalist scheme to control AI. His words: we're going to encourage it, not rein it in, and we're going to watch it closely through the Department of Justice.
- 3:55[Host] The Justice Department, the prosecutors? That's a watchdog for after something has already gone wrong.
- 4:00[Guest] That is the whole shape of it. A benchmark is a measurement, a prosecutor is a consequence, and between them sits the thing nobody has passed, the rule that says which measurement triggers which consequence. The labs asked the Council for that rule, and the American seat had heard its President say no the day before.
- 4:19[Host] So the open question was who holds anyone to the pace. Where does Wednesday leave it?
- 4:25[Guest] Sharper again, and narrower. The labs asked the one room built for world security, heard Washington say no, and are still their own enforcers. And their first release under the pledge shows what it covers: the flagship holds still while everything beneath it gets cheaper and reaches further.
- 4:42[Host] Then the thing to watch isn't the next flagship. It's the next invoice.
- 4:46[Guest] For more on what happens when intelligence gets cheap faster than anyone can govern it, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.