Gemini 4 Argon is our next era of frontier intelligence. It shows significant improvements across benchmarks, setting a new state of the art for real-world long-horizon software engineering tasks.

An extremely instructive development in the use of AI in theoretical physics just emerged from the University of Chicago, thanks to Dam Son-- everybody and anybody seriously interested in theoretical physics should read this and ponder its implications

World seems to have woken up to the idea that AI capex actually needs a return at some point. People now realise that roughly $2-3trn of revenue will be needed annually, quite soon. But the conversation about what counts as "AI revenue" is still highly unsophisticated. People will say e.g. "Amazon's cloud business is doing really well, ergo AI is seeing great ROI". No. What you need is revenue EXTERNAL or DOWNSTREAM to the system, ie from ordinary businesses subscribing to models. @darioperkins good on this. A lot of Amazon cloud revenue is just Anthropic buying stuff How much is "proper" AI revenue? In the summer we made four estimates. We need more work to make these estimates better/more accurate! @hsu_steve @lugaricano @weakinstrument

BOTH THINGS CAN BE TRUE AT THE SAME TIME: 1. AI capabilities are real and advancing rapidly 2. It will take longer than singularity or fast-takeoff people think for AI to affect real productivity in the economy, or generate $trillion revenues for AI labs (although capex investment for the future already drives this much money into data center and semiconductor infrastructure). #2# is related to human gatekeeping and outright resistance to technology, as well as the difficulty in understanding the best way to use AI in practical settings. It can take years for existing companies (let alone educational institutions) to adapt to new technology.

Physicist @hsu_steve warns humanity may surrender power to AI completely voluntarily because the alternative is getting crushed by competitors who use it: "Gradual disempowerment, which I think people are intuitively aware of, but usually they've not seen it fully formulated this way." "Every single element of our society, whether you're a corporation, a military, a government, or even an individual trying to do well on the dating apps, if you have access to a superior intelligence, an AI, that you need to compete against all the other players, you're gonna start using that AI." "Gradually, even if nothing goes wrong, even if there's no evil AI or there's no AI hacking your systems, you literally are gonna surrender power to it because you will get better outcomes. If you don't surrender power to it, your competitors are gonna beat you. They're gonna crush you." "That's gradual disempowerment. It's gonna happen completely voluntarily as people just hand over power to AI."

🦔Oracle's 5-year credit default swap spread just hit a record 230 basis points, more than four times the broader investment-grade index at 55 bps. The spread has quadrupled in the past year and now exceeds financial crisis levels from 2008. For comparison, Turkey's sovereign CDS sits at 245 bps. Oracle is one notch above junk and the credit market is pricing it as a bigger risk than it was during the worst financial crisis in modern history. My Take Oracle's CDS is priced worse than 2008 and Ellison is still signing hundred-billion-dollar AI deals. I think he's operating on the assumption that Oracle is too big and too embedded in government IT systems to be allowed to fail. That may end up being true, but it's a bet that someone else covers the downside rather than a financial strategy. S&P cut Oracle to BBB- in July and said they "underestimated the scale of the investments required." One more notch and Oracle is junk-rated. Every institutional fund with an investment-grade mandate would have to sell, and that's a forced liquidation Oracle isn't prepared for. Half the backlog is OpenAI, which by its own pre-IPO numbers burns $278 billion more than it takes in through 2030. I don't think Ellison has a plan for what happens if S&P downgrades him and OpenAI can't pay on schedule at the same time. And I think both of those are more likely than the stock price assumes. Hedgie🤗

This could be a problem: Goldman sensitivity table shows that in a worst case scenario where hyperscaler ROIC on capex is 0 (token costs collapse, token demand goes to open models, etc) they still need to spend $920BN every year just to cover depreciation and running costs

Wow, I just found out that our team has been scooped by @AnthropicAI See the addendum of the post “How does it feel to be scooped by a machine?” written by my collaborator Lance Dixon (who was under NDA while checking the result) I’m still processing, but not entirely surprised. I do I wish we had the time in the resources to commit to it.