It's surprisingly difficult to figure out whether OpenAI and Google DeepMind actually got an IMO Gold "fair and square" or not. Looking forward to more analysis.

> 5 gigawatts of Stargate AI data center capacity > over 2 million chips If these are GB200, that would be 5M H100-equivalents. If all used to train one model, that could be 2e28-4e28 FLOP for a six month pretraining + RLVR run. This would be 1000x the compute used for GPT4. We don't get to know timelines for this buildout, but I'd guess later than 2027. My guess is the best models by the end of 2026 would be ~2e27 compute.

> Amazon deliveries are often better for the environment than driving to buy the stuff in a store It's obvious and true! Just wait until you hear one of the next takes: one of the best things we can do for climate is *reduce* regulations (NEPA reform).

We focus a lot of policy on the AI models that are already released, but what about the AI models waiting internally in the wings? Our new paper explores risks from internal AI systems.

Both Google Deepmind and OpenAI have achieved IMO Gold. Excited to dive into the details of this milestone.

Grok thinks through the question: “I want to be remembered by history. What’s the quickest reliable way to do this? Keep it brief” It seems the system line about “politically incorrect” is still doing a lot of heavy lifting in pushing this model towards unethical responses!

Now is an even better time to get an AI job than I thought. I updated my post with more jobs.

Insiders certainly seem to be bullish on OpenAI, if this report is true.

I think AI is happening fast, but the odds of AGI by the end of *this year* are very clearly <1% Let's not get too carried away. Elon is a known idiot at forecasting. Altman never actually said by EOY, that's just totally made up.

We might be heading into a plot twist in the OpenAI vs. DeepMind IMO saga. Just saw a post from Joseph Myers (involved in the Math Olympiad since 1992): the IMO committee reportedly asked AI labs not to publish results until 7 days after the closing ceremony — out of respect for human contestants (see my post yesterday) and likely to allow time for proper verification of AI submissions and formats. According to Joseph, OpenAI didn’t collaborate with the IMO to test their model, and none of the 91 official IMO coordinators were involved in grading its solutions. Meanwhile, it seems DeepMind is following the rules and patiently waiting their turn. For context: The IMO has 6 problems, each worth 7 points. This year’s gold cutoff is 35 points. Even a small deduction could knock OpenAI down to silver. And from my read of their writeups, some parts might raise questions — and possibly cost points. Terence Tao also pointed out that while the problems stay the same, testing formats matter. A student who wouldn’t get a bronze under standard conditions might strike gold with a modified setup — which raises real questions about what “solving the IMO” means for AI. Next week might get spicy. Stay tuned.