Some rumors I've heard. I'll start with Google, then Anthropic and OpenAI. Google: Insiders say Google has either cracked RSI (recursive self-improvement) or is very close. Their pace has picked up fast in recent months. They went from GPT-5 level to Opus 5.5 level in months, and speculation is that RSI made that possible. Anthropic: About a week ago, Anthropic started a new pretraining run. It is not Fable 5.5 but what would become Fable 6, or something bigger, to catch up to OpenAI's Bel. OpenAI: Bel (it may have another codename) is the model behind the math results OpenAI published a few days ago, and RSI apparently helped there too. From what I'm hearing, OpenAI, Anthropic, and Google are all very close to cracking RSI or already have. It will get scary and exciting at the same time.

You aren't bad at machine learning math. You were taught by academics who prioritized dense notation over spatial intuition. Traditional math education forces you to memorize symbols before you understand concepts.

I've always assumed stimulants make smart people sharper, but it's the opposite. In a 2023 study, people on Ritalin, modafinil and dextroamphetamine worked harder on a complex problem, but the quality of their solutions dropped. Plenty of people report feeling like their IQ drops by few points on e.g. modafinil. I'd say these might feel useful for tedious work you absolutely disdain, but probably not for creativity.

“None of the humans involved in these proofs are available to take questions, give talks ...” If we turn every talk into a defense exam (torturing speaker by questions), we shouldn’t be surprised when nobody volunteers. Invite them. Let them present several results, with short AI-generated summaries of the proof ideas. Let the speaker and audience work through the mathematics together, suggest new problems, and help each other understand. Even before AI, plenty of talks were mostly advertising: grand claims about the results, plenty of muscle-flexing, and barely a slide on how the proofs worked. We shouldn’t romanticize that format. Now we could have talks that give us a view of several breakthroughs: AI generated ideas behind them, the techniques doing the work, and the connections worth exploring. That sounds exciting to me. Of course, someone can spend weeks digesting one proof and give a beautiful lecture. But why make that the price of admission to every seminar? There are many results posted that people once wondered whether they would see in their lifetime. Getting an overview of five of them may be a better use of someone’s time than preparing to defend every technical detail of one single result. “Here is what I understand. Here is what I don’t. Let’s figure it out together.” That should be a welcome opening to a research talk.