When the former Windsurf employees now at Cognition eventually start their own companies, THOSE are the founders I’d want to work for. They’ve lived the sacrifices early employees make, and I believe they’ll fight to do the right thing for their teams.

Intelligence isn't a collection of skills. It's the efficiency with which you acquire and deploy new skills. It's an efficiency ratio. And that's why benchmark scores can be very misleading about the actual intelligence of AI systems.

this IMO gold will fly past us as quickly as the turing test did. soon normies will say “duh of course they’re good at math, they’re computers”. but the RL breakthroughs the team made to solve math (congrats!!) will likely generalize to environments with much higher direct value.

Their bet allowed for formal math AI systems (like AlphaProof). In 2022, almost nobody thought an LLM could be IMO gold level by 2025.

Wow. Just used @OpenAI agent to find an airbnb for an event in October and ended up booking the agent’s 1st recommendation. Gut checked by scouring airbnb manually before looking at its results (to test it). Agent got it right on the first try. Took agent ~10 min. Normally.

Several months of serious upskilling is often all that stands between boredom vs delight, silly mistakes vs flawless execution, empty aspirations vs a consistent habit. It's just like physical workouts. You can't expect your body or even your attitude towards exercise to.

OpenAI finds itself in a dilemma. It and Sam really want to be the main character in AI every day, and will likely not be satisfied by releasing a nothing burger model after Google and many startups (👋 Kimi) releasing crazy-performant open models. But if they release a better.

"Almost certainly within the next five years, any economically valuable job humans can do, AI will be able to do 80% of it, with a few exceptions like heart surgery or brain surgery. But by and large, 80% of all jobs can be done by an AI."

Also the “intern” trick. Saw it so many time on the “founder slide” and then found out with simple due diligence. I once gave some feedback about it 🤦‍♂️ IMHO it is fine if you had an asterisk for full disclosure but most don’t.

Just saw someone with "worked at Google DeepMind" in their profile. Turns out they participated in Google's Summer of Code

I love the attitude that @obsdmd gives if you try to enable Vim mode in Settings 😆

sad to report that it turns out python environment management didn't have to suck this whole time and UV is actually great.

RT to help Simon raise awareness of prompt injection attacks in LLMs. Feels a bit like the wild west of early computing, with computer viruses (now = malicious prompts hiding in web data/tools), and not well developed defenses (antivirus, or a lot more developed kernel/user.

Congrats to Simon Willison (@simonw) on 23 years (!!) of blogging. Really excellent LLM blog, I sub & read everything:. (e.g. I sub via RSS/Atom on NetNewsWire). If you consistently enjoy the content like I do, sponsor on GitHub:

With OpenAI dropping the price of o3 by 80%, today is a great reminder about how important it is to build for where AI is going instead of just what's possible now. You can now get 5X the amount of output today for the same price you were paying yesterday. If you’re building AI.

improving LLMs on long horizon tasks with RL is the singularly most interesting problem to work on right now.

Apple just GaryMarcus'd LLM reasoning ability

We don't even know how to train humans on long horizon tasks with sparse reward signals effectively.

Yesterday I volunteered at @aiDotEngineer and I'm sharing my AI learnings in this blogpost below. Tell me which one you find most interesting and I'll write a deep dive for you. Key topics.1. Engineering Process Is the New Product Moat.2. Quality Economics Haven’t Changed—Only.