Working on math AI at https://xitter.cf/t.co/u95v5xJnFC and telescope software at https://xitter.cf/t.co/Yx0Z8UGvEc. firstflagPOISONed. Formerly: Parse cofounder, Facebook, Google
Piedmont, California
Joined March 2008
- Tweets19.5K
- Following1.5K
- Followers8.5K
- Likes28.4K
The UI behavior on chatgpt.com is strange right now. If you have a browser tab open in "Codex", open a new tab, and click "Home" in that new tab, then it also switches the UI in your original tab.
Mathematicians used to study Euclidean geometry. It was beautiful. But coordinates and calculus together became a more powerful tool that surpassed it. Even though they were often very ugly and involved a lot of grinding work, they made Euclidean geometry essentially obsolete.
Effective Mathematics: instead of thinking about math directly, you go work in finance or something, make a bunch of money, and then spend that on tokens, asking the AI to solve some unsolved math problems
I agree with Tsimerman's optimistic view of math + AI.
"Pretty soon we're going to have proofs that no one understands... and based on that mathematics, other things are going to work. Does a train run, even if no one understands the mechanics by which it runs? Yes, it does."
Jacob Tsimerman, who won the Fields Medal this year, thinks mathematics will soon rest on proofs no one understands. He says it’s already true of many of his own theorems, which build on accepted results whose proofs he doesn’t know.
Full episode: youtu.be/6uIJdXmB4vE
Terry Tao: "Problem-solving alone is now actively harmful to mathematics."
I would like to politely disagree, and encourage everyone to continue solving math problems.
teorth.github.io/tao-web/sli…
The funny thing about the “moving the goalposts” analogy is that NFL kickers have gotten so good recently, it might actually improve the game to move the goalposts a bit
Traditionally, all math departments are looking for similar qualities and hire in essentially the same way. Now that AI is flipping over the table, I think it would be exciting if different math departments ended up hiring differently, and having different styles.
The OpenAI math drop has led to a flurry of collaboration.
"Human experts have been key to seeding the insights for breakthroughs that push the models to out-of-distribution solutions. Even people without experience are contributing."
The future of mathematics.
We've hit a "wall" κ ≈ 7.004e-4 on integer multiplication:
1) The current best algorithm is built from two parts that have to work together: a “bit” part and a “complex” part, which does the Fourier-transform-style work. Each has its own score, called its "saving."
2) The final κ is set by whichever saving is lower.
3) For days, everyone has been upgrading the bit part: better frame layouts, packing leftovers into “banks”, reusing dead registers, etc.
4) The complex part has been stuck at the same score, 7.009e-4, since PR193. Every PR since then uses that exact number.
5) So the bit part is now within 0.8% of the complex part. Push the bit part past it and κ doesn’t move: the complex part becomes the weak link. That’s the wall. 🧱
A few have proposed solutions beyond the wall using conditions, but haven't solved it yet.
Human experts have been key to seeding the insights for breakthroughs that push the models to out-of-distribution solutions. I've noticed even people without experience are contributing to the leaderboard with persistence.
If you're an expert or just a random person, consider pointing your Codex Astra high at this repo and let it rip on wild new ideas to break us past the wall: github.com/CrocSwap/integer-…
BART management wants you to believe that without more money, they'll have to shut down. I don't believe it. Privatizing BART, requiring the buyer to maintain SF-Oakland rail service but allowing them to redevelop BART's real estate, would generate billions of dollars.
Kevin Lacker retweeted
Reading all those recent beautiful essays by Terry Tao, Kevin Buzzard and @wtgowers, I could not resist, but write up my deeply personal perspective on how I perceive and see things:
docs.google.com/document/d/e…
You know a new feature is good when it has dozens of bugs and nevertheless I still want to use it constantly (cloud codex)
For a personal assistant, I really don't need superhuman intelligence. I would prefer mundane reliability.
Kevin Lacker retweeted
Starship will expand the aperture of physics research by placing much larger telescopes in orbit and giant ones on the Moon
Starship reached orbit for the first time on Monday. One of my favorite potential use cases is launching the next generation of space telescopes faster and with far fewer constraints.
Today, engineers designing the largest space telescopes face extreme mass and volume limits that forces them to make telescopes lighter, foldable, and dependent on incredibly intricate deployment mechanisms, adding years of engineering complexity and increasing the risk of failure.
Starship changes that tradeoff. Thanks to its massive payload capacity and volume, telescope builders could "convert risk into mass": build heavier, simpler observatories instead of optimizing every component around launch constraints.
That could accelerate the next generation of space telescopes, and astrophysics research, by decades!
MongoDB's CEO is leaving to go to Meta.
In the past, I liked MongoDB a lot. For a programmer, it is much easier to use than SQL. But with modern AI, that doesn't matter any more. The format and workflow differences become trivial. So what does MongoDB still have going for it?
Okay here's the b2b SaaS devtool I want. "Testing as a service."
Agent-first. You tell your agent, use product X for testing, and it works.
You don't have to use git branches or anything. The agent just sends a diff and gets back the build output.
Does not exist yet AFAICT
This is the best article on RSI that I have read, using metrics to analyze both the definition of RSI and how close we are to it.
noahpinion.blog/p/wheres-the…
Eventually we will have AI compilers, in the sense that AI will be able to generate:
1. the intent, in human language
2. the algorithm, in a high-level programming language
3. the implementation, in machine code
4. a proof that 2 and 3 are equivalent
