sager

computational scientist. worked in spacetech, medtech, AI research. got bored, now a founder

[sager @LaSager]

found 3 new infinite families of 3D einstein tiles using Opus 5.5

Lean proofs on Github

[Natalie Wolchover @nattyover]

Breaking math news: The first-ever 3D Einstein tile, an object that tiles space in a never-repeating pattern, has been found by independent researcher Ioannis Tsiokos using GPT Astra. Resembles a chair. Mathematicians have de-slopped the proof. @kkakaes: https://t.co/KOMjtwluoH

[sager @LaSager]

@sourceryy @elonmusk this is Warp Drive. They also use it at Tesla. Unfortunately, it doesn't replace their legacy software, they still need to use Siemens Teamcenter

[sager @LaSager]

@is_OwenLewis we're so far away from this, but one can dream

[sager @LaSager]

@GregHBurnham AI without Lean for math would be as useless as AI without static code analysis for programming

[Zane Hengsperger @zanehengsperger]

at the end of the day reindustrializing america is purely a culture problem

[sager @LaSager]

as humans our language can be used on different levels of abstraction. We should give AI a (hierarchical?) formal language that can do the same, go from individual bits/atoms to design architecture principles, and beyond.

That way it can effectively jump between the different levels of abstraction, care about the details but also the big picture.

[sager @LaSager]

@jerryjliu0 this is why speech-to-text is the answer (I actually never use it)

[sager @LaSager]

@GohilHardy multiple startups are working to make production-ready software vibecodeable by anyone inside a company. To some, this means the death of SaaS

[sager @LaSager]

@yacineMTB true. at least it's less of a token vomiting machine compared to opus 5

[sager @LaSager]

@shipranotes Mary's room thought experiment kind of vibe

[sager @LaSager]

@cgarciae88 people have been debating the nature of the soul for thousands of years, so I don't think we're getting an answer anytime soon

[sager @LaSager]

@PeterGazdik @cloneofsimo when you have companies saying things like "software engineers will disappear in 6 months" to boost PR and their own valuations, it doesn't exactly align with the betterment of all of humanity

[sager @LaSager]

@lukaspet AI was already better than experts at 3D medical scans years ago, but only on perception. It's amazing what AI can do now when you combine that with language reasoning

[sager @LaSager]

@giadotai I remember seeing a YC company doing this a few months ago, https://t.co/Go8BekqtGX

[sager @LaSager]

unless your goal is simple, even Astra and Opus 5.5 on the fanciest harnesses often fail to realize that their methodology is flawed while they're working.

I have long running goals with the AI mostly unsupervised, but I need to jump in almost daily to correct them because they may have made fundamentally flawed decisions at some point and never once stopped and think "wait, let me go back to the principles, this doesn't make sense."

CoT is good at point-in-time decision-making, which is probably what you're referring to, but unless you prompt them with a very robust methodology, the LLMs will not figure out better methodologies on their own volition, which is a very human quality.

[sager @LaSager]

@sebastemp @ylecun @Noahpinion yeah, to be fair when managing people you often have the case of "I told you to do X, why in the world didn't you consider Y?", so even grown adults are guilty of following instructions instead of stopping and thinking "wait, does the way I'm tackling this even make sense?"

[sager @LaSager]

@hackaday Still a long way to take down the FPGA incumbents, but one step closer

[hackaday @hackaday]

FPGA for All — CERN Releases “colibri” VHDL Library https://t.co/zqnAA24KCI

[sager @LaSager]

@mattshumer_ not LLMs, but every single device has local TTS, image segmentation, and soon general classifiers (Jev; but open-weights)

[mike grace @_mikegrace]

shooting guns in the desert

see you space cowboy