I have the best XRD agent in the world btw
Would’ve immediately flagged the Cu2S impurity in LK-99 for its phase transition mimicking superconductivity at 104 degrees before publication
Periodic Labs is going to have a lot of false positive claims of superconductivity in the next few years if Neon or some other braindead 1000+ H200 RL slop is what they’re using internally
Comment for access to Dorothy before I open source it next week
@lmtuwu this is fire and genuinely useful. i use a more advanced version of this in-house. people need to stop sleeping on the power of thoughtful surrogates to bootstrap robotics. underrated post
My entire Caltech thesis in physics was about studying phonons, and this is genuinely sad. Is this the bar for "novel discoveries and breakthroughs" now?
In superconducting theory, they prove some known conjectured bound on the critical temperature really is a bound.
1. This was proved twice in the last few months
2. No one uses this to discover materials... The constant in the bound was numerically calculated by physics Gods, Allen and Dynes, in their 1975 paper Transition temperature of strong-coupled superconductors reanalyzed.
This bound is only approached when the coupling, lambda, between electrons and phonons approaches infinity. In practice materials soften so much that they aren't even stable above around lambda = 4...
Any PhD would've told them that this asymptotic bound is rarely used in practice and a better problem would've been approximating some valid pairs of coupling and phonon frequency values before instability. Then you could better filter your candidate superconductors from simulation for testing in the lab.
There's a reason why no one rushed to prove the former problem for 50 years...
Announcing Generalised Superintelligence! (GSI) 🎉
GSI has achieved 250+ novel discoveries and breakthroughs across eight domains, including Quantum, Superconductivity, Chip Design and Mathematics through its Meta Solver Platform. GSI is extremely compute efficient compared to other frontier labs and excels at finding the hard structure in a problem first, before attempting it.
Everyone can contribute to scientific breakthroughs using the kind of superintelligence GSI is building.
Research preview out now at https://t.co/LSwU67gczE
This post is getting some attention so just want to let people know that we're hiring heavily right now! https://t.co/wpxsRjbx7C
We're basically doing what Periodic Labs is trying to do (and more) but we're actually going to achieve it because we're not NPCs.
Tap in if you're interdisciplinary and love robotics, science, or ML. Interview is something like debugging DR. GRPO by hand, some physics problems on the blackboard, an open ended research discussion, and a coding harness utilization screening.
We particularly index on thoughtfulness and people with unique backgrounds, however you might define that.
@sebuzdugan I mean I did say it was already proved twice in the past few months, so its not really novel. Also they highlighted this result in the comments of their launch post about "scientific breakthroughs" with "general superintelligence". All I'm arguing is that that wording is insane
@proofchecker@Akitti wut I think the navier stokes result is a real result LOL they proved the existence of a finite-time blowup with smooth external forcing.
Also I never said GSI's result wasn't a fine math result. They made a post about "scientific breakthroughs." I said this wasn't close to one
@denehoffman true haha... weird parallel but reminds me of the recent trend in the pokemon card space to overvalue "low population" graded cards while the prices corrected this month because people finally realized that low population means "low demand" to grade
I agree AI reviews are helpful, but they’re honestly kind of harsh, even to good work. They can’t seem to discern the fine difference between a limitation and a true flaw. For work that’s below 2 sigma above the mean, it can make the job of a human reviewer doing a final pass harder. I personally believe these are the types of subtitles text models won’t be able to navigate without overhauling the training procedure like we’re doing at https://t.co/1ApTKVqQRj
I’m really not sure why they didn’t just have a scientific advisor look at this post. The result I discuss is 1 of 2 they highlight in the comments of their launch discussing “scientific breakthroughs” so they seemed to think highly of it.
Anyway, beyond knowing what a useful problem to work on is, research in science at this level takes deep tacit knowledge and long horizon creativity. You can’t just throw RL at LLMs/VLMs (or in their case, a giant harness) and hope for the best.
My XRD model is better than Periodic Labs’ and I’m open sourcing it next week. You need to rethink how you inject physics into your models to be extremely data efficient. No one is doing this like we are with our interdisciplinary team at https://t.co/1ApTKVroGR
It’s more just how they label their work: Their company is called “Generalized SuperIntelligence" and I'm calling out 1 of 2 of their main results from their comments from their launch post about "scientific breakthroughs". This is nowhere near any of that so I think the ingenuousness is bad for the field. I’m generally against anyone claiming steps towards ASI without results at least pointing in that direction. This just further confirms that AI doesn’t even know what a useful problem to work on is.
@mm600925909104 the intersection of hardcore scientists that can build frontier models is very low though. that's what we're trying to change with https://t.co/Zre3DNTEAB by bringing these people together
I appreciate the response and open dialogue, most people wouldn't respond so you're the goat for that tbh. But uhhh your company is called "Generalized SuperIntelligence" and I'm calling out 1 of 2 of your main results from your comments from your launch post about "scientific breakthroughs" so I don't think it's really radical for me to believe you unironically think this will transform superconductivity
Yeah this is the kind of intelligence that takes deep tacit knowledge at the frontier of knowledge. Good advisors in academia have this, and can steer students to work on problems that are important/useful but also have enough footing to make headway on. People tapped in in SF also have this. VLMs are so behind even after saturating text benchmarks. Theoretically RL can handle this? But I don't really see it scaling given how long rollouts take and sparse rewards are. Yeah... AGI is far away if people don't change their mindset. I'm trying to change that at https://t.co/1ApTKVroGR