alex peysakhovich
@alex_peys
Followers
5K
Following
10K
Media
216
Statuses
2K
partner sutter hill ventures. ex-facebook ai. interests: dogs, motorsports, ai for science, game theory
Joined February 2013
leaving some generative chemistry models to cook on the gpus over the holiday break
0
0
1
my experience is that i have had many clever ideas fail but it is rare that a stupid idea scaled up fails
My machine learning education has progressed to the point where I lose sleep tossing around a “brilliant” idea, only to find the next day that it doesn’t actually work. This is great! I talked about this a few years ago:
0
0
1
nvidia is really leaning into very cool open source, obviously makes sense given the company mission (sell gpus!) - it's great when public and private incentives align
Introducing NitroGen, an open-source foundation model trained to play 1000+ games: RPG, platformer, battle royale, racing, 2D, 3D, you name it! We are on a quest for general-purpose embodied agents that master not only the real world physics, but also all possible physics across
0
0
1
i don’t even submit papers anymore but i donated to openreview, you should too
OpenReview is a lifeline for progress in the AI research community, and it urgently needs our increased support. https://t.co/HJJNRcl9km In 2025 alone, OpenReview supported over 1,300 conferences and workshops, served 3.3 million active monthly users, handled over 278,000 paper
0
1
7
Now we have high-throughput screening + human interpretability + new capabilities (e.g. dynamics). Stay tuned as we show off more cool stuff at the intersection of aging biology, chemistry, and AI! MONET Paper: https://t.co/IyxmUQvBBI Our blog: https://t.co/li0N1MVNZA fin
integratedbiosciences.com
Most of us can guess someone’s age just by looking at their face. Wrinkles, gray hair, skin tone. These subtle visual cues betray the passage of time with surprising accuracy.
0
0
2
Go short as easily as you go long. Express your market view nearly 24 hours a day with E-mini S&P 500 options on futures (ES) from CME Group.
2
8
53
A special "reference consistency" architecture lets us virtually stain both still images and timelapse videos. Interestingly, this reference consistency also allows for some in-context learning across cell lines and imaging setups. 6/n
1
0
2
We want to keep the human spark in our screening process. So we built MONET (Morphological Observation Neural Enhancement Tool): a diffusion model converting brightfield → virtual cell painting. You can see many examples from the model at: https://t.co/xK5giYkkGH 5/n
thiscellpaintingdoesnotexist.com
Integrated Bio's first generation virtual cell painting playground.
1
0
2
The above images use cell painting—which rich, high-contrast images. It has downsides: labor-intensive (painful for high throughput) and it kills cells (can’t see dynamics). What if we used just brightfield images? We as humans found it much harder, a model worked fine. 4/n
1
0
2
Turns out, in the same way that you can look at someone’s face and tell their age, you can do the same with their cells! Images below show fibroblasts from different ages. There is a very clear pattern. And, if a person can do it, we can train a model to do it to over 90%
1
0
2
Secure your crypto with self-custody Buy a Biometric Wallet up to 37% off Get up to $75 in rewards Airdrop apply starting Dec 17 (UTC)
0
2
11
To implement high throughput screens we needed to actually measure whether a compound had the effect we wanted - did it make cells "younger"? Assays like transcriptomics or methylomics are getting cheaper but are still too expensive if your goal is screening millions of
1
0
1
Working with @IntegratedBio we are finding small molecules that can positively impact aging and age related diseases. Today we're pulling back the curtain on how computer vision can make high throughput screening better while keeping humans in the loop. 🧵1/n
1
7
22
you always learn something cool in an anthropic blog post, and this one is no exception
For much more about phase two of Project Vend, read our blog post:
0
0
1
Still searching? Here's the move: USDA Prime steaks delivered to their door before Christmas. No wrapping. No guessing. Just the "where did you find this?" reaction. 6 FREE Petite Ribeyes + FREE shipping ($240 value) Use code SANTA229 📦
0
3
18
very much looking to glp-1s that are either daily or somehow reversible (it's just a peptide, give me a protease i can take to stop it). i want to not snack at work, but giving up actually good food is not a price i want to pay
@JoshConstine They should get ready for it! It’s not that I don’t want to go out and do things on a glp-1, but the food and drink is no longer the main draw. I used to travel across the world just for specific food 😂
0
0
3
what did you think they meant when they said segment ANYTHING?
Introducing SAM Audio: the first unified AI model that allows you to isolate and edit sound from complex audio mixtures. This could mean isolating the guitar in a video of your band, filtering out traffic noises, or removing the sound of a dog barking in your podcast, all with
0
1
3
the greatest world model builder of all time never used benchmarks, he just looked at it and saw that it was good
0
0
6
strongly recommend deadlifting (with proper technique and not trying to 1rm all the time) to help with back pain
@dpolehn If you find out it isn’t, you can’t unring that bell. Nobody ever hurt their back by not deadlifting. There are other ways to build strength and flexibility.
1
0
6
The economic doom loop represents what can happen when a business takes on excessive debt, then struggles with the burden of financing that debt, and finally loses ways to make money. If we think of a city in the same manner, the situation can be explained by high taxation,
0
1
13
apropos of the whole automating science discussion: as an ai person who works very closely with some amazing chemists and biologists i can very confidently say if you think you will automate these people away, you’re ngmi
0
2
8
but will it solve the "you're on mute" and "can everyone see my screen?" benchmarks?
Zoom achieved a new state-of-the-art (SOTA) result on Humanity’s Last Exam (HLE): 48.1% — outperforming other AI models with a 2.3% jump over the previous SOTA. ✨ HLE is one of the most rigorous tests in AI, built to measure real expert-level knowledge and deep reasoning across
0
0
5
this is also why machine learning works (in the sense of having lots of applications) and causal inference really doesn’t - when something doesn’t work we just collect data until it does (we call that scaling!)
0
0
3