Melissa Du

llms won't solve cancer by themselves wrote about what's missing from the current AI for science discourse and why we need more than just intelligence was an absolute pleasure to work with @swyx on this piece!

@dseeland dm’ing you!

@bearablylight Hi Melissa, I’m a journalist working on a story about jmail.world, we’ve been contacted by a German member of your team who offered to speak to us - could you confirm that he works with you on jmail? Thanks!

llms won't solve cancer by themselves wrote about what's missing from the current AI for science discourse and why we need more than just intelligence was an absolute pleasure to work with @swyx on this piece!

@karinanguyen such a simple and informative setup, super cool work!

Excited to release PostTrainBench v1.0! This benchmark evaluates the ability of frontier AI agents to post-train language models in a simplified setting. We believe this is a first step toward tracking progress in recursive self-improvement 🧵:

@uzpg_ mf has almost 100 likes

when i tell my cofounder he needs to tweet more no one: absolutely no one: my cofounder:

they're all leaving 🤔 monitor-the-situation.com/middle-east

@DKThomp There's commentary in the AI for science space that points out this nuance! latent.space/p/scientist-si… @NOETIK_ai @LilaSciences and other major players understand the nuance and are building with the right thesis. Would love to chat sometime if you'd like a more ML native perspective :)

"AI IS GOING TO CURE CANCER." I've heard this defense, or prediction, from several AI builders and boosters. But I've been frustrated by the lack of specifics. So I asked Eli Lilly CEO Dave Ricks how useful AI is at discovering new drugs. His answer: Today's LLMs are "not particularly good" at the kind of biology necessary to develop therapies. This is not an AI skeptic talking. Lilly has a huge deal with Nvidia to build a biology-focused supercomputer. You don't do that deal if you think the tech is a scam. But as Ricks explains, ChatGPT, Claude, etc work so well because they're trained on the internet and other texts—a universe of human language. But there is no equivalent record of biology. There is no complete internet of human anatomy. And so while AI is promising at answering some diagnostic questions and doing radiology, it's not even close to great at doing the thing that it's boosters claim it will do. The AI discourse is still overpopulated with people living in a future that doesn't exist (AI is thiiiis close to curing cancer) and people living in a present that doesn't exist (AI doesn't have a business model, it's an empty parlor trick that can't "think," etc).

I’ve read a bunch of great blog posts in the last few weeks, and I definitely am not referring to the mainstream viral, AI-written doomer posts sucking the air out of the room. Here’s some smart writing that deserves more love: - The Scientist and the Simulator (@bearablylight for Latent Space) - Heuristics for lab robotics and where its future might go (@owl_posting) - Biology needs to become prospective (@arcadia) - Will Bayesian Statistics Transform Trials? (@kroetscha for Clinical Trials Abundance Blog) Links to my writeup and all pieces in thread🧵

@willdepue will in his just to feel something era

Jwiki just opened up for public contributions. Anyone can sign in, propose edits to articles, and view the full revision history of every change. Edits require admin approval for now. Our goal is Wikipedia-style open editing though, where the articles self-correct. Still figuring out anti-vandalism tools.

@rtwlz ❤️❤️❤️

Wikipedia has different versions for every language, and the same topics don't always use the same pictures. I made a site to show them all walzr.com/in-every-langu…

You can now mark documents for unredaction in Jmail's UI. The DOJ has redacted the names of many perpetrators in the Epstein case alongside the victims. Jmail believes we should call for the unredaction of documents containing incriminating evidence against those who conspired with Epstein or otherwise engaged in illegal activity.

Jemini today is the worst that it'll ever be.

epstein co-conspirator scoreboard: 1 conviction 2 deaths ...

Worked with @TheEconomist to analyze 1.4 million emails from the latest Epstein file drops

@BillCrosby @swyx post has more details but tldr; domain-specific models (simulators) and better data infrastructure! language models are great at reasoning and less great at understanding / predicting the physical world