Only in the Mission π²π½ @HackIterate brought a mariachi band to close out the AI Research Hackathon. Agents, guitars, and culture. cc @Kristopherfloyd
Only in the Mission π²π½ @HackIterate brought a mariachi band to close out the AI Research Hackathon. Agents, guitars, and culture. cc @Kristopherfloyd
We just hit 300 million npm downloads. This is the culmination of a decade of open source work.
301.6M total Β· 58.9M/month Β· 16.1M/week Β· Cloud: 201.1M
Last 30 days ran 14.7x a year ago. Cloud is up 27.4x. 71% of downloads came in the last 12 months.
@constructive_io
Why do AI image generators sometimes copy their training data?
At FRC #24, Anubhav Jain (Applied Researcher at Apple, NYU PhD) walked through his CVPR 2025 paper with collaborators at NYU and Sony AI.
Diffusion models sometimes reproduce training images almost exactly, and deduplication helps but doesn't eliminate it. Anubhav traces memorization to an "attraction basin" early in denoising, where classifier-free guidance pulls every trajectory toward the same memorized image.
The fix happens at inference time:
β’ Withhold guidance until the trajectory leaves the basin, and the copy goes away
β’ "Opposite guidance" escapes the basin sooner
Similarity to training data drops while FID and CLIP stay strong, with no retraining and no weight changes. The honest caveats: it trades off with image quality, and the experiments ran in 2024 on Stable Diffusion.
Thank you to @SignalFire for hosting #24.
Tonight we're back at SignalFire for FRC #30: Worlds for Agents, with @bonniesjli and @jiaxin_ge_. Doors 5:30 PM.
Request a seat π
At FRC in May, @bonniesjli (now at @OpenAI) showed that RL training hits a ceiling, and the recipe sets it. You can't train longer to fix a bad one.
This goes around training entirely. A frontier model edited 0.016% of a small model's weights directly, with no gradients, fixed its reasoning loops and made its math up to 6x better.
If that holds up, the next way past the ceiling may not be a better recipe. It may be a model that reads and rewrites another model's weights.
soon ai agents will handle most business negotiations. currently, the problem is that an ai can be talked into a worse deal if the other side is friendly π
@ericavaneee built a test that catches this.
Before TERMS-Bench, @ericavaneee worked on Amazon's negotiation agent as an applied scientist. That's where the problem came from: math and code have answer keys, but a negotiation doesn't. The deal is objective, the back-and-forth that produced it isn't, and most benchmarks just have another LLM grade it.
At FRC #29 at @arrayvc, she presented TERMS-Bench from @StanfordEng, which makes the environment itself the verifier. Across 13 models, almost every one closed the deal, but the value they kept varied 3.7x. Friendly messages from the other side made them give up more, not less.
Thanks to @gmi_cloud for the support!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! π§΅
Blog: https://t.co/bOWCcFhCVq
This is INSANE. Ciscoβs president Jeetu Patel has given a 27-year-old engineer unlimited AI tokens to try to rewrite 98% of the companyβs legacy code within 24 months, Fortune reports.
Padmanabhan Krishnamurthy thinks he can do it with a two-person team. He has just four weeks to prove the approach.
much of the robotics research I see runs on @trossenrobotics hardware. great to see them pushing for a stronger American robotics industry with @robotforamerica! π¦ΎπΊπΈ
Our robots are American-made, designed, machined, and assembled in-house, and every product comes with lifetime support from the people who built it. Weβre proud to be a member of @robotforamerica, working toward a stronger U.S. robotics industry.
https://t.co/gPZp0tDtsX
FRC #28 at @StanfordMed on Oct 21: two talks on what breaks inside a cell, and what it takes to model it.
Ganesh Pusapati of the @RohatgiLab showed that the molecular machine broken in more than 80% of colorectal cancers isn't dead, just broken. His team's @NatureGenet screen of single-letter DNA edits found one that strengthened a key contact, partly repaired the machine, and slowed colorectal cancer cells in the lab. The next idea is a drug that holds the machine together instead of blocking the growth signal.
@drAOPisco of @ChanZuckerberg argued the virtual cell is a data problem before it's a model problem. Tabula Sapiens 2.0 covers nearly 1.2 million cells from 28 tissues, and that's no longer big enough for the people building models. The next wave has to come from experiments that perturb cells and measure how they respond.
Thanks to @AmplifyPartners and @dhaliwas for dinner and the welcome, @gmi_cloud for compute, and Aaran Vijayakumaran for leading the discussion.
Great to see everyone in the room: @sidpreetham@ananthp_@probell@ChoYehlin@juansearch@jefferey_cai@BroderSchmidt@yalda2009@timothylrn@irishikeshmagar@lisali126@C_Hackenbroch@Maxi_nglrt@AneequaSundus@SanilBaweja@shoobiedoo313@efrazar@ankits0052
Both full talks π
big opportunity area in ai x bio. every cell has a built-in brake that tells it when to stop growing. in most colorectal cancers, one mutation breaks the brake.
most drug discovery models are built to find a molecule that plugs a pocket on a protein and switches it off. fixing a brake means finding a molecule that strengthens a weakened connection between proteins so they work again. that's a much harder design problem, with far less data behind it
and that's where the opening is
At FRC #28 at @StanfordMed, Ganesh Pusapati of the @RohatgiLab asked a different question about colorectal cancer: what if you repair the machine it breaks instead of blocking the signal?
His team's @NatureGenet screen of single-letter DNA edits found one that partly fixed it and slowed cancer cells in the lab. That points to a new kind of drug, one that holds the broken machine together.
Thanks to @AmplifyPartners and @gmi_cloud for sponsoring. Full talk π
@gerta_ma and @alexhapki have helped people add six and seven figures to their offers. tonight they're sharing their playbook with our community for free.
if you're interviewing, weighing an offer, or up for a promotion, come with questions
Your research creates value. Tonight, learn how to negotiate for it.
Free live workshop for the Frontier Research Club community with @gerta_ma and @alexhapki, cofounders of YourNegotiations. They'll cover negotiating the full package (comp, title, scope, resources), protecting your leverage before an offer arrives, and live Q&A on your own situation.
Thank you Gerta and Alex for building this for our community.
Tonight, 5:30 to 6:30 PM PT, live on Zoom. Live only, no recording.
https://t.co/XBxe0pZ9Bd
most of the virtual cell conversation is about models. angela's point is that the bottleneck is the data, and the data we need most comes from experiments.
if you're building in ai x bio, that's where the moat is.
Everyone is racing to build a virtual cell: an AI model that predicts what a drug or a gene knockout does before anyone runs the experiment.
At FRC #28 at @StanfordMed, @drAOPisco, Director of Data Science at @ChanZuckerberg, argued it's a data problem before it's a model problem. A model only learns what its data shows, and most cell data today is snapshots. The next wave has to come from experiments that perturb cells and watch what happens.
Thanks to our FRC #28 sponsors @AmplifyPartners and @gmi_cloud. Full talk π