Ry

In a few hours I’ll be landing in Seoul for #ICML2026 🇰🇷 !! Alongside colleagues from @cusp_ai , I’ll be around all week (6-11 July) and would love to exchange ideas about: → AI co-scientists navigating chemical space → synthesisability and closing the gap between the screen and the lab → evals, guardrails and verification mechanisms for scientific agents → RL, circuit interp, persistent memory, long-horizon reasoning, and #AI4Science more broadly DM/comment if you want to grab a coffee ☕ and meet some of the team… we’re hiring btw 🙌 - jobs.ashbyhq.com/cuspai/916e148…

Today we’re also confirming our $450M Series B funding, valuing CuspAI at $2.6 billion. I am incredibly proud of what our team has built and the scientific rigor they bring to work every single day.

Today, @CuspAI is launching the ‘AI Materials Foundry’ - a global network of data, labs, compute, and deep scientific expertise dedicated to the design of novel materials.

All time year of sports

The first experimental evidence of recursive self-improvement (RSI). Autoresearching the autoresearch agent for eight days. The result beats the harness we hand-tuned for two years, on held-out benchmarks: 🧵(1/7)

Just came back from ICML. Gave a keynote at the Foundation of Deep Generative Models Workshop, in which I stated that Intelligence should be a scientific subject, arguably much more significant than Physics. It is high time we study it with the same scientific methodology and mathematical rigor as modern physics, instead of always at the level of being empirical, meta physical, meta mathematical, philosophical or speculative... This remains as the biggest opportunity ever for young scientists. To my knowledge, this is not the focus of any of the frontier "AI" companies.

It's mind-blowing how fast agentic coding has progressed in the past 6 month. It's a completely different world now.

@BrookeaJoseph Great work Brooke, and team … this stuff is awesome sauce 🔥

All of the positive posts of people experiencing computer use with 5.6 is really heartwarming 😌

the speed of improvement and innovation in the robotics world is incredibly impressive 🔥🤩

SWE-Bench pro pronounced dead 2 hours before its poster presentation 😔

bro using tshirt as the 2nd screen

@Stefania_druga @SakanaAILabs Ah I’ve seen some familiar faces from Imperial College 💙🤗

My brain is still bustling with ideas after all the amazing conversations at @SakanaAILabs #ICML2026 party. Turns out a group of smart curious people is the best cure for jetlag;) great turnout, kudos to the team!

@sarahba1010 @PhilHackemann @LMU_Muenchen @zuseschoolrelAI @icmlconf Congrats! Fun hearing your oral and stopping by the poster session!

💫 Fantastic news: @PhilHackemann and I won the Outstanding Position Paper Award at #ICML2026. If you are in Seoul and want to learn more about the dual-use potential of AI alignment methods, come join us for our oral presentation tomorrow at 4pm, Hall D2.

Reward-guided diffusion: reinforcement learning steers generative models toward novel, stable crystals Anyone who trains generative models knows the tension. A diffusion model fit by maximum likelihood learns to reproduce its training distribution, so it samples what looks familiar. In materials discovery, though, the interesting compounds sit in the low-probability tails, the underexplored regions where genuinely new structures live. Sampling more of the same is exactly what you do not want. Hyunsoo Park and Aron Walsh reframe this as a post-training problem. Instead of nudging outputs with heuristics like classifier-free guidance, they fine-tune the model with reinforcement learning, optimizing explicit rewards rather than data likelihood. The move that makes it tractable: they run diffusion in a compressed VAE latent space, which turns the awkward mix of discrete atom types and continuous coordinates into a homogeneous action space where policy gradients behave. The engine is GRPO, the same group-relative policy optimization behind recent reasoning LLMs, here driving a denoising diffusion policy. The reward is multi-objective and, crucially, verifiable: a creativity term for uniqueness and novelty, a stability term using a machine-learning force field as a fast proxy for DFT energy above the convex hull, and a diversity term that blocks mode collapse. The payoff is a clean shift of the Pareto frontier. On the Alex-MP-20 benchmark, their metastable-unique-novel score jumps from 15.9% to 61.3%, novelty climbs from 62.3% to 97.5%, and metastability still rises from 51.2% to 72.1%. Novelty and stability usually trade off, and here RL buys both at once. The same framework handles inverse design too: targeting a 3 eV bandgap, RL keeps chemical validity intact while classifier-free guidance collapses to near-zero. For discovery pipelines in batteries, catalysis, or semiconductors, the practical takeaway is that a verifiable reward plus policy optimization can redirect a generative model toward candidates worth synthesizing, not merely plausible ones. Swap the reward and the same machinery targets ionic conductivity, formation energy, or a specific gap, giving a modular path from "generates something reasonable" to "generates what the program actually needs." Paper: Park et al., Nature Machine Intelligence (2026), CC BY 4.0 | doi.org/10.1038/s42256…

Come chat about continuous methods for discrete generation at 2715 hall A 🥰

@katherine1ee @icmlconf @OpenAI awesome! will stop by and come have a chat 🤗 👋🏽

Hellooo! I’ll be at @icmlconf with @OpenAI! If you’re interested in safety research at OAI generally or pretraining safety research I’d love to chat! You can find me hosting a Q&A at our booth Wednesday morning at 9:30am!

@jonkhler hmm 🤔

#icml - a quick lesson in thermodynamics 🤓

In a few hours I’ll be landing in Seoul for #ICML2026 🇰🇷 !! Alongside colleagues from @cusp_ai , I’ll be around all week (6-11 July) and would love to exchange ideas about: → AI co-scientists navigating chemical space → synthesisability and closing the gap between the screen and the lab → evals, guardrails and verification mechanisms for scientific agents → RL, circuit interp, persistent memory, long-horizon reasoning, and #AI4Science more broadly DM/comment if you want to grab a coffee ☕ and meet some of the team… we’re hiring btw 🙌 - jobs.ashbyhq.com/cuspai/916e148…

@felixudr Oh wow looks soo good!

Good vibe in Seoul 🇰🇷