1/ Medical students don't become doctors by memorizing facts and passing exams. They do by residency, years of practice across thousands of patient encounters. We wanted to see if we could train models the same way. Introducing ResidencyRL 🧵

Most AI-for-science systems stop at hypothesis generation. We extend Co-Scientist to design experiments, write code, and interface with physical lab hardware – then validate it across materials science, biology, and computer science. Fully unsupervised discovery isn't here yet, but execution-grounded human-AI collaboration is already moving the needle in real labs. Proud to be part of this!

A huge win for science, AI, and humanity Google’s Gemini Co-Scientist generated hypotheses, designed experiments, helped synthesize new materials in physical labs, predicted biological behavior, and autonomously discovered a medical AI architecture that beat multiple frontier models. Closed loop AI science is here

for all the gemini haters out there, you darn degenerates. name anything in the universe that’s a bigger driver of scientific progress than sir demis and his marauding minions.

Exciting new work from our team @GoogleDeepMind! Co-Scientist autonomously designed a safe precursor route for 2D MXene synthesis, a pipeline predicting E. coli swarming morphology at unseen inducer concentrations, and an agent architecture beating six frontier models on length-adjusted HealthBench. Validated against real experiments, and with real caveats worth reading.

Today, we're excited to share early progress on using Gemini to accelerate scientific discovery in the real-world. We present an extension of Co-Scientist which we use to collaborate with scientists across materials science, biology, and computer science.

I am excited to share some of the progress we are making towards using Gemini to accelerate scientific discovery in the real-world. We present an extension of Co-Scientist that transitions from pure in silico ideation to execution-grounded research partner, adapting to multiple levels of autonomy across different scientific domains.

Advancing science is AI's ultimate purpose: not beating benchmarks, but making progress that translates into the physical world and addresses humanity's biggest challenges: accelerating the transition away from fossil energy with new materials for batteries or controlled nuclear fusion, finding new cancer cures, and pushing semiconductor tech further. To get there, AI needs to move from computation into the physical world. This work demonstrates early progress – among other experiments: Gemini designing materials synthesis protocols and interfacing with lab hardware.

My thesis for Google graduation ;) AI coscientist needs to be real and useful. With AI + real lab e2e collaboration, we made a significant progress in 2D material synthesis -> promising candidates to break current limit and build way more powerful chips and batteries

Small Models Have Arrived calv.info/small-models-have-…

I'm hiring! 🚀 Our Frontier Health team at @GoogleDeepMind is looking for a Research Scientist to build foundational AI & physiological world models for human biology. Help us transform healthcare from reactive observation to proactive intervention. 🧬 🔗 Apply: goo.gle/3SPNis2

We created AI Sleep Co-Scientist to analyze 1M hours of rich, multimodal sleep recordings. It discovered many new insights into human physiology: 🧠weaker brain–heart coupling predicts neurodegenerative diseases 🕑sleep "age" associated w/ health risks across multiple organs 😴distinct arousal characterizes comorbid insomnia/apnoea + more! arxiv.org/abs/2607.25175 Awesome collaboration w/ Emmanuel Mignot’s group, led by @connect_thapa!

We are making continuous improvement on Gemini for Biology!

Thrilled to share our new preprint on AMIE, demonstrating expert-level performance compared to primary care physicians in real-time clinical video consultations!

3/ In a blinded side-by-side eval by board-certified clinicians, our trained model was preferred in: → 87.6% of cases for overall clinical impression → 90.7% for completeness of information gathering → 75.3% for management plan quality Missed red flags dropped by a third.

Very happy (and relieved) to see our work on multimodal conversational medical AI accepted in @NatureMedicine nature.com/articles/s41591-0… In the published version, we have substantially expanded on the analysis and evaluation. Kudos to @_cjpark @timstro @JanFreyberg @_khaledsaab This work also formed an important precusor for our more recent work where we explored a similar problem but in real-time interaction: nitter.net/RyutaroTanno/status/20… Both modes of UX (synchronous and asynchronous) are useful but in different ways. Also a nice reminder that a prospective evaluation remains as an important future work.