ScienceBuddy is a free, browser-based workspace that pulls papers, data tables, biological sequences, and scientific images into a single environment for biomedical researchers. The agent inside ships with 224 tools across 22 modules spanning genomics, molecular and cancer biology, pharmacology, bioimaging, literature retrieval, and database queries, and every tool call, generated artifact, and compute step is kept visible in dedicated Compute, Results, and Trajectory panels so researchers can audit the work behind each answer. Access to GPT-6 is included at no cost during the preview, with GPU acceleration and PhAI Labs' JEV framework handling execution.
The launch matters because ScienceBuddy is built to learn from the researchers who use it. An inner loop watches sessions for failures and rewrites the agent's harness (its instructions, skills, and context) while the task model stays fixed, keeping only candidates that win paired evaluations. An outer loop then freezes the harness and retrains the model with reinforcement learning against rubric-based rewards derived from real researcher requests and corrections. PhAI Labs calls this recursive-in-recursive self-improvement, with updates deployed between cycles while the service remains live.
PhAI Labs is based in Palo Alto and was co-founded by Yingcheng Wu (CEO), Zhenfei Yin (president), and Ling Yang (chief scientific officer), with Muchen AI acting as a joint R&D and technical support partner on this release. Preview access is open at science-buddy.io in English and Chinese, with a technical report on arXiv and an MIT-licensed GitHub repository covering the self-improvement experiments and reproduction steps.
Comments (11)
Sign in to join the discussion.
N
Nadia Oyelaran9d ago
GPT-6 for free is either the most generous thing I've seen this year or a very creative way to describe your own fine-tune. Which one is it?
Y
Yuki Tanabe9d ago
Inner loop, outer loop, I'm one more loop away from this being a washing machine manual. The JEV framework sounds cool though, where do I read actual docs?
I
Ingrid Halvorsen9d ago
Staff eng question: is the demo pre-recorded or is there a live endpoint I can hammer? Agent harnesses that refine themselves tend to refine themselves into timeouts.
P
Pritam R.9d ago
ok wait, a free GPU-accelerated workspace for scientific agents? what's the catch, do you harvest my protein folds in exchange
M
Marcos Delgado9d ago
Tagline rewrite, no charge: 'ScienceBuddy: your lab's second brain, running on borrowed GPUs.' You can venmo me later.
Z
zephyr9d ago
The launch video has that calm Dieter Rams energy until the particle effects show up at 0:34 and ruin the vibe. Someone please protect your editor from After Effects presets.
L
Ling Xin9d ago
The thread cuts off right at the outer loop reveal and I refuse to click through. That's either genius pacing or your tweet composer betrayed you.
A
Amara Nwosu9d ago
Intern question that's probably dumb: if the inner loop refines the harness while I collaborate, what happens when I close the tab mid-refinement? Does my buddy get amnesia?
K
Kiran Shetty9d ago
The pool of scientists who want yet another agent workspace is shrinking fast. Everyone I know in a wet lab just wants their PCR to not fail.
S
Sofia Marchetti9d ago
Been building something adjacent for 7 months, 3 weeks out from launch (as always). Genuinely curious how you're handling tool-call grounding on messy experimental data.
T
Tomás Berger9d ago
Would love to see how JEV compares to the agent harness work out of the Allen Institute folks. Any whitepaper or is this vibes-driven research for now?
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
PhAI Labs
Building AI for real scientific discovery. Discovery Foundation Models, scientific agents & infrastructure.
Building AI for real scientific discovery. Discovery Foundation Models, scientific agents & infrastructure.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
ScienceBuddy is a free, browser-based workspace that pulls papers, data tables, biological sequences, and scientific images into a single environment for biomedical researchers. The agent inside ships with 224 tools across 22 modules spanning genomics, molecular and cancer biology, pharmacology, bioimaging, literature retrieval, and database queries, and every tool call, generated artifact, and compute step is kept visible in dedicated Compute, Results, and Trajectory panels so researchers can audit the work behind each answer. Access to GPT-6 is included at no cost during the preview, with GPU acceleration and PhAI Labs' JEV framework handling execution. The launch matters because ScienceBuddy is built to learn from the researchers who use it. An inner loop watches sessions for failures and rewrites the agent's harness (its instructions, skills, and context) while the task model stays fixed, keeping only candidates that win paired evaluations. An outer loop then freezes the harness and retrains the model with reinforcement learning against rubric-based rewards derived from real researcher requests and corrections. PhAI Labs calls this recursive-in-recursive self-improvement, with updates deployed between cycles while the service remains live. PhAI Labs is based in Palo Alto and was co-founded by Yingcheng Wu (CEO), Zhenfei Yin (president), and Ling Yang (chief scientific officer), with Muchen AI acting as a joint R&D and technical support partner on this release. Preview access is open at science-buddy.io in English and Chinese, with a technical report on arXiv and an MIT-licensed GitHub repository covering the self-improvement experiments and reproduction steps.
Comments (11)
GPT-6 for free is either the most generous thing I've seen this year or a very creative way to describe your own fine-tune. Which one is it?
Inner loop, outer loop, I'm one more loop away from this being a washing machine manual. The JEV framework sounds cool though, where do I read actual docs?
Staff eng question: is the demo pre-recorded or is there a live endpoint I can hammer? Agent harnesses that refine themselves tend to refine themselves into timeouts.
ok wait, a free GPU-accelerated workspace for scientific agents? what's the catch, do you harvest my protein folds in exchange
Tagline rewrite, no charge: 'ScienceBuddy: your lab's second brain, running on borrowed GPUs.' You can venmo me later.
The launch video has that calm Dieter Rams energy until the particle effects show up at 0:34 and ruin the vibe. Someone please protect your editor from After Effects presets.
The thread cuts off right at the outer loop reveal and I refuse to click through. That's either genius pacing or your tweet composer betrayed you.
Intern question that's probably dumb: if the inner loop refines the harness while I collaborate, what happens when I close the tab mid-refinement? Does my buddy get amnesia?
The pool of scientists who want yet another agent workspace is shrinking fast. Everyone I know in a wet lab just wants their PCR to not fail.
Been building something adjacent for 7 months, 3 weeks out from launch (as always). Genuinely curious how you're handling tool-call grounding on messy experimental data.
Would love to see how JEV compares to the agent harness work out of the Allen Institute folks. Any whitepaper or is this vibes-driven research for now?