BioTech - AI - Host of Latent Space AI for Science podcast
Mostly SF Bay Area
Joined April 2009
- Tweets205
- Following257
- Followers173
- Likes111
RJ Honicky retweeted
🧪🧭 What if we could ask a model which experiment gives us the cells we want?
Our new preprint, LabCompass, takes a step in this direction through generative inverse design. For me, this is one of the most exciting directions for the field.
Together with many others, we have been working on predicting how cells respond to perturbations. How can we use these models to design experiments that achieve a desired biological outcome?
🎛️ This inverse problem is difficult. Combinations and doses create spaces that cannot be exhaustively screened, and each experiment produces a distribution of cell states. Predictive models are also imperfect, particularly in unexplored regions where we want to discover something new.
LabCompass couples a generative model of experimental designs with a forward model of cellular responses. The desired phenotype guides generation in continuous design space, while a learned prior favors plausible designs. Selected candidates are tested experimentally, and the measurements inform the next round.
💡 The method is not tied to a particular virtual cell model. It can work with different differentiable forward models and cellular readouts, and the same trained models can address different phenotypic objectives.
🩸 We demonstrated this in primary human blood stem and progenitor cells. Across the study, we profiled 78 million cells and experimentally validated conditions enriching megakaryocytic, erythroid, monocytic and early lymphoid populations.
One result I particularly like: distinct culture recipes produced similar target enrichments. Several solutions to the same biological objective give us alternatives to investigate and develop further.
🔬 The opportunity extends beyond blood and cell culture. We also applied LabCompass to single-cell transcriptomic data from viral infection. I’m excited to explore its potential for other perturbation systems, drug combinations, differentiation and organoids.
This raises a question for virtual cell models: alongside prediction accuracy, how effectively can they guide experiments toward desired biological outcomes?
If this connects to a problem you are working on, I’d love to hear from you. We’re keen to see how LabCompass transfers to new systems and happy to explore ideas together.
A wonderful collaboration with Bertie Göttgens and Nicola Wilson at the Cambridge Stem Cell Institute. Huge credit to joint first authors Licyel Paulas-Condori, Alessandro Palma, Lorenzo Consoli and Juan Antonio Rubio-Lara, and the whole team across both labs!
📄 Preprint: biorxiv.org/content/10.64898…
💻 Code: github.com/theislab/LabCompa…
#InverseDesign #VirtualCell #SingleCell
RJ Honicky retweeted
Was fun to be on the @latentspacepod podcast a few weeks ago to talk about AI for physical simulation and understanding. This pre-dates the public launch of @accelerated_u so I couldn't yet talk the exact details but hint at what happens if you scale some of these methods to Trillion parameter model sizes and fully 4D context lengths in the Trillions in a universal model.
The podcast covers some of the ideas that provide the foundation for scaling. It also goes into the promise and successes that were possible even before going really big like building high resolution fully AI based weather models that are tens of thousands of times faster and as accurate as existing forecasts. We talk about predicting plasma behavior in fusion so quickly that one could take corrective action before something bad happens. And we cover how physical understanding doesn't just help with replacing experiments but lets us optimize design directly.
We even briefly touch on the promise of combining those capabilities into one large model which I can now talk about more.
Watch for yourself: youtu.be/79mIutht1f4?si=M3Wd…
RT @realdealpatil: Was fun to tag team @latentspacepod with @mattmcpartlon1! My favorite part was Matt explaining why mice make for such fo…
This quoted post is unavailable.
RJ Honicky retweeted
somehow without me noticing, @latentspacepod has quietly overtaken some of my own podcast heroes like @ProfCalNewport, @TBPN, @gruber, and other AI pod friends! wow!
we are seriously ramping up LS as a new tech media org, thanks to subscribers and donors. @ricmac joins us as our first fulltime Head of Editorial to build out technical coverage of AI news; we will also be FINALLY opening up for sponsorships.
we are hiring writers and show producers. reply/reach out if you are interested!
Today's my first day in a new role: Head of Editorial at Latent Space (@latentspacepod).
From ReadWriteWeb to The New Stack, I've always loved being at the cutting edge of tech x media. So I'm thrilled to join @swyx & his team. We have big plans! latent.space/
RJ Honicky retweeted
AI agents are only as useful as the foundation models they orchestrate.
The existing protein-ligand structure prediction models that performed well on benchmarks weren’t built to handle the challenges of real drug programs, so we built Pearl from the ground up. Pearl cleared a new bar for molecular AI. Sapphire is what we’ve built on top of it.
@edunov and I joined Brandon and @honicky on @latentspacepod to discuss both, including the first public details on Sapphire. Listen below.
Agents 🤝 GEMS. For the first time, we’re sharing details on Sapphire 💎 our agentic drug design system built on top of our SOTA molecular foundation models.
Sapphire works 24/7, autonomously iterating on drug candidates – reasoning about 2D chemical structures and 3D molecular poses, and forming and testing hypotheses with real intelligence along the way. It represents a view we've held from the start: agents become meaningful for drug design only once the underlying molecular foundation models are capable enough to perform on real drug programs. We built those models first.
Our CEO @enfeinberg and CTO @edunov cover both topics on @latentspacepod, including what separates models that hold up in reality from those that only look good on benchmarks.
RJ Honicky retweeted
Create materials that don’t exist yet.
Millions of data points aren’t required to make real discoveries.
Our recent campaign discovered 300 novel compositions in 16 weeks.
RJ Honicky retweeted
Remove the bottlenecks to invention.
“Materials discovery has been stuck on a 20–30 year timeline for generations. Joseph thinks that’s about to change, and Radical AI is putting that thesis to the test in the lab, one sample at a time,” reports Brandon Anderson.
Thank you @latentspacepod for the discussion on how self-driving labs will invent materials for the next wave of technology.
RJ Honicky retweeted
A spicy podcast.
In the movie The Man Who Knew Infinity, Ramanujan became a much stronger mathematician after the Hardy made him write proofs.
Verification is for better generation, for optimal design.
Verified AI accelerates superintelligence, esp agentic superintelligence.
This past March, @CarinaLHong sat down with @latentspacepod for an episode on the Series A announcement.
"Verified AI sounds like eating broccoli and paying taxes, but to Axiom it's about scaling brilliance, compounding brilliance."
Agentic superintelligence needs Verified AI.
RJ Honicky retweeted
Our paper on ESMC, ESMFold2, and mechanistic interpretability for proteins is up on @biorxivpreprint!
We've made a few changes since the initial version went online last week.
1. We found an issue in the way we provided MSAs to OpenFold3. This led us to report lower performance of OpenFold3 on some benchmarks. This issue does not affect any of the other models evaluated.
2. We updated how we report results on Runs N' Poses to more closely match the original paper (counting only ligands with valid SuCOS similarity score). We also add a bar plot to the supplement that stratifies performance by similarity. This mostly changes the absolute values of the pass rate, not the relative performance of models.
3. Added some more BLI data to the supplement.
4. Added some missing citations, fixed typos, etc.
Check out the preprint here: biorxiv.org/content/10.64898…
RJ Honicky retweeted
🔬Doing Vibe Physics
The full story of how GPT‑5.x derived new results in theoretical physics and quantum gravity, live on our Science pod today!
latent.space/p/lupsasca
our conversation with @ALupsasca, an award winning theoretical physicist on his AGI-pilling journey applying GPT5 to physics problems (with a nudge from @markchen90)!
Timestamps
0:00 Introduction to Al's impact on physics research
0:43 Guest introduction: Alex Luposka
2:49 Alex joining OpenAl and the shift in physics research
4:08 The release of GPT-5 and the shift in capabilities
10:05 Explaining Quantum Field Theory and amplitude calculations
14:20 Overview of gluons and the strong force
14:38 Discussing the first research paper on single-minus gluon tree amplitudes
20:56 How ChatGPT helped solve a year-long physics puzzle
23:02 Complexity of manual calculations in physics
26:12 The history and mechanics of Feynman diagrams
27:44 The Parke-Taylor formula and the quest for simplification
31:26 Using ChatGPT to find the simplification in the special phase space region
38:07 Proving the formula from scratch to ensure validity
41:00 Determining the scientific impact and future research
42:27 Introduction to the second paper on graviton amplitudes
45:41 |
Defining particles, irreducible representations, and symmetry
47:46 How GPT Pro generalized the research to gravity
53:57 The epistemological shift: Is this a new way of doing physics?
59:27 The use of Al as a 'scout' for research directions
1:01:44 The role of 'taste' and collaboration with Al
1:10:23 Personal evolution from Al skeptic to resident scientist
1:12:46 Solving a black hole perturbation problem with GPT-5
1:16:34 Discussing whether Al can make original, conceptual leaps
1:20:09 Challenges of 'Al slop' and the future of academic publishing
1:23:13 The bottleneck of writing academic papers
1:30:19 Final takeaways and looking ahead to the next year
GPT-5.2 derived a new result in theoretical physics.
We’re releasing the result in a preprint with researchers from @the_IAS, @VanderbiltU, @Cambridge_Uni, and @Harvard. It shows that a gluon interaction many physicists expected would not occur can arise under specific conditions.
openai.com/index/new-result-…
RJ Honicky retweeted
🔬 Training Transformers to solve 95% failure rate of Cancer Trials
the AI for Science pod is back with @RonAlfa, CEO of @NOETIK_ai, and Daniel Bear, VP Research at Noetik, explaining exactly how their team of top AI x Bio researchers and engineers (shoutout @owl_posting) will use AI to cure cancer, by focusing on key bottlenecks like patient selection, and training large cancer foundation models like TARIO-2, an autoregressive transformer trained on one of the largest sets of tumor spatial transcriptomics datasets in the world... which first required years of blind faith in collecting good data to even get going:
RJ Honicky retweeted
🔬Why There Is No "AlphaFold for Materials"
latent.space/p/materials
Materials Science is a force for good everywhere in our lives, from your clothes to the computers you use.
We catch up on AI for Materials Discovery with Prof. Heather Kulik of @KulikGroup, one of the first materials scientists to realize that there was alpha in combining computational tools with data driven modeling... and we test out some predictions she makes on the pod with Opus 4.6 and GPT 5.4!
RJ Honicky retweeted
Your next cancer drug or gene therapy spends years stuck in a lab before it reaches a pharmacy shelf. Anthropic is testing a tool that goes after the bottleneck: the coding work biologists do before any human trial even starts.
It’s called Operon. Four tasks show up on the loading screen and they tell you everything.
“Design a CRISPR knockout screen” means turning off every gene in a cell, one at a time, across 20,000+ genes, to find which ones cause disease. Imagine flipping 20,000 light switches to figure out which one controls the kitchen. Planning that experiment alone takes a biologist months.
“Analyze scRNA-seq data” means reading what each individual cell in a tissue sample is doing, instead of blending them all together into one average. The code to crunch that data takes weeks to write.
“Rank enzyme variants with PLMs” means feeding an AI trained on 65 million protein sequences and asking: which version of this enzyme will actually work? It predicts the answer before anyone picks up a pipette.
“Build a phylogenetic tree” means mapping how organisms or genes branched apart over evolution. How scientists trace where a virus strain came from.
Every one of these jobs requires a PhD and serious programming chops. Operon wants to turn them into a conversation.
The name itself is a nod. An operon is a cluster of genes that switch on together in biology. Nobody outside a biology department would pick that name.
Operon sits inside Claude Desktop as its own mode, next to Chat, Code, and Cowork. It reads files straight off a researcher’s computer (biology datasets are way too big to upload). It has a planning mode and an auto mode, borrowed from Claude Code.
Anthropic has been laying track here for months. Last October they launched Claude for Life Sciences with plug-ins for PubMed (35 million medical papers), Benchling (where scientists log experiments), and 10x Genomics. In January they added a drug compound database. Claude already beats human experts on a lab protocol comprehension test, 0.83 versus 0.79.
And the real-world pharma results are hard to ignore. Novo Nordisk used Claude to chop clinical report writing from 15 weeks to under 10 minutes. A team of 50 writers shrank to 3. Their annual Claude bill costs less than one writer’s salary. Every day a drug hits the market sooner is worth roughly $15 million to them.
The AI drug discovery market sits around $3 to 5 billion this year, growing 20 to 30% annually. Over 200 AI-discovered drug candidates are in clinical trials right now. But I want to be straight about what AI can’t touch: clinical trials still grind on for years. Regulatory review, manufacturing, all of that stays slow. AI trims the front of a decade-long pipeline.
Anthropic hit $19 billion in annual run rate last month. $380 billion valuation. 80% of that revenue comes from businesses. Operon says a lot about where they think the next enterprise dollar is hiding. The lab.
BREAKING 🚨: Anthropic is working on a new Operon agent for Claude Desktop, built for scientific research in biology!
Operon will have a "private environment" to work alongside you. Users will be able to create different sessions within Operon projects, manage generated artefacts, and work with Skills.
Cowork but for scientists 👀
RJ Honicky retweeted
"We have a lot of interaction between the TPU chip design architecture team and the sort of higher level modeling experts because we really wanna take advantage of being able to co-design 'What should future TPUs look like based on where we think the sort of ML research puck is going?' in some sense.
Because as a hardware designer for ML in particular, you're trying to design a chip starting today, and that design might take two years before it even lands in a data center, and then it has to sort of be a reasonable lifetime of the chip to take you three, four, or five years."
Jeff talks more about the process of co designing TPUs and how they think through building for the future.
@JeffDean
From rewriting Google’s search stack in the early 2000s to reviving sparse trillion-parameter models and co-designing TPUs with frontier ML research, Jeff Dean has quietly shaped nearly every layer of the modern AI stack. As Chief AI Scientist at Google and a driving force behind Gemini, Jeff has lived through multiple scaling revolutions from CPUs and sharded indices to multimodal models that reason across text, video, and code.
We sat down with Jeff to unpack what it really means to “own the Pareto frontier,” why distillation is the quiet force behind every generation of faster, cheaper models, how energy not FLOPs is becoming the true constraint on AI compute, what it takes to co-design hardware and models 2–6 years into the future, why unified multimodal systems will outperform specialized ones, what it was like leading the charge to unify all of Google’s AI teams, and his prediction that deeply personalized models with access to your full digital context will redefine what useful AI looks like.
@JeffDean @GoogleDeepMind @Google
RJ Honicky retweeted
Today we introduced AlphaGenome, a new tool that can more comprehensively predict the impact of single variants or mutations in DNA 🧬
How, you ask? 🤔
tldr; Our AlphaGenome model takes a long DNA sequence as input, processes that data, and predicts thousands of molecular properties by characterizing its regulatory activity.
For the full read ➡️ deepmind.google/discover/blo…
RJ Honicky retweeted
“2026 in AI for Science is going to look a lot like 2025 for Software” - @kevinweil
An exclusive first look at Prism, @OpenAI for Science’s first major product launch, telling the story of how Kevin headhunted @vicapow from obscure reddit threads, and first demo of Prism! We dig into:
- how Prism uses GPT-5.2 to propose clean diffs, proofread and restructure sections
- turn whiteboard photos into TikZ diagrams
- run parallel reasoning chats to verify math without touching the manuscript
- what “AI in the workflow” really unlocks: closing the impedance mismatch from thought to LaTex (there’s a startup buried in every million copy-pastes into ChatGPT)
- why AI for Science is poised for explosive takeoff in 2026.
We are big believers too! Look out for the first ever AI for Science pod coming to a Latent Space near you.
RJ Honicky retweeted
we are going thru OpenAI’s sCM paper on the LS Paper Club today with @honicky!!
(search our youtube or discord to join)
[24 Oct 2024]
Let me see your image models 1-2 step!
buttondown.com/ainews/archiv…
Congrats to @clu_cheng and @DrYangSong
RJ Honicky retweeted
thanks to @honicky and @vibhuuuus for another great Paper Club on ReFT: Representation Fine Tuning by @aryaman2020 et al!
signups below
