@dieworkwear Had instant association with South2west8, although after gorpcore boom outdoorsy photoshoots were everywhere. This year also had some great items (love their bags)
@RimShayakhmetov At least one manifesto about AI that is good & not a 150 page read about the future dominance of AI-power-users-tech-bros master race. Although some wordings are so vague, they remind me of the mentioned Agile manifesto (which basically led to the overcomplicated development)
HUGE!🚀 We just released v2 of our ADMET specialist language model, trained in MMAI Gym. 4B, without any tools, and strict SOTA on 17 (!) of 28 tasks: lower MAE than all four strong ADMET specialist baselines (and better than general purpose LLMs, too, of course).
The point is the transformation and that you can actually train a language model for such previously thought specialist/tool-only tasks. MMAI Gym turned a pure language model into a specialist molecular-property predictor: the model itself learned the prediction, no external tools needed to get SOTA results.
The 17 wins span selected CYP inhibition, plasma protein binding, microsomal clearance, solubility, permeability, transporter inhibition and safety assays, volume of distribution, etc. In drug discovery terms: how a molecule is metabolized, how much free drug is left to act, and early safety flags, the calls that help decide which molecules move forward. If one count tasks where it is better or within 5% of the best baseline (SOTA-level), then this new model update is SOTA-level on 21 of 28.
The four strong specialist baselines: Chemeleon, MapLight, MapLight+GNN and MiniMol, each retrained on the same MMAI Gym ground-truth data and splits. Our model is InsilicoMMAI-4B-Chem-ADMET-2.0.
And where do general-purpose LLMs stand? Way behind any specialist model without tools/cheating. Our model has lower MAE than any model.
📊 DDDBench: https://t.co/ytR5OTcx6V
▶️ Try the new model for free now: https://t.co/KfOugvVF3c
SLM: the S keeps standing for specialist and small model at the same time💪
#insilicoSOTA
Every parent wants a happy child. So we built MDD-PGT, the first genetic predictor of depression for embryo screening.
Kids with a lower MDD-PGT tend to sleep better, fight less and enjoy jokes more. With 10 embryos, picking the lowest-risk one cuts depression risk by over 40%.
Orchestration is the low-hanging fruit in AI drug discovery. Right now, everyone is overobsessed with it.
OpenAI recently acquired Orchestra Bio for program coordination. Anthropic Mithril is the same pattern: one operator, agents running contracts, labs, logistics.
Useful - yes. But it's not the hard problem nor the final goal.
Compressing drug discovery and development timeline is, including success rates.
We designed potent DDR1 inhibitors in 21 days and had mouse PK in 46 back in the good old 2019 (Nature Biotechnology). It was the same field, years earlier chatGPT or COVID for that matter, with different obsession on using AI to discover drugs and humans operating around it at an excellent speed.
Building the operations layer with agents is a must, no doubt in that. But don't forget about the main goal of compressing the clock (longevity: pun intended) and having pharmaceutical superintelligence to really move the needle in clinical outcomes.
How do you make a foundation model fluent in drug discovery? 🧬💊
Excited to present our paper at #EMNLP2026:
“MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery”
400+ tasks. One goal: teach Liquid AI models the language of drug discovery.
Come chat!
@MelindaBChu1@RimShayakhmetov I don't, I work in Insilico Medicine but I don't see why bother about incentives, while the message is ok.
But I agree with Rim about the nuances, and that basically our guys did this all in 2019. And, TBH, all the pharma uses ML approaches, mostly unbeaten by LLMs, for decades.
@MelindaBChu1@RimShayakhmetov Well, why wouldn't they, some investment analytics firm (ok, with the help of specialists) did thing that for most people is almost equal to nobel-prize-winning breakthrough.
And they actually mentioned bottlenecks/reality in their primary report.
Instant SAR → pharmacophores MCP 🤯
Ask your agent, get an interpretable model: what drives potency, and what your series needs just to bind.
Our brand new SAR Pharmacophore MCP. No training, no GPU.
Coming soon to our MCP Hub.
#AgenticPharma
1.5 years ago, we bet the company on benchmarking and evaluating all things AI for science.
Today, we’re #1 in many areas. One of them is AI for longevity. Just ask your favorite LLM.
Here’s what GPT Astra says about our Cell cover paper 👇
Link in the comments
Our Special Issue on AI in Biology is out 👉https://t.co/7FRErzTs9T
The cover introduces LongevityBench, an open benchmark comprising 17 tasks across five biodata domains to evaluate AI systems in the context of human aging.
Aging has a language. 🧬 We’re teaching AI to read it.
On the cover of Cell: compact Longevity-LLMs, jointly trained by Insilico Medicine and @liquidai, demonstrate SOTA performance on LongevityBench.
Powered by these models, Longevity Claw nominated 328 candidate gene targets for aging research.
Small model. Big questions. Open source. 🧵👇
#insilicoSOTAFM