Inherent, founded by three former DeepMind researchers, comes out of stealth with a $50M seed co-led by Index and Radical
What happened: Tantum Collins, Edward Hughes and Louis Kirsch, all from Google DeepMind, founded the London public benefit corporation with Kaloyan Aleksiev of Reka AI and Microsoft. Index Ventures and Radical Ventures lead; NVentures, Ex/Ante, Metaplanet, Macroscopic Ventures and Mythos Ventures join. It is building Faraday, in which researchers work with self-improving AI agents. Danny Rimer of Index says the gap is AI that picks which questions are worth asking.
Why it matters: It is the money behind a result already on the record: three months later Inherent published Faraday, a 27B model trained to direct coding agents in replicating papers. The round shows investors backing a small lab whose bet is on how scientists and agents divide the work, not on the largest model.
Evidence and limits: The round says nothing about whether Faraday works; that is the later paper's question. Index confirms its lead; the other participants are from TNW's report.
Unreasonable Labs' ProjectionBench asks models to predict a recent paper's results from its research question alone, then with more detail; GPT-5.4 scores highest
What happened: Andrew J. Lew, Yuan Cao and Markus J. Buehler took 45 open-access Springer Nature papers from the previous six months on bioactive, mechanical and nanomaterials. Models get the topic and research question, then the null hypothesis, then the procedure, and predict the results at each stage, offline. GPT-5 breaks predictions and the paper's conclusions into atomic claims and scores their overlap. GPT-5.4 led (area under the curve 1.56, against 1.33 for Gemini 2.5 Pro) and kept an F1 of 0.70 with minimal context.
Why it matters: Most science benchmarks test what a model already knows; this one asks it to anticipate a result it cannot have read, by using papers newer than its training data. It also shows which context helps most: adding the null hypothesis raised scores more than adding the experimental procedure did.
Evidence and limits: GPT-5 extracts the ground truth and grades every answer, GPT models' included, which the authors flag as a possible bias. 45 papers in three materials areas, and the benchmark had not been released. Matching a paper's stated conclusions measures agreement, not whether a prediction is right.
Unreasonable Labs' LEIA predicts how 3D lattice materials deform and where stress builds as loads are applied step by step, checked only against simulation
What happened: Haiqian Yang, Yuan Cao and Markus Buehler of Unreasonable Labs trained a model that predicts, load step by load step, how a lattice material deforms and where stress builds. On MicroPlate, their new set of 63 lattice plates with meshes of up to 442,000 nodes, its stress predictions correlated 0.94 with simulation on shapes it was trained on and 0.66 on eight held-out shapes. A design search ranked 553 candidates in 30 minutes, each 100 to 300 times faster than finite-element simulation.
Why it matters: It is the first result on the record from a company that says its engine proposes and tests hypotheses, and every test here is against the authors' own finite-element runs: nothing was built or loaded. The authors say it covers only cubic lattice plates under four kinds of load, and the benchmark is their own.
Biohub opens ESMFold2 under an MIT licence, and reports its designed binders hitting their targets in the lab for 36% to 88% of the small ones
What happened: The ESM team, which joined Biohub in November 2025, released three things at once: ESMC, a language model trained on about 2.8 billion protein sequences; ESMFold2, which predicts all-atom structures of proteins and their complexes; and ESM Atlas, a map of 6.8 billion sequences and 1.1 billion predicted structures. In the preprint, Sal Candido, Alexander Rives and 39 colleagues designed binders for five cancer and immune targets, then made and tested them: 36% to 88% of the compact minibinders bound, and 15% to 29% of the antibody-derived ones.
Why it matters: A structure model this open changes what a lab without a deal can do: MIT weights mean a group can fold complexes, run the binder search and keep what it designs, and Biohub says the search ran in days where it would otherwise take months or years. The atlas is the other half: organising 6.8 billion sequences by what the model learned about them makes searchable the biology Biohub says has never been annotated. Biohub's own FoldBench runs put ESMFold2 at or above AlphaFold 3 on antibody-antigen and protein-protein complexes.
Evidence and limits: Biohub designed, made and tested the binders itself, and the benchmark comparison with AlphaFold 3 is its own run of FoldBench; no outside group has reported either. Binding in a dish is not a drug: Biohub reports affinity, specificity and stability, not activity in an animal or a person. The 36% to 88% is the compact minibinders; the antibody-derived designs, the format closer to a therapeutic, ran 15% to 29%.
Perceptic, founded by three former Palantir engineers, comes out of stealth with a $12M seed round led by Accel to tie together the data and AI tools behind pharma's drug development decisions
What happened: Tilman Flock, chief executive, Martin Copes and Zaki Trache helped build Palantir's life sciences practice. Their software runs AI agents over a drug company's public, internal and purchased data, tracing each claim to its source, for three jobs: scouting assets to license, choosing which indications to test, and organising clinical data for trial design. Sonali De Rycker led for Accel, with Air Street Capital, whose founder Nathan Benaich gave a statement, and Elder Gull. The money is for engineering and customers.
Why it matters: Flock says most AI startups in drug development improve one step, such as a structure, a binder or trial recruitment; Perceptic pitches the layer that connects them, the decisions about which asset or indication to pursue. Every result is the company's own: due diligence cut from weeks to hours and a 50-fold rise in clinical data extraction come with no method, and of several large drug companies it says use the software it was allowed to name only CSL. Fortune reported it, quoting the chief executive and both investors.
AlphaProof Nexus resolves nine open Erdős problems with Lean-checked proofs
What happened: Google DeepMind's agent has a language model write proofs in Lean, whose compiler checks every step. Run on all 353 Erdős problems formalised in the open Formal Conjectures repository, it resolved nine, two of them open for 56 years, at a few hundred dollars each in inference, and proved 44 of 492 open conjectures from the OEIS integer-sequence encyclopedia. A basic loop of model edits and Lean checks matched the same nine, at higher cost on the hardest. The peer-reviewed version appeared in Science on 8 October 2026.
Why it matters: Lean checks every step, so a proof that compiles is right without a referee, and one agent could be set loose on hundreds of open problems at once; the team's own experts checked, for each of the nine, that the Lean statement said what Erdős asked. The check proves a statement true, not that it mattered, and nine of 353 leaves most of these problems out of reach. The cost per solved problem also leaves out the larger bill of trying all 353.
Kemira says CuspAI's generative design searched 300 trillion structures and left about 20 priority candidates for PFAS removal, now going into development and testing
What happened: Kemira, the Finnish water-treatment chemicals company, set the brief: remove GenX, PFBS and PFOS from water at sub-parts-per-billion concentrations, with chemistry that is water-stable, environmentally compatible, synthesisable and cost-effective. CuspAI's platform designed metal-organic frameworks from scratch across about 300 trillion possible structures and delivered over 5,000 designs with property data for the three molecules, narrowed to about 20 priority candidates, in six months. The project now moves to development and testing, with further material classes being scoped under the framework agreement.
Why it matters: A named industrial buyer has written down the brief it gave an AI materials company, target molecules, concentration, stability, synthesisability and cost, and what came back against it. Today's leading remediation technology is granular activated carbon, and Kemira, which is in the activated-carbon regeneration business, says it wants an alternative that is more selective and lasts longer; the next phase has to show at least one candidate made and measured against that bar.
Evidence and limits: Everything is Kemira's and CuspAI's own account, with no paper or data released. The release says the designs carry "full property data" without saying whether any of it was measured, and records no candidate synthesised or tested. Who runs the next phase, the deal's terms and who owns the candidates are not stated.
Incyte pays Genesis Molecular AI $120M upfront and shares its own lab data to train Genesis's drug-design models
What happened: Incyte and Genesis expanded the collaboration they began in February 2025. Genesis receives $80M in cash and a $40M equity purchase upfront, recurring research funding for model training and inference compute, and up to $232M in milestones per programme. Incyte adds at least five targets and holds exclusive rights to any drug. It will also share proprietary experimental data to train GEMS, Genesis's platform. Pablo J. Cagnoni, Incyte's head of R&D, and Evan Feinberg, Genesis's founder and CEO, spoke for the two companies.
Why it matters: The drugmaker is paying for more than designs: its own experimental data will train a startup's foundation model, and it is funding the compute as well. Data on how real molecules bind and behave is what public datasets lack, and Feinberg calls proprietary data among the most valuable inputs for molecular AI. Here a buyer puts money on that view.
Evidence and limits: Committed are the $120M and research funding of an undisclosed size; the milestones depend on drugs advancing toward the clinic and the market. The release names no targets, gives no results from the first programmes beyond Incyte calling them strong, and does not say what data is shared or who owns models trained on it.
What happened: A Nature paper reports multiple biomedical case studies and experimental follow-up for hypotheses produced by the multi-agent system.
Why it matters: Peer review and documented follow-up make the claim more inspectable than the original product preview, even if broad autonomy remains unproven.
Robin links hypotheses, experiments and updated hypotheses
What happened: The system combines literature search, hypothesis generation, experimental strategy and data analysis; human researchers run the physical experiments and return the data.
Why it matters: Unlike one-shot hypothesis agents, experimental evidence returns to the system and changes what it proposes next.
Google's ERA, in Nature: a language model searching over code writes single-cell and COVID-19 forecasting software that beats the best human methods on their benchmarks
What happened: Eser Aygün, Michael P. Brenner and colleagues at Google Research and Google DeepMind turn writing scientific software into a search: a language model rewrites the code and tree search keeps what raises the task's score. On held-out data from the OpenProblems single-cell benchmark, 40 of its methods beat the best published one; the top, a reworked BBKNN, scored 14% above ComBat. Rerun over the 2024-25 COVID-19 season on data available on 1 May 2025, its model's average error was 26, against 29 for CDC's ensemble.
Why it matters: These are real benchmarks and real public forecasts, not toy tasks, but every test was scored on data that already existed, and Google chose and ran them. The COVID-19 result is a retrospective rerun, and the 14 models the abstract says beat CDC's ensemble were scored over three weeks. It shows that where a task has a score, search over code can reach expert software; it says nothing about tasks without one. The forward test came later, in CDC's scoring of the flu season.
Mistral AI buys Emmi AI, a Linz start-up whose neural networks stand in for engineering simulations
What happened: Emmi AI, founded in 2024 by Johannes Brandstetter, Dennis Just and Miks Mikelsons, builds models that it says predict fluid flow, heat transfer and other physics in seconds, where a numerical solver takes days or weeks. Its founders and more than 30 researchers and engineers join Mistral's science and applied AI teams, and Linz becomes a Mistral office. Arthur Mensch, Mistral's CEO, pitched the deal at aerospace, automotive and semiconductor makers. Terms were not disclosed; Emmi had raised a EUR 15M seed round in April 2025.
Why it matters: With the purchase, a general-model company took on a physics-AI team, making simulation a line of work at Mistral beside language models. Five months later Brandstetter, by then a Mistral vice-president, spoke for the company on its plan with CERN for a collider-physics model.
Evidence and limits: No price was given. 'Seconds instead of weeks' is Emmi's own description of its models; the announcement shows no benchmark.
MIT and UBC's WorldParticle uses one transformer design to simulate fluids, cloth, sand, elastic solids and protein motion
What happened: Caoliwen Wang, Minghao Guo, Peter Yichen Chen, Wojciech Matusik and colleagues at MIT, the University of British Columbia and elsewhere built a transformer that moves particles forward one step at a time, and trained it on six kinds of physics: two kinds of fluid, cloth, sand, elastic solids and protein molecular dynamics. Each was learned from a conventional solver's output (OpenMM for proteins) and tested on new materials, shapes and forces. Against three earlier learned simulators, it kept long runs stable where they left sand floating and cloth exploding.
Why it matters: Simulating sand against a soft gripper, or liquid against a flexible valve, today takes a separate solver for each material and code to couple them. One design that covers all of them is a step toward a simulator that couples them by default. A learned simulator can also be run backwards for design, which the authors use to fit friction parameters. For proteins it advanced side-chain motion 50 femtoseconds per step, 100 times the step molecular dynamics uses.
Evidence and limits: Each kind of physics has its own trained weights, so it is one architecture, not yet one model, as the authors say. Every test is against the solvers it learned from, real-world data enters only as three manipulation sequences compared by eye, and the protein test covers side-chain rotation, not folding. MIT News's "100 times faster" is that larger time step, not a timed comparison. A scene tops out near 50,000 particles on one GPU.
METiS, which designs drug formulations with AI, lists in Hong Kong
What happened: METiS priced at HK$10.50, raising about HK$2.11B gross, and began trading on 13 May as 7666.HK. Eighteen cornerstone investors took 55% of the shares.
Why it matters: Unlike most AI drug companies on this record, METiS designs delivery: lipid nanoparticles and formulations. Its 6,911-times oversubscription measures demand for shares, not whether the platform works.
WhiteLab Genomics says AAV carriers designed by its AI platform put about 50 times more DNA into the brain than AAV9 after one injection into a vein in mice, with none found in the liver
What happened: At an ASGCT 2026 oral presentation in Boston, the Paris company reported that shells designed end to end by its ALFRED platform reached roughly 50-fold higher DNA enrichment in the brain than AAV9, the standard carrier for brain disease, after one intravenous injection in wild-type mice, and were undetectable in the liver. ALFRED works out which cell-surface receptors control entry to a tissue and designs around them rather than screening random variants. The work was done with Françoise Piguet's GENOV lab at the Paris Brain Institute.
Why it matters: Many gene therapies for brain disease are injected into the brain or the spinal fluid because the blood-brain barrier stops medicines carried in the blood, and AAV9 given by vein also reaches organs such as the liver, which the company says creates potential safety and dosing limits. A designed shell that gets into the brain from a vein, and not into the liver, would change how such therapies are given, and it was designed from receptor biology rather than found by screening.
Evidence and limits: Everything here is the company's own account: the release gives no dose, no number of animals, no timepoint and no figures, and there is no paper. What was measured is DNA in the tissue; the release also claims functional targeting, with no data shown for it, and reports no effect on an animal's disease. Mice only, and the liver claim is that nothing was detected, which rests on an assay limit the company does not give.
Recursive Superintelligence launches with $650M co-led by GV and Greycroft to build AI that improves itself, valued at $4.65B
What happened: Richard Socher, formerly Salesforce's chief scientist, founded the San Francisco and London company with seven co-founders including Tim Rocktäschel, who led open-endedness work at Google DeepMind, Jeff Clune and Alexey Dosovitskiy. GV and Greycroft co-lead; NVIDIA and AMD Ventures join. Tom Hulme writes for GV that the first goal is a system with the capacity of '50,000 PhDs', aimed first at the science of AI itself.
Why it matters: It funds the most direct version of automated science: a system that proposes experiments, runs them and keeps what improves its own code and training. The company says it will turn the system to drug discovery, batteries and fusion once it works on AI research, so its first results will be about machine learning, not the natural sciences.
Evidence and limits: The valuation is the company's, repeated by GV, without a pre- or post-money basis. No result has been published; the plan for other sciences is a stated aim.
BiomniBench: on 100 real biomedical analysis tasks, changing the agent harness moved scores more than a model generation did
What happened: Phylo and Stanford researchers built 100 data-analysis tasks from Nature, Cell and Science papers with their authors or domain experts, and scored each agent's whole trace against expert rubrics. Claude Opus 4.7 scored 73.34 in Claude Code and 63.94 in Terminus-2; moving from Opus 4.6 to 4.7 within Claude Code added 3.8. No setup scored above 73.34 of 100.
Why it matters: A model's score on science work means little without its harness. The cleanest gap is Opus 4.7 at the same reasoning effort; GPT-5.4's larger 13.5-point gap also changed effort and turns (13 to 82). An LLM judge scores the traces (Gemini 3.1 Pro, κ 0.70 against experts), each setup was run three times, and the tasks redo published analyses rather than find anything new.
Isomorphic Labs raises $2.1B in a Series B led by Thrive Capital to move its AI-designed programmes toward the clinic
What happened: Alphabet and GV returned; MGX, Temasek, CapitalG and the UK Sovereign AI Fund are new. The money goes to its drug design engine, IsoDDE, moving its programmes toward the clinic, and hiring. Demis Hassabis, founder and CEO, says the approach is "fundamentally sound"; Max Jaderberg, its president, says the engine has hit milestones in internal programmes. Joshua Kushner of Thrive and Krishna Yeshwant of GV speak for the investors.
Why it matters: The round is 3.5 times its first outside round, $600M thirteen months earlier: capital is paying for the step from models to medicines before any clinical result. The milestones and speed it cites are not shown in the release, and it gives no valuation.
Microsoft Research's MatterSim screen picks tantalum phosphide as a heat conductor, and a sample made at UT Dallas measures 152 W/m/K, close to silicon
What happened: Microsoft Research AI for Science screened over 240,000 candidate crystals for thermal conductivity with its MatterSim-v1 atomistic model and picked tetragonal tantalum phosphide (TaP). It was made at the University of Texas at Dallas and measured at the University of Illinois Urbana-Champaign: 152 W/m/K on the best samples. The compound had been made before, but not considered as a heat conductor. The same post announces MatterSim-MT, a model that also predicts charges, magnetic moments and dielectric response.
Why it matters: A screen that ranks hundreds of thousands of crystals is useful to a buyer only if a pick survives being made and measured, and this one did, at a value near silicon's. The post names where better heat conductors are wanted: cooling in computing, power electronics and aerospace. For teams choosing which candidates to send to a lab, it is one worked case of an atomistic model narrowing the list for a property as sensitive as thermal conductivity.
Evidence and limits: One material, and the post is the only source: no paper reports the synthesis or the measurement, and the model's predicted value for TaP is not printed beside the measured one. 152 W/m/K is for the best samples. The labs that made and measured it are collaborators on the project, not an outside check. The MatterSim-MT preprint says its code and weights will be released on publication.
Periodic Labs is in advanced talks to raise at least $500 million at a $7.5 billion valuation, led by AMP
What happened: Per three sources who spoke to Forbes, the round is led by AMP, the investment vehicle Anjney Midha founded after leaving a16z. Periodic did not respond and AMP declined to comment. Its seed round last September was $300M at $1.3B.
Why it matters: Liam Fedus and Ekin Dogus Cubuk founded Periodic to train models on data from its own autonomous labs; this would value it at nearly six times its seed price. Talks only, and media-reported: the price says how much capital is betting on that route, not that it works.
MolRecBench-Wild shows old OCSR benchmarks overstate how well models read real chemistry papers
What happened: 5,029 drawings from 820 papers, balanced for difficulty and labelled across 37 classes, split OCSR into seeing the image and representing the chemistry. 42% carry both; the best of 18 models reaches 41.05% SMILES accuracy.
Why it matters: Over-90% exact match on patent benchmarks like USPTO overstates OCSR on paper drawings; the shares reflect how drawings were picked, not the literature, and this is not production performance.
Ge Lei and Samuel Cooper at Imperial show that an LLM guiding experiments changes its predictions and confidence with the wording and format of the question, not only with the data
What happened: Ge Lei and Samuel J. Cooper used LLMs as surrogates, the model a Bayesian optimizer consults to choose the next experiment, on sparse data from known functions. Correct hints about a function's shape acted like a prior and wrong ones misled; asking for each point separately or all at once gave different predictions and uncertainty; as observations were added, confidence dipped and recovered, the dip moving with their order. In three optimization tests, including a simulated cathode design, these choices changed which experiments were picked.
Why it matters: LLMs are increasingly used as the surrogate in low-data optimization, where only a handful of measurements exist and the model can bring in domain knowledge a Gaussian process lacks. This paper shows that the prompt and the query format are part of that model: two groups using the same LLM on the same data can get different suggestions, so a lab reporting an LLM-guided campaign needs to report how it asked. In the cathode task, a pointwise prompt carrying domain guidance gave the lowest final regret, narrowly ahead of the SMAC3 optimizer.
Evidence and limits: The tasks are mostly noise-free functions chosen for control, and each optimization study used one model (GPT-4o, Qwen3.5 or GPT-5.4), so it does not compare models. On the standard Branin test the LLM beat a Gaussian process, but not once the optimum was moved; on hyperparameter tuning SMAC3 reached the best final result. The cathode task is a simulation. A preprint.
Joachim Baumann and colleagues at Stanford find that one AI rewrite of a paper, costing about $0.25, raises AI reviewers' scores
What happened: One rewrite by GPT-5.1 or GPT-5.4, with no new experiments run, raised AI reviewers' scores on 60 ICLR 2026 submissions by 0.45 points on a 1-10 scale on average across 24 prompt and model combinations, Joachim Baumann, Jiaxin Pei, Sanmi Koyejo (Stanford) and Dirk Hovy (Bocconi) report in an ICML 2026 position paper. Among all 75,800 ICLR 2026 reviews, fully AI-written ones were more alike across papers than the rest; on the 8,015 papers with both kinds, averaged AI scores predicted acceptance worse than human ones (AUC 0.71 against 0.82).
Why it matters: Conferences already let AI into review: ICLR 2026 allowed LLMs to help write reviews, and AAAI 2026 issued fully AI-generated reviews alongside human ones. If one cheap rewrite moves AI reviewers the same way, the authors argue, papers will be written for the model; the rewritten papers here already became 6.5% more alike. The authors set two tests an AI reviewer must pass before use, that it keeps reviews diverse and cannot be gamed by style, and find today's fail both.
Evidence and limits: The rewrite test used 60 papers from one venue, four prompts (one telling the model to jailbreak the reviewer), two GPT rewriters and three reviewer models; GPT reviewers rose more than Claude, which the authors link to models favouring their own style. Similarity is measured on text embeddings, which track wording and topic rather than distinct arguments, and which ICLR reviews were AI-written comes from a detector that can err.
