Big shoutout to all of our partners for making Nemotron 3.5 Super VL their own through post-training 🙌 Here’s what they’ve built 🧵
In Nature this week, Satya Nadella says his organization’s new research shows how machine learning can improve weather forecasts two to six weeks out, a timeframe that has historically been very difficult to predict. He highlights the work’s potential to help people better prepare for extreme weather.
Senior AI reporter James O’Donnell interviews Stanford University bioengineering PhD candidate Samuel King about using a generative AI model to propose genetic blueprints for microscopic viruses. The session covers King’s work, his recognition as one of MIT Technology Review’s Innovators Under 35, and new ways of seeing biology. The source says this is not yet an example of AI-generated life.
Seems very important, as tons of people are tweeting about the end of bigger open models from China due to this from MOCOM, but there's a statement intending to make it easier for open source technology. Much hinges on implementation and how "near automatic" approval would work. https://twitter.com/ZilanQian/status/2111073916194013274
NVIDIA and AI4Bharat at IIT Madras introduce Nemotron-pretraining-Indic-v1, approximately 2.8 trillion tokens across 22 Indian languages, releasing soon alongside its open, GPU-accelerated generation pipeline. The corpus includes native-script text for all 22 languages and romanized text for twelve, translated from curated English sources with quality checks and source tracing. An Indic continued-pretraining stage improved five reported benchmarks, but combined synthetic and natural data, so the comparison does not isolate this corpus’s contribution.
Why it matters: The workflow offers a reusable approach to multilingual data generation by separating translation, quality assessment, repair, and script conversion while preserving links to source documents.
Mira Murati says one of the best parts of this year, inside Thinky and out, has been watching people follow their own ideas into weights they own. https://twitter.com/thinkymachines/status/2110851669915570685 In the quoted post, @thinkymachines marks one year since Tinker’s launch with a look back at what it has shipped and what users are building. https://thinkingmachines.ai/news/one-year-of-tinker/
The New York Times visualizes the staggering amount of math needed to train AI models, using data from Epoch's research teams 📊 https://t.co/Cz21cuf9oQ Our training compute estimates: https://t.co/X7jSJHtWnz The Epoch Capabilities Index (ECI): https://t.co/RbP83Metg6
Tinker lets customers turn their expertise into cutting-edge models that they own. One year after launch, we look back at what we’ve shipped and what our users are building. https://t.co/fWTOG5U3k1
We're expanding access to GPT-Rosalind for academic research. The new Rosalind Academic Plan helps scientific labs bring AI into the work of understanding biology, exploring ideas, and deciding what to test next—with sponsored access for eligible labs at U.S. universities. Academic research starts with questions whose impact may take years to impact human health.
Tang Duyu, former head of multimodal model development for Huawei’s Xiaoyi assistant, founded Guanxing Shuzhi to develop physical AI that identifies needs and decides when to intervene without explicit commands. Its ProAgent architecture pairs StreamingJev, a lightweight event monitor, with a Response Model that plans interventions and selects devices. The team reports testing guidance, warnings, and interaction on Reachy Mini across 60 cases, and introduced ProactiveXBench to evaluate intervention timing, progress, and outcomes.
Brandon Anderson interviews Leash Bio’s Ian Quigley about generating reproducible protein–small molecule interaction data using DNA Encoded Libraries and detecting misleading signals in experiments and models. They discuss a Kaggle competition with about 4B measurements where nobody beat chance on an unseen chemical library, and an agent sandbox searching 6.5M molecules on a budget of 100. Anderson reports that the sandbox has matured into Minos, with a preprint, blog, and open dataset being released today.
10 new papers in health science research. all written with muse models in under 4 weeks. health is one of the areas where we believe muse is going to have the greatest positive impact on humanity ❤️ https://t.co/eejFDFd2Co
Meta says it is exploring how AI can accelerate health science research. Medical experts used a human-led research scaffold powered by Muse Spark 1.3 and Muse Wonder, with access to 5M retrospective de-identified electronic health records and publicly available data, to conduct original research and speed hypothesis generation, code execution, and manuscript drafting while retaining authority over scientific decisions. The 10 manuscripts span cardiology, oncology, infectious disease, emergency medicine, and more.
Researchers used AI-based protein design tools to retain the catalytic cores of two types of protease while replacing the rest of each protein. The resulting enzymes could be tuned to cut specific proteins in half, and the researchers demonstrated their effectiveness by cutting proteins associated with neurological diseases.
Microsoft researchers and collaborators published research in Nature on Probabilistic Bias Correction (PBC), an open-source machine learning tool that corrects systematic bias in existing weather models to improve forecasts two to six weeks ahead. In the study, PBC improved ECMWF’s physics-based model accuracy for 98 percent of precipitation and 92 percent of temperature targets, and doubled the accuracy of its AI forecasting system. The tool can run on CPUs, though ECMWF cautioned that operational implementation remains challenging.
Scimplify raised $25 million in an all-equity strategic round led by Hitachi Ventures to expand its AI-powered specialty chemicals and critical materials business. Its Scinode platform uses specialized AI agents to identify production routes and facilities and track quality and regulatory requirements. Co-founder Sachin Santhosh said the startup plans to allocate around $7.5 million to U.S. expansion and another $2 million to Scinode development.
MemSilicon announced on October 12 that it completed tape-out of MountTai 2026, its first-generation high-density RRAM chip based on the 1TnR approach, with validation underway. The company reports a 28nm CMOS process, a 1T4R structure, 1Gb capacity, 30Mb/mm² density, access latency below 20ns, and on-chip hardware error correction. Its architecture shares one selection transistor across four resistive memory cells to reduce transistor area overhead per cell.
The ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious
Trillium Labs is hiring for several roles, with near-term starts preferred, to build an open scientific ecosystem for LLMs. The roles include full-time post-training researchers, engineers experienced with open-source libraries such as megatron, MILES, sglang, and Tinker, mission-aligned generalists, nonprofit operations staff, and student researchers. Nathan Lambert says he is also interviewing for MATS and asks applicants to reply, email or DM him, and submit the form on the company website.
AI labs are buying defunct startups’ code and workplace archives through companies such as SimpleClosure and Sunset to train models and build environments where AI agents practice tasks. SimpleClosure’s Asset Hub has handled over 150 deals, while Google bought Spirit Airlines’ decades of data for $10 million this past August. The market raises questions about employee privacy, consent, and pricing, and no company has released tests comparing training results from real-world archives with alternatives.
Why it matters: The sales expose a gap between who generates workplace data and who receives its resale proceeds, with employees’ private information entering transactions that typically benefit creditors and investors.
Nathan Lambert calls a piece on China’s LLM data economy the best he has read so far on how it compares with the US. He says Chinese labs seemed 6–12 months behind US labs in willingness to spend on data, but are rapidly catching up, and notes that Tencent, Alibaba, and ByteDance lead Chinese spending rather than OpenAI and Anthropic.
Robotics and physical AI companies raised nearly USD 48 billion through October 8, 2026, about 1.7 times the total for 2025, according to PitchBook data cited by the Financial Times. Funding increasingly concentrates in leading companies while supporting robot hardware, general-purpose models, training data, and deployment. The report examines paid video collection, robot training facilities, and commercial deployments as ways to gather data, while acknowledging that data quality, task coverage, and transfer across robots also affect learning.
Researchers from Sakana AI and UC Berkeley introduced MASS, a method that uses copies of Qwen3.6-27B to solve tasks, evaluate results, and optimize workflows without a stronger teacher model. After two rounds, its test-task win rate against the base model reached 69.9 percent, although benchmark gains largely reflected more completed work and some scores declined. Shared blind spots allowed flawed results to pass evaluation and enter training, prompting the researchers to recommend independent checks, limits on autonomy, and clear intervention conditions.
The U.S. Air Force awarded WindBorne Systems a contract worth up to $49 million for atmospheric data collected by its autonomous weather balloons, with $1.7 million ordered so far. The Air Force pays as needed and can request data for specific regions and times. WindBorne operates about 700 balloons globally and also feeds their readings into its AI weather forecasting model; executives say the contract can support fleet expansion and hardware research and development.
Microsoft Research introduces StudentSim, a research system that trains personalized student simulators from publicly available, de-identified learner records to evaluate and improve AI tutors. StudentSimEval measures behavioral fidelity and guidance responsiveness across chess, second-language English writing, and mathematics, where StudentSim outperforms leading baselines. In a chess tutor study, training with StudentSim feedback achieves 90.5% accuracy versus 75.7% without reinforcement learning and 71.6% with GPT-5.4 feedback; real learner studies remain essential, and the paper and code are publicly available.
Why it matters: Evaluating both a simulated learner’s starting ability and response to guidance offers a concrete method for assessing tutor feedback beyond the correctness of an explanation.
Starting October 1, 2026, arXiv limits submissions to two papers per person per calendar month, keeping the existing cap of three active submissions. The platform says AI tools helped push monthly submissions past 40,000, with the cs.AI category growing more than sixfold in two years and nearly 9,000 support tickets generated. Editorial advisory council chair Thomas G. Dietterich calls the rule a temporary fix until moderation tools scale.
Data from Ornn, Silicon Data, and Bloomberg as of August 2026 shows what a16z calls a textbook Jevons paradox in the AI market: token prices keep dropping while H100 GPU rental prices hold steady or climb. Cheaper tokens unlock AI agents, automation, and new applications, so volume grows faster than per-unit costs fall. The system depends on AI usage growing fast enough to offset falling token prices; if demand flattens, chip makers, memory suppliers, energy providers, and cloud companies take a hit.
On October 7, OpenAI released 722 mathematics manuscripts, including two on the Kakeya conjecture: a 97-page paper claiming a three-dimensional maximal-function result stronger than Wang and Zahl's, and a 175-page paper claiming the four-dimensional Hausdorff dimension case. Both are attributed to OpenAI, are not Lean-formalized or peer-reviewed, and OpenAI's README says some unformalized results may be flawed.
Why it matters: The piece traces how two OpenAI manuscripts extend Wang and Zahl's three-dimensional Kakeya work to a stronger maximal-function version and to four dimensions, while noting the proofs lack peer review.
Nathan Lambert says the data industry has grown, but low-quality output can amplify behaviors such as reward hacking. He is advising Mercor on research into sample-efficient data for RL and says the field needs stronger evaluation science and benchmarks tied to areas of clear economic value. Lambert expects major investment around frontier labs in open inference, open post-training and data, with opportunities to build on open models.
Standard Bots, which calls itself "America's largest AI-native industrial robot manufacturer," described its AI stack for industrial robot arms used in machine tending, welding, and assembly. Its largest model is in the low billions of parameters, trained on over a billion images for zero-shot perception, with training in the cloud and inference run on-premises via edge GPUs. The company recently raised $200 million at a $1 billion valuation in a Series C led by General Catalyst and RoboStrategy.
TypeSafe announced its Series AI round, and Sequoia disclosed that the company crossed $100M ARR in its first week, three weeks after launch, according to this AINews roundup. The same issue reports several vendors shipping typed decision models on one day, including OpenAI's Decisions API on GPT-6 Luna at $0.10/M input tokens and Perplexity's pplx-decider-v1.1-27b claiming 94.5% on Decision Bench.
Columbia University's Li Yunzhu team presented progress on FlexiTac, an open-source flexible piezoresistive tactile sensor platform, at an IROS 2026 workshop. The sensor assembles in 3 minutes, ships with a full BOM and readout software, and integrates with the LeRobot robot-learning framework. Analog Devices is developing a commercial version with tactile resolution five times that of a human finger, produced via semiconductor manufacturing.
At IROS 2026, University of Queensland researcher Jen Jen Chung argued in a keynote that data volume is not the same as information, and that blindly piling up data can degrade robot policies. Drawing on drone soaring, multi-robot communication and contact-rich assembly work, she said data value is unevenly distributed and that sharing or collecting only decision-changing information beats full broadcasts or uniform sampling.
StepFun's flagship base model Step 5 Preview reached the top of OpenRouter's Trending list and second place on its daily ranking, with an official open-source release set for October 15. The sparse MoE model has 600 billion total parameters and about 27 billion activated per token, and Artificial Analysis gives it a composite intelligence index of 44, placing it in the global open-source top three.
At IROS 2026, running September 27 to October 1 in Pittsburgh, Leiphone reports embodied AI research is converging on four trends: foundation models internalizing physics, "brain-cerebellum" architectures becoming standard, touch as core to dexterous manipulation, and cheaper data collection. Papers cited include ULTRA, LITHE, VOFA, KineFuse, VTAP Gripper, ViHaTeleop, and HALO, alongside workshops on physical world models and uncertainty.
Mifeng Technology launched Mifengpai, which it calls the world's first full-category high-quality physical AI data crowdsourcing platform, at a Shanghai event on September 23. The platform pairs MEgo capture devices and the MEgo Engine processing platform with a collector network, and covers 22 categories, over 5,000 task types and more than 50,000 real scenarios. The company also announced a robot trainer program and a 100 million yuan subsidy plan.
As of March this year, China's daily average Token call volume exceeded 140 trillion, and large-scale inference is driving SSDs to extend from capacity and persistent storage toward the compute side. Samsung demonstrated a Z-NAND architecture, placing model weights in Z-NAND and keeping KV Cache in DRAM; large North American internet companies have begun replacing some HDDs with high-capacity QLC SSDs. Data management is also shifting from fallback-style tiering to proactive tiering at the outset of system design and dynamic scheduling during operation.
In their conversation, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido pointed out that AlphaFold's breakthrough is only the beginning, protein structure prediction is far from solved, and protein dynamics and disordered structures remain difficult problems.
TypeSafe AI completed an $870 million funding round weeks after the launch of Jev, reaching a $7.5 billion valuation, led by Andreessen Horowitz with participation from Sequoia and existing investor DCVC.
This week, OpenAI suddenly released nearly 400 AI-generated mathematical results, spread across more than 700 manuscripts and covering directions including combinatorics, geometry, number theory, theoretical computer science, algebra, topology, probability and statistical mechanics, and mathematical physics.
Why it matters: Through first-hand reactions from dozens of mathematicians, this presents the impact of OpenAI's one-time release of nearly 400 results on the academic ecosystem and researchers' career paths.
