Anthropic says Claude discovered a previously unknown enzyme system in bacteriophage DNA, next to a stretch of repetitive DNA resembling CRISPR. The system's function is not yet known, but the few known systems with similar features can cut, copy, and paste DNA. Anthropic says more research is needed to determine what it does and whether it could be used for gene therapies like CRISPR.
NVIDIA validation engineer Sakeena Fiza works in the data-center systems engineering lab, validating hardware before new products enter mass production. One of her earliest and most memorable moments was seeing the NVIDIA Rubin GPU enumerate successfully at the system level for the first time. Validation covers the full path from tray, rack, and cluster to production line and customer AI factory, aiming to find problems before customers do.
Radical Numerics co-founder and CEO Eric Nguyen argues that biosecurity is an AI arms race: models that increase biological capabilities can also help defenders keep pace with the offensive capabilities of frontier models.
Google's Empirical Research Assistance (ERA) uses Gemini to maintain a tree of experimental notebooks, select and mutate candidate approaches using an Upper Confidence Bound rule, and automatically solve scientific problems that can be expressed as scoring functions. It has produced at least ten papers.
Xiaomi has released the MiMo-V2.6 family, including the multimodal MiMo-V2.6-Pro and MiMo-V2.6-Flash, and introduced MiMo-V2.6-Pro-UltraSpeed, which delivers up to 20× faster output.
Why it matters: Xiaomi's first entry in the frontier open-model lineup, with open RL environments and training recipes, provides a concrete example of post-training cost strategies.
Microsoft has published RetroChimera, a retrosynthesis-prediction framework, in Nature and open-sourced its implementation and weights. It combines the Transformer model R-SMILES 2 with the GNN model NeuralLoc, using a learned ensemble to rerank their predictions and performing strongly on both common and rare reaction types. In blind tests, chemists preferred its single-step reaction predictions. It found complete routes for nine of ten challenging targets, compared with five for a de novo model, four for an editing model, and two for NeuralSym.
Over a 15-month investigation with The Times of San Diego, MIT Technology Review created the first comprehensive map and analysis of deaths near surveillance towers along the U.S.–Mexico border.
At New York Climate Week, NVIDIA highlighted five companies using AI to advance clean energy. ThinkLabs AI uses digital twins and agents to reduce the evaluation time for grid-connection applications from 30–45 days to two minutes.
Google Research has released MilleMiglia, a C++ instance generator that creates realistic, privacy-preserving benchmark data for middle-mile logistics and delivery problems. Its source code and documentation are available on GitHub.
Researchers from @BroadInstitute, @UniofExeter, and other institutions are using AlphaGenome Atlas to better identify potentially disease-causing DNA variants and interpret their effects. 🧵
Latent Space's AINews roundup of developments from September 15–16 reports that Steve Yegge shut down Gas Town and acknowledged spending thousands of dollars a month on coding-agent subscriptions to build just that one project.
NVIDIA submitted preview results for Vera Rubin NVL72 for the first time in MLPerf Inference v6.1, reporting up to 3.7× the throughput of GB300 NVL72 on Qwen3-VL and up to 2.5× on DeepSeek-R1.
Why it matters: MLPerf Inference v6.1 reports throughput, scaling efficiency, and software-optimization gains for Vera Rubin NVL72 versus GB300 NVL72, offering a basis for judging inference infrastructure value.
Hex's data agent uses GPT-6 Astra to turn analytical answers into interactive visual reports that employees can share.
At Salesforce Dreamforce, NVIDIA founder and CEO Jensen Huang appeared alongside Marc Benioff as Salesforce released Koa, its first CRM reasoning model, built on NVIDIA Nemotron 3 Super.
Why it matters: Huang discussed AI safety and shifts in the share of open models at Dreamforce, while the article explains how Salesforce's first reasoning model was trained and when it will be deployed.
Good Start Labs turned the 1830 railway board game into a training environment for experiments with 30B models. Both single-turn question-and-answer and multi-turn terminal-agent training improved performance in the game, but only the terminal-agent setup improved results on the Finance-Agent benchmark. The company, spun out of Every and funded with $3.6 million, argues that how a game is presented determines which skills a model learns and whether they transfer to work outside games.
Google Research proposes Retrieve-for-Train, a framework that uses offline RL to discover reward-aligned query fan-outs and compile them into supervision, then distills them into a 53.9M-parameter diffusion retriever. At inference, it generates a complete result set in one non-autoregressive pass without chain-of-thought tokens. The method fine-tunes fan-out language models based on Gemma3-4B and Qwen3-4B, using set-level attribute rewards to address paraphrase collapse in zero-shot LLMs and autoregressive latency.
At the AI Infra Summit, NVIDIA presented its DSX platform. DSX MaxLPS was validated on Lambda's five-rack, 19-node HGX B200 cluster, where it increased token throughput by 24% under the same power budget—from about 4 million to 5 million tokens per second—and improved performance per watt by 23%.
Why it matters: The article reports DSX MaxLPS measurements on a Lambda cluster, allowing readers to judge possible throughput gains under a fixed power budget.
At the AI Infra Summit, NVIDIA announced several advances in Vera Rubin and its DSX platform, shifting AI infrastructure measurement from peak performance to useful agentic tokens per megawatt.
Why it matters: The article reports several measured token-output figures per megawatt for Vera Rubin and DSX, showing how efficiency metrics for AI factories are changing.
Weather forecasting plays a key role in the transition to renewable energy. WeatherNext 3 provides grid operators and wind and solar power companies with forecasts of wind speeds at turbine height and solar irradiance, supporting smarter, cleaner energy planning. 🧵
Our own @johnschulman2 talks with Dwarkesh about where human judgment still matters as models improve and self-improve: teaching them to handle messy real-world tasks, applying taste to what works in the long run, and, above all, specifying what we actually want. Quoted post by @dwarkesh_sp (context, not the main author's work): New episode with @johnschulman2, @oneill_c and @BerenMillidge. I got together with some of the most insightful AI researchers I know who are at the openish companies, because I wanted to hear the details of what's actually happening at the frontier and what comes next. 0:00:00 – Steelmanning the case against RSI 0:18:39 – What's driving the Chinese labs' progress 0:28:06 – How will automated AI researchers be trained 0:33:51 – Will long-horizon RL elicit AGI? 0:45:24 – The sim-to-real gap 1:00:33 – How much progress is explained by data? 1:18:03 – Why is RL working so well? 1:24:54 – Move 37 and entropy collapse 1:28:31 – Rapid-fire timelines
Google Research proposes ToolGrad, a “answer-first, question-second” approach to generating tool-use data: first create a ground-truth tool-call chain, then work backward to infer the user's prompt, using just one LLM call.
John Schulman says training on user data is exceedingly unlikely to contribute much to frontier models’ math gains, which he attributes to scaling pretraining and RLVR. He says user data may instead help identify failure modes that hired annotators cannot easily recreate, and calls for stronger disclosure norms about how companies train on user data and what capabilities they aim to improve.
César de la Fuente's lab uses Codex and ChatGPT to search the genomes of living and extinct organisms for candidate antimicrobial molecules to fight drug-resistant infections.
Training Agents 4: From reward functions to environments. https://x.com/i/broadcasts/1OxwbngOoXEJB
OpenAI has launched Data agent in ChatGPT Work, letting users connect company data and ask AI in natural language to surface insights and build interactive dashboards.
TRL v1.14's AsyncGRPOTrainer can now train and sync only LoRA adapters to vLLM. Rank-1 adapters are just a few megabytes and can be transferred through a Storage Bucket mounted across multiple Hugging Face Jobs, without NCCL.
Google DeepMind has released AlphaGenome Atlas, a platform containing predicted effects for 9 billion single-nucleotide variants in the human genome. At 1 PB, it is more than 30 times the size of the AlphaFold Database. From today, academic researchers can access it free through the web portal, AlphaGenome API, and a skill in Google Antigravity.
Why it matters: The article gives the scale of predicted effects for 900 million single-nucleotide variants and free academic access points, helping readers assess the availability of genome-interpretation resources.
In a paper published in Cell, Google, HHMI Janelia, Cambridge, and collaborators have published a connectome of the complete male fruit-fly brain and central nervous system. It contains more than 166,000 neurons and 125 million synaptic connections, making it the largest brain map to date by neuron count.
Why it matters: Google, HHMI Janelia, and partners have released the largest brain map by neuron count, showing a concrete application of AI in connectomics.
Google DeepMind and Google Research have released WeatherNext 3, which they call the most advanced and accurate global weather model so far, based on independent real-time evaluations by Brightband.
Why it matters: The article reports specific measures for resolution, update frequency, and precipitation accuracy, helping readers assess AI weather forecasting on real-time data.
A publicly available, low-cost recipe fine-tunes LFM2.5-350M with TRL's GRPO using about 500 samples and 100 training steps, raising the pass rate on IFStruct from 22.6% to 29.7%. Training can run on a free Colab or Kaggle GPU, while evaluation uses llama.cpp on a local MacBook. The code is open on GitHub.
A Hugging Face blogger reproduced Surya Narreddi's idea of getting a language model to paint watercolors using TRL and OpenEnv. The model writes about 150 lines of JavaScript with p5.brush to create the painting.
Hugging Face has released BenchMIRT, a multidimensional IRT method for auditing LLM benchmarks at the individual-prompt level and separating mixed ability signals within a benchmark. Trained on results from 100 LLMs across 16 benchmarks and more than 34,000 questions, it independently recovered two dominant dimensions—safety and general reasoning—without being told which benchmarks measured them. The analysis finds that scores on safety benchmarks such as BBQ and WMDP correlate more strongly with general reasoning, suggesting a single benchmark score may combine several signals.
Google Research has introduced MAPL-EMIT, a Swin-S transformer-based deep-learning framework that automatically detects, predicts, and localizes methane plumes from NASA EMIT hyperspectral satellite data. It achieved 84% recall on expert-labeled plumes.
Google has introduced TimesFM-3, a time-series foundation model pretrained natively for multivariate forecasting. It has 330M parameters, was pretrained on more than one trillion time points, and forecasts in a single forward pass. It supports multiple targets, historical covariates, and known future covariates, and ranks first for both point and probabilistic forecasting on the three public benchmarks Gift-Eval, FEV-Bench, and Time.
Microsoft has released GigaPath-Flash and GigaTIME-Flash, using a 22M-parameter ViT-S backbone distilled from the billion-parameter GigaPath encoder to sharply reduce the compute requirements of pathology foundation models. Both are open-source under Apache 2.0.
Cleaning data and aligning the reward function for RLVR takes expertise and effort upfront, but the result is a model that's state-of-the-art on a complex task. Guest post by researchers at UIUC and Bridgewater, in collaboration with our team. https://thinkingmachines.ai/news/putting-task-expertise-into-rl
Google Research has introduced the experimental Planetary Prediction Engine (PPE) under Google Earth AI. From a natural-language query, it can autonomously discover geospatial data, engineer features, train and evaluate models, and produce a report—compressing weeks of manual data engineering into minutes.
Why it matters: The article reports quantified improvements for PPE in public health, food security, and epidemic forecasting, giving a view of the practical limits of geospatial modeling automation.
Google has released GlucoFM, a self-supervised foundation model with a dual-stream design that separates slow glucose trends from short-term fluctuations while preserving timing and missingness information. Across 14 evaluations spanning four cohorts and seven clinical tasks, it improved average PR-AUC by 5.8 percentage points over the best GluFormer variant and achieved the lowest MAE for predicting post-meal glucose response. It was pretrained on 109,066 hours of unlabeled CGM data and performed best at cross-dataset transfer and few-shot adaptation.
Sentence Transformers v6.0 adds a fourth model type, MultiVectorEncoder, supporting ColBERT-style late-interaction retrieval and providing a complete training recipe.
IBM's Granite team has released Granite 4.2, its first family of dense decoder-only reasoning models, in 3B, 8B, and 30B sizes. They were post-trained from the Granite-4.1 base and are all licensed under Apache 2.0.
Why it matters: The article discloses Granite 4.2's architecture, SFT data mix, and multi-stage RL curriculum in detail, offering a look at concrete post-training methods for reasoning models.
