Viewpoints, commentary, analysis, interviews, and discussion of trends.
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.
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
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.
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
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.
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.
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.
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.
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.
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.
Ben Affleck went viral this week for several interview clips. At GQ and Bloomberg Screentime 2026, he explained concepts including machine learning, neural networks, transformers, tensors, GPUs, and inference, and said he can write Python.
In an interview, Periodic Labs' Liam Fedus and Ekin Dogus Cubuk proposed a "synthesis superintelligence" approach: reinforcement learning in physical experiments, AI-driven materials characterization, density functional theory (DFT), and high-throughput labs. The aim is for models to learn the full process of doing science, rather than only studying published results.
ONTIME Mobility COO Han Feng believes that as world model capabilities grow stronger, the importance of real data will not diminish: synthetic data is responsible for expanding known long-tail scenarios, while real data is responsible for discovering unknown problems and providing calibration and validation. ONTIME Mobility does not train large models; instead, it collects real data through routine mobility services, Robotaxi, and offline operations and maintenance networks, and delivers training materials to autonomous driving, world model, and embodied intelligence companies through a data closed loop.
Based on a survey of 300 data and AI executives, MIT Technology Review found that only about 34% of enterprise agentic AI projects reach production on average. A lack of knowledge is one of the main reasons they fail.
The focus of enterprise AI competition has shifted from whether predictive models can outperform statistical forecasts to how prediction systems can make autonomous decisions without straying from business intent. Deep learning and generative AI enable real-time training, allowing AI to keep evolving rather than wait for quarterly refreshes, and expand its data sources from structured numerical records to unstructured interaction data. Vishal Gupta, a partner at Everest Group, says the word 'analytics' is giving way to AI.
Of the 45 layers in GLM-5.3-Flash, 34 use Kimi's KDA linear-attention approach and 11 use DeepSeek's KPool-DSA sparse attention. Linear and Sparse attention appear together in one model for the first time. The combination reflects a shift away from finding a single best mechanism toward assigning different roles: Linear maintains long histories at low cost, while Sparse retrieves specific information as needed.
ECCV 2026 received 10,473 submissions and accepted 2,834. The 3D topic fell from a leading area to third place, but questions around 3D representation, geometry, and physical consistency are spreading into video generation, embodied AI, and world models. Of 93 workshops, about 10 were directly related to 3D, 11 to spatial intelligence, nine to physical AI, and five to world models. Best-paper and honorable-mention awards also focused on 3D surface representation, object reconstruction, and geometry perception.
Language models have greatly sped up coding and data analysis, but the bottleneck in scientific research remains physical experiments that can take days or weeks to validate: “thinking got cheaper; doing didn't.” The biotech industry is splitting into two approaches. Foundries use next-generation sequencing, multiplex assays, and lab automation to produce experimental data an order of magnitude faster; Navigators embed AI in day-to-day company workflows to accelerate decisions.
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.
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
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.
Jim Fan attributes GEN-1.5’s promise to naturally repetitive human demonstrations, retained failures that teach recovery, and UMI, which lets humans collect data directly while wearing a robot gripper. He argues that teleoperation loses physical intuition and that sufficient data can enable zero-shot behaviors such as picking up novel objects without finetuning. He remains cautiously optimistic, noting that in-context learning depends on the distance between training and test conditions and that current demos are too simple to draw conclusions.
Hugging Face has released a report on the open-model ecosystem from January through August 2026, drawing six findings from Hub data. Public model repositories grew from 2.43 million to 2.96 million and datasets from 711,000 to one million, but 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of downloads.
Why it matters: Hugging Face analyzes its Hub data on the open-model ecosystem; findings on downloads versus likes and the scale of Qwen derivatives can inform model selection.
A Hugging Face blog argues that GPU utilization, not model intelligence, is AI's next real constraint. GPU costs accrue by calendar hours, but output is produced only during compute hours; after buying GPUs, companies face the challenge of keeping their clusters busy. Using aviation as an analogy, the article says even fully loaded clusters may waste much of their potential because GPUs run all day while demand does not.
Berkeley's EPIC Data Lab argues that as inference costs have fallen from around $30 per million tokens for GPT-4-level models in early 2023 to under $1 today—and below $0.10 at some vendors—data systems face three new challenges: systems designed for agents, systems operated by agents, and systems built by agents.
Import AI 457 focuses on three studies. SentinelOne analyzed fast16.sys, a virus from more than 20 years ago that corrupted results from high-precision software such as LS-DYNA 970, PKPM, and MOHID by tampering with in-memory calculations, sabotaging scientific and engineering work.
In Import AI 455, Jack Clark estimates a greater than 60% chance that by the end of 2028 AI R&D will run without human involvement, with frontier models autonomously training their own successors. He puts the chance of this happening by 2027 at about 30%.
