A study details the first human recipients of Precision Neuroscience's minimally invasive brain-computer interface, designed for recording and stimulation without penetrating brain tissue.
Thinking Machines Lab published 'Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL' on its company newsroom. Reinforcement learning is the training approach underpinning the architecture of the company's Tinker product, and the scale of its compute agreements with NVIDIA and Google Cloud has been attributed to RL workloads.
On 27 August 2026 Anthropic opened a research preview of the Model Hardware Standard (MHS), which it describes as "a shared specification for AI agents to safely operate physical devices", to a first group of scientific research labs and advanced manufacturers. MHS lets agents operate multiple instruments in parallel — Anthropic names microscopes, liquid handlers and robotic arms — across tasks ranging from routine drug discovery experiments to laser calibration on a quantum computer. Development began as a collaboration between Anthropic and HHMI Janelia Research Campus. Anthropic states integration work that typically takes a lab or facility weeks or months is reduced to "hours or minutes" via a standardised driver using read/write primitives plus natural-language tags describing device characteristics and enforced safety limits. Agents control hardware through three mechanisms: MCP, the command line interface, and code files (APIs). Critically for the competitive read, MHS is model-agnostic, works with any device exposing a programmable interface, and Anthropic says it is sharing an early version with partners across science, robotics, electronics and manufacturing to build safety evaluations "ahead of making the standard open source". Genentech is a named partner, having implemented MHS to automate the BCA protein assay across a liquid handler, a robotic arm and a plate reader, with the stated goal of building autonomous labs.
Globes reported that Atreides Fund founder Gavin Baker, an investor in Safe Superintelligence, said SSI intends to launch its first language model by early the following week - the first attributable, on-the-record timing signal from inside the cap table after roughly two years in which SSI has shipped no model, API, paper or demo. The report puts total capital raised at about $7bn across offices in Palo Alto and Tel Aviv with over 50 employees, and names Alphabet, Atreides, Yuri Milner, OurCrowd and VentureIsrael alongside the previously disclosed backers. Investors are described as pressing Sutskever to show product progress. Experts quoted speculate the technical differentiator is real-time updating of model weights during reading - attacking the 'knowledge stagnation' limit in current frontier models - though Globes stresses none of this is confirmed by SSI, which remains publicly silent. Treated as a single-source report of intent, not a completed launch.
xAI made Grok Bot available with SuperGrok, Cursor Pro and all Cursor Teams plans, widening access two weeks after launch.
Figure disclosed Index, a Figure-exclusive data-collection pipeline it had been running in stealth for four months to build what it calls the world's largest and most diverse physical dataset. Disclosed metrics: over 264,000 app downloads across 100+ countries, over 44,000 weekly active users, over 16 million videos uploaded by its network of "Creators", uploads running at 30 minutes of video every second (which Figure characterises as 4.9 years of human work uploaded per day), and $15m already paid out to contributors. Figure states it is "on a path to 100x" and is "committed to spend over $1B the next 12 months on data and compute". The argument is that generalisation is a data problem and that the data needed to train a general-purpose robot does not exist on the internet — it has to be sampled from the real world — so Figure is buying that distribution directly rather than scraping it. Why it matters: a stated $1bn+ 12-month data-and-compute commitment is a capital-scale disclosure from a private company, and it reframes Figure's moat as a proprietary data pipeline rather than hardware alone. All figures are company-stated and unaudited.
Figure AI unveils Index, a global dataset collected from real-world environments via an app with 264,000 downloads across 100+ countries, to train general purpose robots.
Thinking Machines Lab announced a Safety Research Grants programme on its company newsroom, its third external grants programme after the Tinker research and teaching grants (Oct 2025) and the Interactivity Research Grants (May 2026).
xAI brought Grok Build beyond the terminal to web and mobile.
xAI released Grok 4.6, which the company frames as building on Grok 4.5 with particular focus on long-running agents and more ambitious interactive and visual work.