plus 9 more stories
Also: PSI reports an AI-found proof of a quantum computing conjecture; and Botpress discloses breach affecting customer bot configurations. Every story links to its primary source.
Curated by Gaurav · November 6, 2026
Thinking Machines introduced Inkling 2 Small, a model with 594 billion total parameters, 17 billion active parameters and image input support. It is available on Tinker, with Apache 2.0 weights planned for the following week after external safety testing; the initial announcement was November 5.
Physical Superintelligence says its research system, using a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture, open since 2008. Lean accepted the formal proof, and researchers are reviewing the result, which concerns limits on quantum speedups for unstructured inputs.
Botpress disclosed that an attacker used an internal authentication flaw and an employee API credential to read bot configurations in 16 customer workspaces. Those configurations can contain third-party secrets; the company contacted affected customers, closed identified access routes and is continuing credential rotation.
TechCrunch reports that Palo Alto Networks is in talks to buy Clutch Security for $500 million in cash plus milestone-based payments. Clutch manages AI agents' access to company systems; the transaction remains under discussion, and neither company responded to the publication's request for comment.
Meta introduced spaces inside Muse, letting people describe and refine reusable experiences such as planners, gift trackers and small games. Users can arrange spaces in the app and connect other services for background updates; the feature is available wherever Muse is offered.
Amazon SageMaker Inference now supports NVIDIA Multi-Instance GPU partitioning, allowing different models to use hardware-isolated portions of one accelerator. AWS says the capability works through inference components, helping fit smaller models to available compute without running a separate Kubernetes platform.
MIT highlighted CheMeleon, an open-source model developed with BASF researchers to predict molecular properties when experimental data are scarce. It learns from descriptors of about a million molecules before task-specific fine-tuning; the November 6 article describes research published earlier in 2026.
Brookfield reported $25 billion of third-quarter fundraising and outlined its global AI factory project pipeline. The fundraising total spans its businesses, including credit and infrastructure, and should not be read as an AI-only raise; combined AI infrastructure and infrastructure debt commitments reached $10 billion over the year.
Baseten made Serverless Loops generally available for supervised fine-tuning and reinforcement learning on supported open models. Its managed service bills serverless training and sampling by token, while users control their training signal and can move selected checkpoints into production inference.
Ai2 released ToyBlocks, an open robot simulation and the first environment in its FoundationWorlds framework. The tabletop block world generates training demonstrations and repeatable evaluations, helping researchers distinguish planning errors from motion failures and test how robots handle unfamiliar conditions.
Tools in this briefing, reviewed in our directory:
Track the tools behind the news:
→ AI Tool Leaderboard → Browse the directory