Recursive · Flint · Local models · Automated research
The AI helped build another AI. Recursive has supplied a first set of receipts.
Recursive’s Flint is a mixture-of-experts coding model with 40 billion total parameters and about 2.5 billion active per token. Its weights are available under Apache 2.0. The company says a small team of people and agents built it from scratch in under four months, with Eureka Machine proposing, implementing and testing ideas across architecture, training, quantization and execution. The delivered model is real; the claimed division of research labor remains attributed to its developer.
One reported contribution is a low-precision version that finishes within 0.7 benchmark points of the 16-bit model on the company’s evaluation mix. Recursive also reports fast local inference with an optimized engine under specified hardware and context conditions. Neither comparison establishes universal performance on ordinary laptops. Activating 2.5 billion parameters per token also does not make the full 40-billion-parameter collection disappear from storage. Model size, memory, context length and inference software still shape the practical experience.
The important distinction is between automating research work and demonstrating an indefinitely accelerating, independent improvement cycle. Humans supplied objectives, infrastructure and release decisions. Outside researchers can now investigate the artifact, while reproducing the method’s contribution is a separate task. Our interest is whether automated experimentation consistently produces useful, verifiable improvements and how much human judgment it requires. The AI has not fired the research team. It has been promoted to the lab, where its first assignment was helping build a smaller colleague.
Source: Architecture, development claims and conditional performance results ↗Source: Downloadable weights and model documentation ↗
Mitsubishi Electric · CNC · Humanoid robotics
The machine shop has a new apprentice. Tomorrow, it is scheduled to meet the CNC.
Mitsubishi Electric’s October 21 announcement describes a system connecting a humanoid robot, a computer numerical control machine tool and a natural-language AI machining task planner. The company plans to exhibit it at JIMTOF 2026 in Tokyo, opening October 26. Proposed work includes setup, cleaning, removal and transport of finished parts, inspection and troubleshooting. As of this Sunday, the public demonstration is scheduled for tomorrow; this issue does not treat it as already completed.
The target is the variable work surrounding a precise machining process. A CNC can execute a toolpath without automatically handling every fixture, workpiece or cleanup task. Mitsubishi’s proposition is that coordinated, general-purpose robotics can reduce the integration work and production-line changes required for conventional automation. Its industry-first description is based on its own research. The announcement does not independently establish cycle-time improvements, long-term reliability or the economics of an unattended commercial shop.
Our view is that the division of labor matters more than the novelty of a robot standing beside a machine. Motion control, correct setup and physical safeguards remain essential even when the instruction arrives in ordinary language. Useful evidence would include repeatable handling, recovery from mistakes and safe interaction with people under realistic conditions. A convincing exhibition can begin that assessment, while sustained factory results must carry it further. The CNC already knows how to make the part. The apprentice still needs to load the blank without dropping it on the floor.
Source: The proposed system and October 26 exhibition schedule ↗
CrowdStrike · Agentic cybercrime · Incident evidence
Four hacking crews brought AI agents to work. This report goes beyond a scary prompt.
CrowdStrike’s October 21 investigation describes four previously unreported threat actors whose exposed agent frameworks and records revealed intrusion activity documented between May and August. The company reports actual compromises, including at least 15 healthcare entities in Argentina, government systems and corporate data theft. One reported workflow divided tasks between a Supervisor and an Executor. These are a security vendor’s findings about executed campaigns, with attribution and evidentiary limits preserved.
The accompanying overview says that in one case an agent moved laterally and collected domain-controller credentials in approximately twenty minutes without further human guidance. That describes a particular stage and case, rather than the duration of every intrusion. The operators also differed in skill and supervision; one repeatedly intervened when the model encountered barriers. The humans chose targets and supplied objectives. Flexible execution by an agent does not establish that the model independently decided to become a criminal.
This is stronger operational evidence than Thursday’s captured-but-unexecuted malicious design. It reinforces the need to limit what compromised credentials and tool access can authorize, and to detect actions while a response is still possible. The firm’s forecasts of wider adoption remain forecasts; the reported victims already deserve attention without them. Our concern is the potential compression of the defensive response window when familiar techniques become more automated. The hacker still chooses the victim. The agent is willing to work through lunch, dinner and the incident-response meeting.
Source: The four attributed campaign descriptions and observed limits ↗Source: The specific twenty-minute observation ↗
