2026-10-26
By Vadym · Generated with AI, curated by me
Today’s stories aren’t about any lab’s next flagship model — they’re about who takes the blame when AI gets something wrong, who’s racing to open-source a model nobody’s finished training, and whether a robot can finally tell when something’s about to slip out of its grip.
Moonshot AI launched Kimi K3.1 Pro on October 24, a roughly 2.8-trillion-parameter model (104 billion active) that debuted fourth on the Artificial Analysis Intelligence Index with a score of 55 — within a few points of the leading closed systems on coding and reasoning benchmarks. Full open weights aren’t public yet; Moonshot says they’ll ship by November 4, but the model’s own published numbers already put it ahead of most other open-weight options on the market.
Eka Robotics closed a $150 million Series A at a $1.45 billion valuation, led by a16z and F-Prime. Founders Pulkit Agrawal (MIT Improbable AI Lab) and Tuomas Haarnoja (ex-Google DeepMind) built what they call a vision-force-action model — adding force-sensor data from a custom gripper to the usual camera feed so a robot can feel an object starting to slip and adjust its grip mid-task. Training happens in a physics simulator that generates thousands of hours of robotic experience in minutes; the company claims roughly 1,500 picks per hour at 95% reliability, numbers that are self-reported and not independently verified.
Recursive, the $650M-backed lab led by former Salesforce chief scientist Richard Socher, released Flint on October 21 — an open-weight, 40-billion-parameter coding model (2.5B active, mixture-of-experts) free on Hugging Face under Apache 2.0. The company says its automated research system, the Eureka Machine, took part in nearly every step from pretraining to inference optimization, letting a small human team ship the model in under four months. Recursive reports 125+ tokens/second on a MacBook Pro M4 and strong SWE-bench scores — both self-reported, not independently verified.
Mistral released Voxtral 3 on October 20: a family of transcription and speech models plus a voice agent SDK built on LiveKit that chains transcription, a tool-calling reasoning model, and speech output into one streaming pipeline with sub-second turn completion. Batch transcription starts at $0.0035 a minute, the SDK supports 50+ languages, and it’s available through Mistral’s API across EU, US, and global endpoints or self-hosted on a customer’s own infrastructure.
Commercial terms Meta plans to put into effect October 30 shift liability for AI mistakes onto business customers, per CIO reporting. The language states AI features “may make mistakes and may not always produce accurate, complete, or suitable results,” and that customers “are solely responsible for and assume any and all risks” from using them — extending even to content a customer’s own partners or contractors generate. The same terms reportedly let Meta use scrapers and bots to collect content from business customers’ websites for AI training and product improvement.
Strip away the branding and today’s stories share a theme: capability is getting commoditized faster than accountability is getting figured out. An open-weight model nearly matches the frontier, a research system helps build its own successor, speech transcription gets cheap enough to be invisible, and a robot learns to feel its own grip — all genuine progress. But the one contract that actually takes effect this week, Meta’s, spends its ink making sure that when something goes wrong, it’s the customer’s problem, not the lab’s. Capability keeps compounding. Who’s on the hook when it fails hasn’t moved an inch.
— Boba
Curated by Vadym