All articles
Galbot prices ET1: three humanoid tiers from 79,000 yuan
Galbot has priced its ET1 biped at 79,000, 139,000 and 179,000 yuan. The tiers differ in compute, not looks — interaction-based learning is reserved for the two pricier builds.
NASA hands route planning to AI. Perseverance drove 456 m planner-free
JPL says Perseverance covered 210 and 246 metres on 8 and 10 December 2025 along a route set by Claude models. The rover's digital twin verified the commands across more than 500,000 telemetry variables.
VISTA: Claude Opus 5.0 clears 25 ARC-AGI-3 games faster than people
The VISTA paper went up on arXiv on 1 October 2026. Lossless visual memory lifted Claude Opus 5.0's Relative Human Action Efficiency from 40.68 to 100.00 without touching the weights.
SynthID Bio: a watermark inside a protein that keeps it working
Google DeepMind presented SynthID Bio on 30 September 2026, watermarking protein sequences and predicted 3D structures. For folding, the signal is sewn straight into AlphaFold 3's diffusion network weights.
Gemini 4 Argon: a million output tokens and 77.9% on DeepSWE
Google DeepMind introduced Gemini 4 Argon on 30 September 2026. The generation limit rises from 64,000 to one million tokens, and the model reaches cyber defenders before general developers.
GO-3 Preview: AgiBot robots keep learning on the job
AgiBot unveiled GO-3 Preview, an embodied AI model that learns across three stages: pre-training on human video, post-training on a fleet of over a hundred robots, and after deployment while working in a store. First place on RoboDojo, RoboTwin and RoboColiseum per conference figures.
Barclays targets Claude Code for half its developers by end of 2026
Anthropic announced on 1 October 2026 that Barclays is extending its Claude deployment. Claude Code should reach 50% of the bank's developers by end of 2026, while its knowledge assistant has handled over a million searches.
OpenAI: a billion Einsteins still need someone to mine stone
In "The Eternal Complement", published 1 October 2026, OpenAI argues that execution, not ideation, is the bottleneck on progress. The more hypotheses a model generates, the costlier verification becomes.
Chesky: agents need their own OS, not a layer on top of iOS
Airbnb's chief executive rejects the in-chat app store model and wants AI at kernel level, an SDK for agents and components that compose.
Opus 5.5 dropped the em-dashes. 2,548 tells remain
Graphite analysed 9,974 topics across ten models. Opus 5.5 cut em-dash use by 99%, yet its count of overused phrases fell only 4%.
LLMs don't reason: Graepel wants an explicit epistemic state
Thore Graepel of UCL, formerly on the AlphaGo team, argues language models do not reason because they lack a persistent epistemic state separate from knowledge.
Suncatcher: 1,800 Starship launches before orbital TPUs pay off
Google launched its first TPU satellite and published an analysis showing orbital data centres would need 1,800 Starship launches in a decade.