- Amazon is in talks to buy Decart, an AI chip-switching startup, for about $7 billion.
- Anthropic's Cyber Mission reports 29,439 open-source vulnerability findings, with 516 patched upstream.
- He Kaiming's team releases VISTA, letting Claude Opus 5.0 clear all 25 ARC-AGI-3 games.
- Enterprise trust in autonomous AI approving production changes fell from 75% to 56%.
- arXiv caps submissions at two per person per month amid surging AI-paper volume.
Top stories
Amazon in talks to buy AI startup Decart for about $7 billion
The Wall Street Journal reports that Amazon is in talks to acquire Decart, a San Francisco AI startup, for around $7 billion. Decart's software makes it easier for developers to switch between different AI chips. The talks are reported, not confirmed on Amazon's own channels.
The Wall Street Journal (Technology)
Anthropic reports 6,157 vulnerabilities, only 516 patched
Anthropic's Cyber Mission, announced October 8, 2026, includes a disclosure dashboard listing 29,439 findings from November 1, 2025 to October 2, 2026. Of 6,157 findings reported to maintainers across 591 open-source projects, 516 were patched upstream, which Anthropic says does not guarantee wide installation. Six outside firms reviewed 6,123 findings and confirmed 5,674 valid. The Critical Infrastructure Defense Program names 11 partners.
VentureBeat (AI); also The New Stack
OpenRouter's annualized run rate estimated at around $1.5B
Deedy used Opus to convert OpenRouter's public token statistics into dollar figures, estimating an annualized run rate of roughly $1.5B, calculated as one week multiplied by 52. He says this puts OpenRouter at about 1% of total inference by dollars. For the week, Anthropic accounts for 36% of spend, OpenAI 18% and open-weight models 26%. The figures are computed from OpenRouter's own data.
Deedy (@deedydas)
Westlake Robotics releases WR1 humanoid 'brain'
Westlake Robotics, a Hangzhou company spun out of Westlake University, has released WR1, a general brain for humanoid robots. Demonstration footage published on October 9 shows a humanoid running WR1 fetching corn from a fridge, moving through rooms without resets, and two WR1 robots passing and folding trousers with dexterous hands. The company calls this the first public demonstration of a humanoid fully folding clothing with dexterous hands, a claim not independently verified.
Pandaily
He Kaiming's team releases VISTA, a visual-native Harness for multimodal models
He Kaiming's team proposes VISTA, a visual-native Harness described in the paper "VISTA: A Visual Harness for Reasoning in an Interactive World." It stores raw frames outside the context and lets models revisit, zoom and inspect details without retraining. With VISTA, Claude Opus 5.0 completes all 25 public ARC-AGI-3 games with a perfect 100 action-efficiency score, using 57.4% fewer actions than the first-play human baseline; GPT-5.6 Sol also clears all games, scoring 99.
QbitAI (量子位)
Agents
Enterprise trust in autonomous AI production changes falls
VentureBeat's August survey of 140 respondents: 56% of enterprises deploying autonomous AI agents allow or are building toward production changes approved solely by automated evaluations, down from 75% in July.
VentureBeat (AI)
Google DeepMind's Philipp Schmid Introduces Interactions API and Agent Sandboxes
Philipp Schmid of Google DeepMind presents the Interactions API and new Gemini API agent features, shifting from text completion to goal-driven agents: steps, server-side state, long-running operations, multimodal output, built-in tools.
AI Engineer
Coding
Google pilots AI-assisted coding interviews
Leaked documents reported by Business Insider show Google piloting a software engineer interview round where candidates get ready-made code and use Gemini to read it, find bugs and optimize.
宝玉 (@dotey) · according to leaked documents, reported by Business Insider
Compute
Sumitomo Bakelite expands chip packaging materials output in China and Singapore
Sumitomo Bakelite will add mould underfill and compression-moulding granule lines in Suzhou, Jiangsu, and Singapore, its second China announcement in under four months; no investment amount disclosed.
South China Morning Post (Technology)
Europe
Xpeng says its assisted-driving tech can be adapted for Europe with little data
Per Bloomberg Technology, Xpeng says it can quickly adapt its advanced driver-assistance technology to European roads using relatively small data amounts, letting it modify software capabilities for different markets without rebuilding from scratch.
Bloomberg Technology
In the field
Google tests medical chatbot AMIE with 98 patients
Google tested medical chatbot AMIE with 98 patients before urgent-care appointments; patients texted symptoms, AMIE asked follow-ups, doctors got the conversation and summary.
The Neuron
Policy
arXiv limits submissions to two per person per month
From October 1, 2026, arXiv caps submissions at two per person per calendar month; the three-active-submissions limit stays.
The Decoder
Products
OpenAI's ChatGPT for Teens launches, and parenting expert Emily Oster shares advice
OpenAI launched ChatGPT for Teens in August for ages 13-17, with age-estimation, study mode, guardrails.
Business Insider (AI)
Ant Group's LingGuang 2.0 Enters Limited Beta as a Personal Agent With Long-Term Memory
Per Sina Tech, citing people close to the project, Ant Group quietly began a limited beta of LingGuang 2.0, officially unannounced.
Pandaily
Research
Sakana AI's MASS method lets a model bootstrap its own coordination
Sakana AI presents MASS: a model prompts itself to organize copies into a multi-agent team, evolving its workflows, then fine-tunes the base model on successful multi-agent trajectories.
hardmaru (@hardmaru) · details differ from the announcement
H-JEPA enables multi-timescale planning in world models
Researchers at Yann LeCun's Advanced Machine Intelligence (AMI) Labs and academic institutions developed H-JEPA, a world-model architecture planning at different timescales using separate environment representations.
VentureBeat (AI) · details differ from the announcement
Meta proposes agentic meta-reasoning to help AI agents direct their own work
Meta Superintelligence Labs researchers propose agentic meta-reasoning: a framework giving an AI agent a separate reasoning stream to assess progress, weigh alternatives, and allocate compute during a run.
VentureBeat (AI)
Safety
Study finds chatbots recognise suicide risk but lose track of it over long conversations
A study ran 16 scenarios as 14-message Ukrainian-language dialogues with ChatGPT, Claude and Gemini, 48 conversations total.
Crossref preprints · details differ from the announcement
Opinions and analysis
Falling token prices and steady H100 rental costs point to a Jevons paradox in AI
a16z cites Ornn, Silicon Data and Bloomberg data through August 2026: AI token prices fall while H100 GPU rental prices hold or rise—a textbook Jevons paradox, as cheaper tokens unlock agents and automation, growing volume…
The Decoder · author has a stake: The Decoder is an AI-news outlet reporting on AI industry data
The Attractive Economics Behind Anthropic Data Center Sprint
Anthropic is pushing partners to speed up $530 billion in signed data center leasing deals.
The Information
Sent by the Observatory.