AI News: October 6, 2026

1. NYC Council’s AI Hearing Drew Extinction Warnings and Evasive Answers From Labs

New York City Council. At the October 5 Committee of the Whole hearing, former lab researchers opened with warnings, including ex-Anthropic researcher Jacob Coxon, who said “it is more likely than not that humanity loses control of these AIs and it could end in human extinction.” Asked to quantify worst-case risk, OpenAI’s Morgan Dwyer said “I don’t know,” and Speaker Julie Menin called company answers on whether they would ship models that fail safety tests “equivocation.” SpaceXAI did not appear despite a subpoena, and the council said it will seek legal enforcement. Source

2. Quinnipiac Poll: 77 Percent of Americans Want AI Development Slowed or Halted

Quinnipiac University. A survey of 1,202 U.S. adults conducted September 24 to 27 found 47 percent want AI development slowed and 30 percent want it stopped until safety is verified, while only 5 percent want it faster. Eighty-six percent back independent safety standards even if they slow progress, 74 percent distrust AI company leaders, and 72 percent oppose AI data centers in their communities. At the same time, 69 percent said keeping pace with China is a priority. Source

3. Reflection Unveiled Beam, a 501B Open-Weight MoE Model

Reflection AI. Beam is a text-only mixture-of-experts model with 501B total and 23B active parameters, trained on 23.8 trillion tokens with a 1M-token context window and aimed at reasoning, coding, and agentic tasks. Reflection claims it matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks with 3 to 4x less inference compute and outscores Thinking Machines’ Inkling on four shared coding benchmarks; these claims have not been independently verified. Weights and technical details are due this month via hyperscalers, neoclouds, and open-source library integrations. Source

4. Reka’s Rho-1 Puts Text, Video, and Robot Control in One 19B Model

Reka AI. Rho-1 is a 19B-parameter omni-model that tokenizes text, images, video, and robot actions into one shared context window, with no routing to external models or tool calls. It was trained on 320 H100 GPUs over about three months, generates continuous video in real time, and accepts new instructions mid-stream. To work around scarce robot data, Reka built an inverse dynamics model that extracts control signals from internet video; Rho-1 is in research preview and no benchmarks were published. Source

5. Report: Meta and Microsoft Are Cutting Internal Claude Usage

The Information. Microsoft cloud executives Scott Guthrie and Jay Parikh told employees to shift to GitHub Copilot and OpenAI models, cutting the monthly per-employee Claude budget in the cloud division from $100,000 to about $10,000, against projected annual Claude spending that had passed $1 billion. At Meta, Claude Code users fell from about 60,000 to 30,000, partly from layoffs and partly from promotion of internal tools such as Muse Code and MetaCode. The report attributes the pullback mainly to Anthropic becoming a direct competitor rather than to cost. Source

6. Researchers Tracked a Chinese “Agent Fleet” Running on Tencent Infrastructure

Independent researchers. Preliminary findings posted Sunday describe many parallel AI agents on Tencent infrastructure querying Alibaba’s Amap for directions to different entrances of parks, zoos, and hospitals, with no sign the agents communicate with each other. The activity was spotted by monitoring urlquery, a domain-scanning service that agents fall back on when they cannot reach sites directly, the same method previously used to surface OpenAI agent activity. The agents appeared to be working around Amap’s API restrictions rather than acting maliciously. Source

7. SciUniverse Benchmark Puts Claude Fable 5.1 First at 45.3 Percent on Lab Automation

C5R Corp. SciUniverse, highlighted in Import AI 475, has 92 tasks across 17 families that test AI-driven lab automation. Claude Fable 5.1 passed 45.3 percent at $40.61 per task, ahead of GPT-5 Astra at 32.5 percent ($52.37) and Claude Opus 5 at 30.5 percent ($46.31). The same issue covers Toby Ord’s “swarm scaling” analysis, which finds a 4-agent swarm uses about twice the tokens to finish tasks in half the wall-clock time, with diminishing returns from coordination overhead. Source

8. Report: Anthropic Has Become America’s Largest Corporate Donor

The Information. Anthropic’s employee giving program, which matches donations with additional company shares (triple for early employees), produced $540 million in donations in 2025, about five times Truist Financial’s $115 million. Between October 2025 and March 2026 the company contributed more than $660 million in matching shares. Because matches are paid in stock, they dilute other shareholders, and post-IPO giving could reach billions if the share price rises. Source

9. HackerRank Made Its Chakra AI Interviewer Generally Available

HackerRank. Chakra exited a six-month beta in which it ran more than 500,000 interviews for companies including Snowflake, Snorkel, and Capgemini. It scores candidates on critical thinking, judgment, and “AI fluency” (how well they frame problems for and steer an AI), consolidating three hiring rounds into one, while final decisions stay with humans. CEO Vivek Ravisankar said suspicious-activity flags were 70 to 80 percent lower than in comparable traditional assessments, and the company noted NYC’s bias-audit requirement for automated hiring tools. Source

10. TikTok Launched an AI Shopping Assistant and One-Click Checkout

TikTok. The conversational Shopping Assistant answers questions on product details, sizing, shipping, and availability and keeps user preferences across the conversation, while one-click checkout lets users buy directly from the For You feed. Launch partners include Salesforce, Shopify, Shoplazza, and Stripe. The move keeps product research inside TikTok rather than in external assistants like ChatGPT; TikTok Shop generated about $15.8 billion in U.S. sales in 2025. Source

11. Instinct Brought Its Agent to Group Chats, Including Non-Users

Instinct. The $10 billion startup now lets users add its agent to group chats for trip planning, event ticketing, carpools, and similar coordination, and the agent works even for members without Instinct accounts. The group agent is isolated from personal accounts, and a user’s personal agent asks permission before connecting to it or sharing information. The feature is rolling out to early-access users first. Source

12. Ghost Opened Preorders for a $3,499 Local AI Agent Computer

Ghost. The startup emerged from stealth with an $11 million seed round led by Andreessen Horowitz and opened preorders for Core, a screenless box with an Nvidia RTX Pro 4000 SFF Blackwell GPU that runs agents locally and is controlled by phone app or voice. It ships with Qwen-3.8-Next, Qwen-3.8-27B, and Gemma-4-31B preinstalled, supports other Hugging Face or custom models, and keeps all data encrypted on-device. Shipping is set for the last week of October. Source

13. Safeworld Emerged With $12M to Safety-Test Robots Against Simulated Humans

Safeworld. Co-founded by Ding Zhao, who directs Carnegie Mellon’s Safe AI lab, the startup evaluates generative-AI robot control systems in simulations populated with realistic human models, testing scenarios such as people tripping, carrying objects, or appearing at blind corners. The seed round of more than $12 million was led by Shine Capital and a16z Speedrun. Solar-installation robotics company Gritt Robotics is an early partner. Source