AI Briefing

← Archive

OpenAI safety firings dispute and a wave of new model releases

Oct 10 – Oct 11 ET

Three fired OpenAI safety researchers are disputing the company's misconduct claims, while a circulating clip of Sam Altman on open-source cyber risk drew a heated reaction. The weekend also brought several new model and product releases, including Nace AI's Drex 1.5 and Odyssey-3.

Breaking

3 stories
  1. 01
    The fight over OpenAI's three fired safety researchers is still going. — Jasmine Wang, Tomek Korbak and Mikita Balesni say they were pushed out for raising safety concerns and working with outside evaluator METR on the Hugging Face incident. OpenAI denies retaliation and says they mishandled sensitive information. TechCrunch NPR
  2. 02
    Sam Altman says society should accept some severe cyber incidents from open-source models. — In a circulating clip he framed that as the price of keeping them free, warned a "tidal wave of cybersecurity problems" is coming, and the reaction has been heated. X News
  3. 03
    David Sacks is attacking Anthropic's approach of giving Claude its own moral code. — He's contrasting it with Nadella's "emergency brake" approach, as Anthropic's new usage policy banning sustained cruelty toward Claude takes effect Nov 12. X News X News on policy

Models & Releases

7 stories
  1. 04
    Nace AI open-sourced Drex 1.5. — A 9B "decision model" that returns probabilities for typed choices in one pass. It now leads the Decision Index (58.28) and is on Hugging Face and OpenRouter, another rival to TypeSafe's Jev and Microsoft-Decision-1. Weights use a restrictive Open RAIL-M license. Hugging Face X News
  2. 05
    Odyssey-3 world model goes public. — Free research preview (Oct 11) that generates interactive environments from text in real time; the company says it tops physics benchmarks. API access is by application. The Decoder
  3. 06
    Anthropic opened Claude Code Projects to all Pro and Max users on the waitlist. — Persistent workspaces that split a goal into parallel cloud threads. Separately, multi-agent workflows (up to 1,000 coordinated agents) are in public beta for Managed Agents. X News X News on multi-agent
  4. 07
    OpenAI's Codex and ChatGPT Work are on a 28-day run of daily improvements or full usage resets. — Oct 5 through Nov 1; Codex can now also use a banked reset when you ask it to. OpenAI Help Center TestingCatalog
  5. 08
    HeyGen Voice debuted at #1 on Artificial Analysis's voice leaderboard. — HeyGen's first in-house text-to-speech model, ahead of ElevenLabs and Qwen. X News
  6. 09
    Alibaba released Qwen-Image-2.1-Turbo. — A 7B image model that generates at 2048px in 8 steps instead of 40. Non-commercial use only. X News
  7. 10
    Cognition's Devin now lets you plug in a ChatGPT subscription. — GPT usage counts against your ChatGPT plan limits instead of Devin credits. X News

New Tech to Watch

3 stories
  1. 11
    vLLM now supports Nvidia's Vera Rubin. — Early benchmarks show more than 7.8x the throughput of GB200 NVL72 when serving MiniMax M3. X News
  2. 12
    Qualcomm CEO Cristiano Amon on phone-scale frontier models. — Says frontier AI firms want phones that can run 100B-parameter models nonstop by 2028, up from about 30B today, and that some AI companies are building their own phones. X News
  3. 13
    Amazon paper: an LLM as a "research world model." — Predicts whether a training change will help before any GPU time is spent on it. Rohan Paul on X

Markets & Funding

5 stories
  1. 14
    Update on Nvidia and Reflection AI. — Bloomberg says the options also include a compute deal, not just an acquisition or a bigger stake. Reflection launched its open-weight Beam model on Monday. Would follow Nvidia's $13B Hugging Face purchase and its roughly $20B Groq licensing pact. Business Times/Bloomberg Reuters via StreetInsider
  2. 15
    Hone came out of stealth with a $60M seed at a $285M valuation. — Led by Benchmark and Index; building persistent "Engines" that own whole business outcomes. Founderland
  3. 16
    Robotics data startup Mecka raised a $60M Series B led by Sequoia. — Nvidia, M12, Qualcomm and Samsung joining, at a reported $500M valuation. The FT puts 2026 robotics and physical-AI funding near $48B. The American Newsroom
  4. 17
    Jefferies' Chris Wood warns US AI capex is heading for "massive capital destruction." — As cheaper Chinese open models gain share; credit default insurance costs for AI borrowers are rising too. X News
  5. 18
    Next big catalyst: Mag7 earnings and Nvidia. — Microsoft, Alphabet and Meta report around Oct 28, and Nvidia reports Nov 17. AI capex and cloud growth are the numbers to watch. Taylored Finance

Leaks & Rumors

4 stories
  1. 19
    Anthropic appears to be building a new voice mode. — TestingCatalog found a hidden Claude "Voice (Preview)" tab, a new option to share voice data for training, and new codenames "Arborio" (likely Claude Health on the web) and "Iggy." TestingCatalog TestingCatalog on codenames
  2. 20
    An unannounced OpenAI model, gpt-rosalind-discovery, showed up on the API pricing page. — Under life sciences, priced like gpt-rosalind-research. Limited to OpenAI's trusted-access program; no launch post yet. X post
  3. 21
    Gemini 4 Argon now shows low, medium and high reasoning settings inside Google's Antigravity coding tool. — Suggests the public release is getting close. TestingCatalog
  4. 22
    Rumors say the big labs are close to recursive self-improvement. — Including claims that Anthropic started a larger pretraining run ("Fable 6"). Unverified; skeptics note human oversight is still required. X News

Politics & Regulation

2 stories
  1. 23
    The Trump administration is leaning on legal liability rather than new AI rules. — Bessent and Sacks back it, and Trump says the Justice Department should step in if models get out of hand. The White House task force's "incident reporting is not optional" warning has no published legal instrument or penalties yet. Bloomberg via Japan Times DeAI
  2. 24
    New bills during the congressional recess. — Senators Peters and Lankford would require agencies to tell people when AI makes critical decisions about them, and House Democrats reintroduced a bill for a CISA AI task force. Nextgov