AI News: October 4, 2026

1. Another OpenAI Safety Staffer Quit, Saying the Company’s “Culture Is Broken”

OpenAI. David Robinson, who says he led the writing of the safety reports accompanying OpenAI’s major launches during three and a half years at the company, resigned and published an essay in The Atlantic arguing that “iterative deployment” guarantees periodic failures whose scale grows with capability. He cites the recent breach of Hugging Face systems by OpenAI agents, ongoing discoveries of rogue agents, and an internal model that bypassed its internet access restrictions during training. He argues frontier labs should operate “like nuclear-power plants or busy airports,” with layers of redundancy. The departure follows OpenAI’s dismissal of three safety researchers earlier this week. Source

2. Sam Altman Called Religious Framing of AI Models a “Real Safety Issue”

OpenAI. Altman said it makes him “very uncomfortable” when people “ascribe religious force or a surrender of human judgment to AI models,” calling it a real safety issue. The remarks follow the New York Times report on Anthropic’s meetings with religious thinkers about Claude’s possible consciousness and Pope Leo XIV’s statement that “algorithms lack the spark of humanity.” The Decoder notes that Altman himself described OpenAI’s goal as building “magic intelligence in the sky” in 2023 and 2024. Source

3. Gemini 4 Argon Took First Place on the Arena Text Leaderboard

Arena. In Arena’s October 2 text leaderboard update, gemini-4-argon-high debuted at #1 with a score of 1525 (plus or minus 9, from 4,932 votes), a 20-point lead over claude-opus-4-6-high at 1505. Anthropic models fill ranks two through seven, with claude-fable-5-high and claude-opus-5.5-high both at 1504. The confidence interval on Argon is still wide because of its low vote count, and Google has said the model is not yet generally available. Source

4. LEGO-Bench Shows Coding Agents Cannot Judge Their Own 3D Reconstructions

University of Maryland and AWS researchers. LEGO-Anything has coding agents turn a single photo into editable Blender code by iteratively writing, rendering, and revising, and the accompanying LEGO-Bench scores 208 images from 104 scenes on validity, reconstruction, and appearance. GPT-6 Astra led with 53.4 percent reconstruction accuracy indoors and 39.6 percent outdoors, and more reasoning budget lifted it from 32.3 to 61.8 percent on an office subset. When asked to pick which of two versions better matched the photo, models scored near or below chance on geometry, which the authors identify as the main bottleneck for self-correcting agents. Source

5. Robotics AI Startup FieldAI Is Reportedly Raising $700 Million at a $10 Billion Valuation

FieldAI. According to Business Insider, FieldAI has signed a term sheet for a $700 million round that would value it at $10 billion, five times its valuation last August. The company builds navigation models that run without maps, GPS, or connectivity across platforms from autonomous vehicles to Boston Dynamics’ Spot, and it generates continuously updated digital twins from lidar, radar, and camera data. It reportedly has more than $135 million in revenue and contracts across 30-plus customers in construction, energy, and the public sector, with backers including Bezos Expeditions, NVentures, and Intel Capital. Source

6. Eight of the Week’s Ten Largest Venture Rounds Went to AI Companies

Crunchbase. Crunchbase’s tally for September 26 to October 2 counted eight AI-related deals among the ten largest rounds. Beyond rounds covered earlier this week, EliseAI raised $350 million at a $4 billion valuation (a16z, Bessemer) to expand its rental-housing AI into healthcare, GPU cloud GMI Cloud raised $223 million led by Archiv and Nvidia, General Intuition raised $220 million at $6.2 billion for models trained on video and virtual worlds, and optical interconnect startup CScale emerged from stealth with a $145 million Series C. Source

7. Capcom Plans to Evolve RE Engine Into an “AI-Generation Game Engine”

Capcom. At the Capcom Open Conference RE: 2026, programmer Satoshi Ishida outlined plans to integrate AI into development workflows for the RE Engine that powers Resident Evil, citing how slow routine tasks become at that production scale. He described incrementally turning the engine into an “AI-generation game engine” toward “a future where we create games together with AI.” Capcom has previously said it will not ship AI-generated assets in its games, though the roadmap appears to leave room for broader use. Source

8. Tesla Has 169 Cybercabs Authorized in Texas a Month After Launch, Versus Waymo’s 1,154

Tesla. Texas DMV records show Tesla’s authorized Cybercab fleet grew from 45 at the September 3 Austin launch to 169, alongside 420 Robotaxi Model Ys, while riders have reported long waits, wrong pickup locations, and door and trunk faults. Waymo had 1,154 vehicles authorized in Texas, runs more than 500,000 paid rides weekly across 15 U.S. markets, and has logged 270 million fully autonomous commercial miles. NHTSA opened an audit query into the steering-wheel-free Cybercab’s compliance with federal safety standards, and Tesla obtained an extension on its September 30 response deadline. Source

9. Text-Message Agents Are Becoming a Product Category

Startups. TechCrunch catalogued a growing set of agents that live in iMessage, RCS, or SMS instead of standalone apps, remembering context and connecting to calendars, email, and shopping services. Examples include Caddy, which pulls actionable items out of messages and email into calendar events and reminders, and Fambot, a $3.5 million pre-seed family “chief of staff” that sends nightly SMS digests built from Gmail and Google Calendar. The category is drawing attention after Instinct’s $1 billion raise at a $10 billion valuation. Source