1. Investigation Finds ChatGPT Adds Real New Yorker Cartoonists’ Signatures to AI-Generated Cartoons A cartoon showing Dolly Parton and Tim Curry at heaven’s reception desk spread widely across social media, with one post receiving 25,000 likes.

2. LibreOffice pledges to keep AI out of default installation, while users can still connect local models through extensions LibreOffice, the open-source document editor developed by the nonprofit Document Foundation, says it will not add artificial intelligence to its default software for the foreseeable future.

3. Google Launches 740-Million-Parameter EmbeddingGemma 2 for Offline Multimodal Search on Phones Google has released EmbeddingGemma 2, a lightweight model that converts text, code, images, video and audio into comparable numerical representations called embeddings.

In Brief

• Mistral previews one-trillion-parameter Large 4 Mistral launched a public API preview of Large 4, a multimodal model with one trillion total parameters and 49 billion active parameters. The company plans to release its weights by the end of October after additional security testing.

• Google moves Gemini Flash and Pro behind pricier subscriptions Starting October 9, free Gemini users will be limited to Flash Lite, while the $4.99-per-month AI Plus plan will retain Flash but lose Pro. Gemini Pro and Deep Think will require Google AI Pro or Ultra, priced at $19.99 and $99.99 per month respectively.

• Lambda reportedly seeks up to $4 billion before planned IPO AI cloud provider Lambda is reportedly raising as much as $4 billion at a $14.5 billion pre-money valuation ahead of a possible 2027 IPO. Its reported $50 billion backlog includes a $35 billion commitment from Anthropic.

• Anthropic expands reduced-safeguard access for cybersecurity teams Anthropic expanded its Cyber Verification Program into Defense, Red Team, and Specialized access tiers for verified security professionals. Higher tiers permit more offensive testing under tighter eligibility, monitoring, and authorization requirements.

• Atlassian and OpenAI deepen enterprise AI partnership Atlassian and OpenAI expanded their partnership to connect OpenAI models with enterprise knowledge and workflows. The companies say the integration is intended to help teams plan, build, and deliver work.

• Open-source OpenTPU runs ten language models on an FPGA OpenTPU released an end-to-end AI accelerator project spanning SystemVerilog hardware, an instruction set, simulator, compiler, and host software. Its developers report running ten models with real weights on a Kintex-7 FPGA card, with outputs matching the simulator token for token.

• Commerce companies work on a standard for trusted AI agents Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE, and Decagon are developing an open protocol for agent-to-agent commerce. The effort follows problems with personal AI agents being rejected by retailer policies, human-verification checks, and conventional anti-bot systems.

• Hark releases its computer-using personal assistant Startup Hark broadly released Hark Pro, an assistant that can access connected services and operate websites on a user’s behalf, with free and paid tiers. Hark says it emphasizes privacy and shows users the agent’s web activity in a visible window; it also plans an AI-native device for 2027.

• Musubi opens a 1.7-billion-parameter moderation model Musubi released PolicyLM-1.7B with open weights for applying plain-English content policies to messages in real time. The company says the model produces binary decisions in under 50 milliseconds and can adopt policy changes without retraining.

• Kandinsky 6.0 generates synchronized video and audio Researchers released code and model weights for Kandinsky 6.0 Video, a family of 3-billion- and 29-billion-parameter diffusion models that generate five-second clips with synchronized 44 kHz audio and lip-sync. An accompanying super-resolution model can raise output to 1080p.

• OpenAI publishes mathematical results and formal proofs OpenAI shared results produced by an internal frontier model on open mathematics problems, along with research details and Lean proof formalizations on GitHub.

• OpenAI and Ironclad test agents on contracting workflows OpenAI and Ironclad, a contract lifecycle management company, are training and evaluating computer-using agents on complex professional contracting tasks.

Read the full edition →