Current developments in generative AI, software engineering, programming, and digital products, with links to original sources.
Mistral Releases 1-Trillion-Parameter "Le Chonk" — OpenAI Updates Model Spec and Doubles Pro Limits, Plus Reports of Spying Targeting NVIDIA's CEO
Mistral AI has released "Le Chonk" (Mistral Large 4), a 1-trillion-parameter open-weight model. The large model was trained entirely in Europe, with an eye toward self-operated cyber defense. OpenAI announced an updated Model Spec and a doubling of Pro usage limits. We also covered reports of Chinese spying targeting NVIDIA's CEO and Thrive Holdings' acquisition of ADT.
Qwen4 as Soon as Next Week — API Access Comes to ChatGPT Pro/Business, and Reports Emerge of Claude Code Sending Email Addresses to External Sites
At the Apsara conference, Alibaba Cloud's CTO mentioned a Qwen4 release next week. OpenAI is adding up to $500 in monthly API credits to ChatGPT Pro/Business plans. In earnings news, Google admitted its AI search is less profitable, and ServiceNow plunged on AI fears. Reports on how Claude Code handles email addresses are also making the rounds.
DOJ Opens Inquiry into NVIDIA's Hugging Face Acquisition — Anthropic Discloses Pre-IPO Risks, Alphabet's Earnings Miss Reignites AI Spending Concerns
Reuters reports that the U.S. Department of Justice has opened an inquiry into NVIDIA's roughly $13 billion acquisition of Hugging Face. Anthropic, meanwhile, disclosed tensions with the administration as a risk ahead of its IPO. In earnings news, Microsoft's Azure surged 48%, while Alphabet missed revenue forecasts, reigniting concerns over AI spending. On the community side, a 10GB pruning of a 30B speech LLM and LIMe, which cuts training compute by 15%, are generating buzz.
Qwen4 Heads Toward a 1M-Token Context — Liquid AI's 2B Omni-Modal Model, Samsung to Boost Capex 60%
A pull request in qwen-code reveals that the Qwen4 series will support a 1M-token context, with Qwen4-max possibly landing on November 1. Liquid AI released LFM3-Omni-2B, a 2.2B model that sees, hears, and speaks. Samsung is raising capex 60% and Advantest hiked its forecast, putting hard numbers behind the AI semiconductor megacycle, while Microsoft disclosed Azure revenue in dollars for the first time.
Gemini 4 Carbon Arrives Two Weeks After Argon — Microsoft Books $10 Billion a Quarter from OpenAI, and Android Studio Opens Up to AI Agents
Google DeepMind has unveiled Gemini 4 Carbon, a new frontier model specialized for cybersecurity, and it now stands alone at the top of the AI Intelligence Index v5. Microsoft recognized $10 billion in quarterly revenue from its contract with OpenAI. Android Studio has opened up to more than 40 external agents, and NVIDIA has open-sourced a multilingual transcription model.
AI Agents Break Into Anthropic's Internal Systems — Coding Interviews to Be Scrapped, “Hunger beats tenure,” and 35B MoE at 262k Context on a Single P40
Your evening AI news roundup for October 28. Cathedral's AI agents breached Anthropic's internals via third-party software, with access to sales, HR, and IT data acknowledged. Anthropic has announced the end of coding interviews. Also tonight: a warning about attacks via AGENTS.md in the downloads folder, a 35B MoE put through its paces on a Tesla P40, the release of diffusion LM “Clover,” and AI bots flooding a UK government consultation.
Anthropic Moves Into Drug Discovery, Acquiring a One-Person Bio Company — Intelligence Index v5 Widens the Open-Weight Gap, and Agent Memory Falls Short of Naive RAG
Anthropic is moving into drug discovery targeting rare pediatric muscular dystrophies and has acquired a one-person bio company that ran on Claude. The Artificial Analysis Intelligence Index v5 widens the open-weight gap to 12 points. We also round up two studies questioning the net value of agent memory, a practical update to Qwen3.5-4B-3D-Coder, and Discord conversations becoming searchable.
UEFI Boot Vulnerability Strikes Servers Worldwide — OpenAI Sued Over GPT-4o Medication-Tapering Advice, Memory Shortage Seen Lasting Through 2030
A severity 9.5 code execution vulnerability has been found in EDK II, the open-source code that underpins server boot, putting most of the world's servers within its reach. OpenAI has been sued over GPT-4o's medication-tapering advice. SK Hynix says excess demand will persist through 2030. Also covered: the 42x cost gap in reasoning effort and 'RSI Arena,' where 10 models post-trained themselves through autonomous learning.
Google DeepMind's Genie 4 Sustains Audio-Equipped 3D Worlds for Two Minutes — AI Claims to Refute the Exponential Time Hypothesis, Opera's Local AI Standard, and the GMO Data Leak
Google DeepMind has unveiled Genie 4, a world model that generates audio-equipped 3D worlds in real time, with scene consistency lasting up to two minutes — twice as long as the previous generation. Meanwhile, frontier AI labs claim to have refuted the Exponential Time Hypothesis, a long-standing assumption in computational complexity theory. Opera has proposed WebAI, a new standard for running AI locally in the browser, and the GMO cloud breach has raised suspicions that customer data was published.
NVIDIA Invests $3 Billion in Meta — Cloudflare Discloses Workers AI Data Leakage, and OpenAI's Math Proofs Go to Outside Experts
NVIDIA invests $3 billion in Meta, which plans to spend even more on NVIDIA chips. Cloudflare discloses a bug where Workers AI responses could contain other users' content. OpenAI sends math, CS, and physics proofs derived by its internal AI for verification by Nobel laureates and other outside experts. Plus: a community rebuttal of the Qwen exfiltration claim, the release of Ling-3.1-flash-1019 and Unlimited-OCR-2, and a record of an AI agent adopting an OSS project left untouched for 8 years.
