Weekly AI: all issues

The week in AI, in one email.

Week #4

Weekly AI #4: Inkling 2 Small, GPT-6 Astra weight editing and Anthropic's IPO portal

This week Mira Murati's Thinking Machines released Inkling 2 Small, an open-weights 594B MoE model, while MIT and OpenAI demonstrated Weight-Space Programming on GPT-6 Astra without retraining. Anthropic launched its official investor portal ahead of a planned Nasdaq listing, and Claude achieved scientific milestones in autonomous protein design and 40-year-old molecular structures. Google introduced Gemini Omni 1.2 Flash for cheaper video generation, alongside three major funding roundups covering TEKEVER, Hypersonica, and Trajectory. Every item below links to the full story on Has Brains AI, with sources.

  1. Thinking Machines releases 594B MoE model Inkling 2 Small

    Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling 2 Small, an open-weights Mixture-of-Experts (MoE) foundation model totaling 594 billion parameters with 17 billion active per token.

  2. GPT-6 Astra fixes model reasoning by editing 0.016% of weights without retraining

    Researchers at MIT CSAIL demonstrated Weight-Space Programming, using GPT-6 Astra to directly inspect and edit neural network weights in Python without fine-tuning, backpropagation, or training loops.

    OpenAI

  3. Google releases Gemini Omni 1.2 Flash with 30% cheaper video generation

    Google released Gemini Omni 1.2 Flash, expanding video generation clips up to 15 seconds and reducing generation costs by 30%. The model is available immediately via the Gemini API, Google AI Studio, Google Vids, and Flow.

    Google

  4. Baidu open-sources ERNIE 5.5 Flash Base 280B MoE model

    Baidu Research open-sourced ERNIE 5.5 Flash Base, a 280-billion parameter mixture-of-experts (MoE) model that activates only 14 billion parameters (5%) per token. Model weights are available immediately on Hugging Face.

  5. AI Roundup: OpenAI GPT-Image 2.6, SWE-2 Flash, Apple AI budgets, and Agility Robotics IPO

    OpenAI released GPT-Image 2.6 with faster generation and SVG export, while Cognition unveiled SWE-2 Flash and Agility Robotics completed its Nasdaq listing. Meanwhile, Apple lowered internal daily AI developer spend limits to manage computing expenses.

  6. NVIDIA releases Nemotron 3.5 Bolt E4B for on-device AI applications

    NVIDIA published weights for Nemotron 3.5 Bolt E4B, a compact 8-billion parameter model optimized for on-device deployment across GeForce RTX graphics cards, Jetson embedded boards, and Spark edge computers.

  7. Tencent introduces Hyra 2.0 autonomous research agent

    Tencent unveiled Hyra 2.0, an autonomous research agent designed to conduct end-to-end scientific and machine learning experimentation with persistent cross-task tool accumulation.

  8. Anthropic launches dedicated IPO portal ahead of Nasdaq listing

    Anthropic published an official investor portal at ipo.anthropic.com ahead of its planned Nasdaq listing under the ticker ANT. The site outlines retail participation pathways and partnerships with brokerage platforms including Interactive Brokers and Revolut.

    Anthropic

  9. Anthropic spends more on internal Claude usage than employee payroll

    Anthropic spends more on internal Claude model usage than on employee base salaries, according to disclosures in the company's IPO registration filing.

    Anthropic

  10. US AI hardware stocks tumble following Democratic datacenter moratorium proposal

    US technology and semiconductor stocks recorded sharp declines following the conclusion of the US elections as congressional Democrats outlined aggressive plans to regulate frontier artificial intelligence and place a moratorium on new AI datacenter construction.

  11. OpenAI weighs revisions to 2027 datacenter roadmap amid local power constraints

    OpenAI leadership is actively discussing revisions to its 2027 computing capacity forecasts as local community opposition and utility grid bottlenecks delay new datacenter construction across the United States.

    OpenAI

  12. 600,000 NVIDIA H200 GPUs held in Hong Kong amid Chinese import freeze

    Approximately 600,000 NVIDIA H200 AI accelerators are parked in Hong Kong storage facilities awaiting clearance as Beijing regulators hold import permits for mainland Chinese hyperscalers, according to reporting by Reuters.

  13. AI chipmakers and infrastructure stocks slide in sharp October sell-off

    Semiconductor manufacturers, AI cloud infrastructure providers, and hardware suppliers recorded double-digit declines across October as investors repriced capital intensity and timeline expectations for artificial intelligence infrastructure.

  14. Google DeepMind employment agreements included lifetime non-disparagement clause

    Google DeepMind included lifetime non-disparagement provisions in employment contracts for UK researchers that prohibited public criticism of the company even after leaving their roles, according to a report by The Washington Post.

    Google

  15. Nucleus Genomics launches AI-assisted polygenic screening for embryo selection

    Peter Thiel-backed startup Nucleus Genomics launched an AI-powered polygenic screening platform for in vitro fertilization (IVF) under the tagline "Have your best baby," offering parents predictive scores for cognitive traits, disease liabilities, and physical characteristics.

  16. Claude autonomously designs proteins binding scorpion neurotoxins

    Anthropic demonstrated that an internal Claude model orchestrated an autonomous protein engineering pipeline to design de novo minibinder proteins that neutralize scorpion neurotoxins. Experimental designs and validation data have been released publicly on Hugging Face.

    Anthropic

  17. Claude resolves 40-year vault particle protein structure puzzle

    Anthropic announced that a specialized variant of Claude resolved the atomic structure of the human major vault protein, deciphering a 40-year-old cellular biology mystery. The structural coordinates and analysis have been published on bioRxiv.

    Anthropic

  18. Anthropic study finds AI could accelerate computational science research up to 16x

    Anthropic analyzed 207,049 scientific sessions across Claude and Claude Code alongside survey responses from 528 active researchers, projecting up to a 16x expansion in scientific output for computational fields. Empirical work showed lesser gains, with researchers in chemistry and materials science reporting 1.5x improvements.

    Anthropic

  19. CLEAR benchmark measures up to 12x tokenizer cost disparity across languages

    The Coalition for Language Equity in AI Research (CLEAR) published an empirical audit revealing that non-English languages face inference cost premiums of up to 12x due to tokenizer bias. The benchmark, dataset, and analysis tools are open-sourced on GitHub and Hugging Face.

  20. Top AI funding rounds: Trajectory, 42AI, Tutor Intelligence, and Larridin

    Trajectory raised $45 million for automated model retraining, while 42AI secured $30 million for autonomous software testing agents. Tutor Intelligence raised $24 million for robotic manipulation and Larridin took in $18 million for synthetic data generation.

  21. Top European AI and robotics funding rounds: Hypersonica, Conlio, CyberKord, and Sortrace

    Hypersonica raised €100 million for autonomous hypersonic defense systems, while Conlio secured €42 million for enterprise workflow agents. CyberKord raised €28 million for defensive security agents and Sortrace added €15 million for warehouse sorting robotics.

  22. Top European AI and robotics funding rounds: TEKEVER, Potential Sciences, Syn2core, and Extend Robotics

    TEKEVER raised $580 million at a $6.4 billion valuation for autonomous uncrewed aircraft, and Potential Sciences secured €65 million for biological foundation models. Syn2core raised €34 million for neuromorphic processors while Extend Robotics took in €18 million for teleoperated robotics.

  23. Anthropic eliminates Claude Code token drain with Haiku 5.5 status updates

    Anthropic released an update to its Claude Code command-line interface that eliminates excessive token consumption from background subagents by delegating periodic status reporting to Claude Haiku 5.5.

    Anthropic

  24. Anthropic open-sources FlashProve for provable AI inference

    Anthropic has released FlashProve, an open-source cryptographic tool and protocol that lets AI providers prove which model generated an answer without revealing model weights. The code and formal Lean 4 verification proofs are available under the Apache 2.0 license on GitHub.

    Anthropic

  25. Google migrates 40-year-old Fortran ocean model to Rust in 59 minutes with Gemini 4 Argon

    Google engineers used Gemini 4 Argon inside Google Antigravity to migrate FVCOM, a 174,000-line Fortran and C ocean simulation, into memory-safe Rust in 59 minutes. The migration workflow has been released as an open-source Antigravity skill.

    Google

  26. AI agents write Rust GPU kernels to make speakrs 60x faster than pyannote

    Developer Praveen Perera migrated the open-source speakrs speaker diarization library from ONNX Runtime to native GPU kernels written in Rust with NVIDIA cuda-oxide, speeding up diarization by up to 60x. AI coding agents authored the majority of the custom GPU kernels.

  27. OpenAI evaluates specialized CAD and electronics engineering model

    OpenAI is conducting closed evaluations of an experimental reasoning model tailored for computer-aided design and printed circuit board layout. The model achieved a 0.875 score on the BenchCAD 2.0 evaluation benchmark, outperforming general-purpose frontier models.

    OpenAI

  28. Hume AI launches Audio Decisions API for native speech evaluation

    Hume AI released Audio Decisions, an API endpoint that analyzes voice recordings directly using native audio modeling to answer questions with structured judgments and confidence scores. The API is available now in developer preview.

  29. DeepSeek releases seven interactive prompts for Harness Creator Mode

    DeepSeek published a collection of seven system prompts for Harness Creator Mode, enabling users to build interactive tools and custom interfaces directly within the application.

  30. Andrej Karpathy vibe codes HackerTube video platform in two hours

    AI researcher Andrej Karpathy launched HackerTube, a customized video recommendation platform curated from two decades of community discussions on Hacker News.