The AI Wire
The firehose

Everything we scan. Filtered to signal.

The wire reads a flood every day so you read a page. Here is the whole intake, the filter, and the last seven days in the raw.

0

Sources

arXiv, HF, GitHub, more

0

Scanned · 30d

items read and scored

0

Signal · 30d

scored 7+ out of 10

We read about 7 so you read one.

What comes down the wire

last 30 days

1,608

News

601

Papers

318

Models

14

Tools

The last 7 days, in the raw

open to everyone
  1. @claudeai: Introducing Claude Opus 5. It's a thoughtful and proactive model that comes clowire10
  2. @ArtificialAnlys: Google has released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Both halve time wire9
  3. Kimi K3 Now Available via Telnyx Inference APIwire9
  4. Kimi K3: Open Frontier Intelligencewire9
  5. Kimi K3: Open Frontier Intelligencewire9
  6. @Intoshapeoff: Kimi K3 Model Weights Released: > Moonshot AI is releasing the full weights of wire9
  7. Deepseek V4 Flash ~105 t/s on two Nvidia 4090d 48G (ada) in vLLMwire9
  8. AMD partners with Claude creators Anthropic, investing up to $5 billion to deploy 2 gigawatts of data center GPUswire9
  9. @OpenAI: ChatGPT Voice is now in the desktop app. Control your computer and direct multiwire9
  10. AMD commits up to $5 billion to Anthropicwire9
  11. @Arielladailyai: A Chinese startup just released the largest open-weight AI model ever built, anwire8
  12. @himanshu9132: 🚀 Anthropic just shifted the AI landscape from "smartest at any cost" to "frontiwire8
  13. Opus 5's effort dial is not monotonic. Above "high", coding scores go down, and Anthropic's own migration guide says so.wire8
  14. Claude Opus 5: The System Cardwire8
  15. Introducing Claude Opus 5wire8
  16. OpenAI Model Hacks Into HuggingFace During Cybersecurity Evaluationwire8
  17. SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPODwire8
  18. Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyberwire8
  19. Kimi K3 text-only for llama.cppwire8
  20. Nvidia, Microsoft launch open AI security alliance, without OpenAI, Google, or Anthropicwire8
  21. Microsoft launches its first cybersecurity model, plus a new agentic cybersecurity systemwire8
  22. Benchmarked every spec-decode method on Qwen3.6-27B across vLLM and SGLang (single RTX PRO 6000 Max-Q)wire8
  23. Open-weight 4B models approach o3-level medical question answering in Swedish [P]wire8
  24. More On An Internal OpenAI Model Hacking Into HuggingFacewire8
  25. Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learningwire8
  26. @onusoz: I have not been as excited for a model release since o1 came out Pivotal momentwire8
  27. I built a compiler that turns computation graphs into the weights of a vanilla transformer, no training anywhere [P]wire8
  28. Anthropic releases Opus 5 with ‘close’ to Fable 5’s capabilitieswire8
  29. GPT-5.5 Scores 10.6% on ActiveVision, Humans Hit 96.1% [R]wire8
  30. Prompt Injection in NeurIPS 2026? [D]wire8
  31. AI #178: A Fire Alarm For General Intelligencewire8
  32. Claude’s voice mode is now available for Opus and Sonnetwire8
  33. LLMs Get Lost in Evolving User Intentwire8
  34. ICAE-Bench: Evaluating Coding Agents as Interactive Project Builderswire8
  35. an AI agent got prompt-injected into moving $175K on-chain. first documented case of this actually happeningwire8
  36. Introducing OpenAI Presencewire8
  37. Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimizationwire8
  38. @MistralAI: Mistral is announcing an expanded global strategic partnership with @Microsoft twire8
  39. @pankajkumar_dev: Claude Opus 5: Launching Today - Claude Opus 5 is scheduled to launch today (Thwire8
  40. OpenAI and Hugging Face address security incident during model evaluationwire8
  41. Gemini 3.6 Flash looks better on paper. What would make you block the upgrade?wire8
  42. Europe just forced Google to open Android to every competing AI. And Gemini 3.5 Pro missed its deadline for the third time this week.wire8
  43. CircuitKIT : Circuit Discovery, Evaluation, and Application Toolkit for Mechanistic Interpretabilitywire8
  44. @testingcatalog: Google released "Gemini 3.5 Flash Cyber" on CodeMender, a new model for finding wire8
  45. @bally_kehal: 1/ The quietest release of the month is the one that matters most. Mistral shipwire8
  46. A $500 RL fine-tune of a 9B open model beat frontier models on catalog reviewwire7
  47. Evaluated 6 frontier LLMs (GPT-5.4, Claude Sonnet 4.6, Claude Opus 4.7, Gemini Pro/Flash, Grok 4.3) on political, gender, and racial bias across 8 benchmarks (~20,600 examples) [R]wire7
  48. Private Claude chats exposed on Google search resultswire7
  49. Why China is giving away its best AI modelswire7
  50. PSA: Your Claude shared chats and Artifacts may have ended up on Googlewire7
  51. OpenAI’s Hugging Face breach has reignited the debate over alignment and controlwire7
  52. A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Foreverwire7
  53. @k44yej: New Model launch at @MicrosoftAI! You can read more about the launches in our twwire7
  54. Final class-action settlement approval granted, judgment entered, and attorneys' fees awarded in the Bartz v. Anthropic AI copyright casewire8
  55. Safety and alignment in an era of long-horizon modelswire8
  56. Anthropic’s landmark $1.5B copyright settlement is approvedwire8
  57. Google is working on a new AI chip designed to make Gemini more efficientwire8
  58. @sarbjeetjohal: Kimi K3 has sent the shockwaves through investor community since Friday, last wewire8
  59. We compared different LLMs on IMO 2026 [R]wire7
  60. Sources: OpenAI and Anthropic quietly lobby Washington regulators to restrict open-source AI models, even as Sam Altman publicly says he supports open source AIwire7
Behind the curtain

A subscriber tier is coming.

The firehose is the raw intake. The next layer is judgment: the handful of things I think are worth your week, annotated, plus the full decode and dossier library and the complete archive.

  • The curated picks, with the why
  • Decodes & dossiers, full library
  • The complete searchable archive
  • The week ahead, before it’s news