<< Go back to HN Digests
Warning : This document is still a draft

2026 W1 - 02-xx




First time here ? Check the introduction here

What you Missed this Week

  • OpenAI and Microsoft end their exclusivity agreement. (Link)
  • Talkie, a “vintage” LLM, stuck in 1930 (Link)
  • Mistral released a new coding model, as good as sonnet, for the price of gpt-oss-120B (Link)
  • IBM also released some models, Granite 4.0 (Link)
  • Manus.ai aquisition blocked by China (Link)

Articles

State of the Art

New Models

Model improvement or other low level optimization

  • 768 pt [Link / HN] Talkie: a 13B vintage language model from 1930
  • 498 pt [Link / HN] Mistral Medium 3.5
  • 386 pt [Link / HN] VibeVoice: Open-source frontier voice AI
  • 326 pt [Link / HN] Granite 4.1: IBM’s 8B Model Matching 32B MoE
  • 101 pt [Link / HN] Laguna XS.2 and M.1

Tunning

  • 200 pt [Link / HN] Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs
  • 138 pt [Link / HN] Advanced Quantization Algorithm for LLMs

Hardware

  • 276 pt [Link / HN] How fast is a macOS VM, and how small could it be?
  • 239 pt [Link / HN] Show HN: Auto-Architecture: Karpathy’s Loop, pointed at a CPU
  • 171 pt [Link / HN] Eka’s robotic claw feels like we’re approaching a ChatGPT moment
  • 127 pt [Link / HN] Show HN: Utilyze – an open source GPU monitoring tool more accurate than nvtop

Coding Agents

Claude new functionalities, tips and tricks dedicated to a specific agent, …

  • 1495 pt [Link / HN] VS Code inserting ‘Co-Authored-by Copilot’ into commits regardless of usage
  • 1336 pt [Link / HN] Claude Code refuses requests or charges extra if your commits mention “OpenClaw”
  • 1249 pt [Link / HN] HERMES.md in commit messages causes requests to route to extra usage billing
  • 634 pt [Link / HN] DeepClaude – Claude Code agent loop with DeepSeek V4 Pro
  • 555 pt [Link / HN] Who owns the code Claude Code wrote?
  • 498 pt [Link / HN] Mistral Medium 3.5
  • 422 pt [Link / HN] A desktop made for one
  • 401 pt [Link / HN] Uber torches 2026 AI budget on Claude Code in four months
  • 392 pt [Link / HN] Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview
  • 312 pt [Link / HN] GitHub Copilot code review will start consuming GitHub Actions minutes
  • 277 pt [Link / HN] Specsmaxxing – On overcoming AI psychosis, and why I write specs in YAML
  • 258 pt [Link / HN] Reverse Engineering SimTower
  • 252 pt [Link / HN] Regression: malware reminder on every read still causes subagent refusals
  • 226 pt [Link / HN] Open Design: Use Your Coding Agent as a Design Engine
  • 175 pt [Link / HN] The agent harness belongs outside the sandbox
  • 138 pt [Link / HN] A good AGENTS.md is a model upgrade. A bad one is worse than no docs at all
  • 110 pt [Link / HN] EvanFlow – A TDD driven feedback loop for Claude Code
  • 106 pt [Link / HN] We decreased our LLM costs with Opus
  • 102 pt [Link / HN] Flue is a TypeScript framework for building the next generation of agents
  • 98 pt [Link / HN] Show HN: AI CAD Harness
  • 90 pt [Link / HN] Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell

Personal Blogs / Opinions

Investigation

Trend investigation, hacking, data analysis, …

  • 598 pt [Link / HN] 4TB of voice samples just stolen from 40k AI contractors at Mercor
  • 508 pt [Link / HN] How ChatGPT serves ads
  • 471 pt [Link / HN] Opus 4.7 knows the real Kelsey
  • 464 pt [Link / HN] Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library
  • 382 pt [Link / HN] Apple accidentally left Claude.md files Apple Support app
  • 373 pt [Link / HN] Kimi K2.6 just beat Claude, GPT-5.5, and Gemini in a coding challenge
  • 346 pt [Link / HN] Using “underdrawings” for accurate text and numbers
  • 332 pt [Link / HN] AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights
  • 299 pt [Link / HN] TurboQuant: A first-principles walkthrough
  • 277 pt [Link / HN] Specsmaxxing – On overcoming AI psychosis, and why I write specs in YAML
  • 258 pt [Link / HN] Reverse Engineering SimTower
  • 241 pt [Link / HN] He asked AI to count carbs 27000 times. It couldn’t give the same answer twice
  • 232 pt [Link / HN] I won a championship that doesn’t exist
  • 153 pt [Link / HN] Show HN: State of the Art of Coding Models, According to Hacker News Commenters
  • 143 pt [Link / HN] Ramp’s Sheets AI Exfiltrates Financials
  • 143 pt [Link / HN] I Got Sick of Remembering Port Numbers
  • 129 pt [Link / HN] Running local LLMs offline on a ten-hour flight
  • 131 pt [Link / HN] Softmax, can you derive the Jacobian? And should you care?
  • 129 pt [Link / HN] Running local LLMs offline on a ten-hour flight
  • 116 pt [Link / HN] Refusal in Language Models Is Mediated by a Single Direction

Tool for AI / Made with AI / showHN

Any Github / ShowHN for new tools / product

  • 671 pt [Link / HN] The gay jailbreak technique (2025)
  • 634 pt [Link / HN] DeepClaude – Claude Code agent loop with DeepSeek V4 Pro
  • 392 pt [Link / HN] Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview
  • 252 pt [Link / HN] Regression: malware reminder on every read still causes subagent refusals
  • 226 pt [Link / HN] Open Design: Use Your Coding Agent as a Design Engine
  • 207 pt [Link / HN] Mike: open-source legal AI
  • 206 pt [Link / HN] OpenWarp
  • 172 pt [Link / HN] Text-to-CAD
  • 168 pt [Link / HN] Understand Anything
  • 127 pt [Link / HN] Show HN: Utilyze – an open source GPU monitoring tool more accurate than nvtop
  • 113 pt [Link / HN] Show HN: DAC – open-source dashboard as code tool for agents and humans
  • 110 pt [Link / HN] EvanFlow – A TDD driven feedback loop for Claude Code
  • 97 pt [Link / HN] Show HN: Agent-desktop – Native desktop automation CLI for AI agents
  • 90 pt [Link / HN] Show HN: Pu.sh – a full coding-agent harness in 400 lines of shell
  • 83 pt [Link / HN] Voice-AI-for-Beginners – A curated learning path for developers

Opinions

  • 655 pt [Link / HN] DeepSeek V4 – almost on the frontier
  • 679 pt [Link / HN] The Zig project’s rationale for their anti-AI contribution policy
  • 555 pt [Link / HN] Who owns the code Claude Code wrote?
  • 471 pt [Link / HN] Opus 4.7 knows the real Kelsey
  • 413 pt [Link / HN] Agentic Coding Is a Trap
  • 333 pt [Link / HN] To my students
  • 303 pt [Link / HN] “Why not just use Lean?”
  • 236 pt [Link / HN] AI’s economics don’t make sense
  • 232 pt [Link / HN] 10Gb/s Ethernet: what I did to get it working in my home
  • 232 pt [Link / HN] I won a championship that doesn’t exist
  • 226 pt [Link / HN] For thirty years I programmed with Phish on, every day
  • 210 pt [Link / HN] The ‘Hidden’ Costs of Great Abstractions
  • 190 pt [Link / HN] Security through obscurity is not bad
  • 169 pt [Link / HN] “People who don’t use AI will be left behind”
  • 155 pt [Link / HN] LLMs Are Not a Higher Level of Abstraction
  • 133 pt [Link / HN] If I could make my own GitHub
  • 129 pt [Link / HN] Good developers learn to program. Most courses teach a language
  • 129 pt [Link / HN] Running local LLMs offline on a ten-hour flight
  • 108 pt [Link / HN] Your CEO is suffering from AI psychosis
  • 107 pt [Link / HN] Your biggest vulnerability is your shitty compensation
  • 102 pt [Link / HN] Why are neural networks and cryptographic ciphers so similar? (2025)
  • 90 pt [Link / HN] NHS goes to war against open source

  • 80 pt [Link / HN] AI, Intimacy, and the Data You Never Meant to Share

  • 407 pt [Link / HN] AI uses less water than the public thinks

  • 326 pt [Link / HN] OpenAI models coming to Amazon Bedrock: Interview with OpenAI and AWS CEOs

Officials

Politics / Strategy

Merge, acquisition

  • 988 pt [Link / HN] Microsoft and OpenAI end their exclusive and revenue-sharing deal

News

Anything from “official” journals

  • 490 pt [Link / HN] OpenAI’s o1 correctly diagnosed 67% of ER patients vs. 50-55% by triage doctors
  • 400 pt [Link / HN] China blocks Meta’s acquisition of AI startup Manus
  • 372 pt [Link / HN] Warp is now open-source
  • 316 pt [Link / HN] Google and Pentagon reportedly agree on deal for ‘any lawful’ use of AI
  • 312 pt [Link / HN] GitHub Copilot code review will start consuming GitHub Actions minutes
  • 288 pt [Link / HN] Why AI companies want you to be afraid of them
  • 285 pt [Link / HN] Spotify adds ‘Verified’ badges to distinguish human artists from AI
  • 256 pt [Link / HN] Anthropic Joins the Blender Development Fund as Corporate Patron
  • 154 pt [Link / HN] Claude for Creative Work
  • 143 pt [Link / HN] After dissing Anthropic for limiting Mythos, OpenAI restricts access to Cyber
  • 125 pt [Link / HN] The More Young People Use AI, the More They Hate It

Unrelated

Not related to AI, but worth reading

[OUT-OF-SCOPE] - 32 items

  • 401 pt [Link / HN] Why TUIs are back
  • 123 pt [Link / HN] Group averages obscure how an individual’s brain controls behavior: study
  • 122 pt [Link / HN] A more efficient implementation of Shor’s algorithm
  • 109 pt [Link / HN] Show HN: Ableton Live MCP
  • 109 pt [Link / HN] Snowball Earth may hide a far stranger climate cycle than anyone expected
  • 107 pt [Link / HN] Coffee with a splash of physics: how to make the most out of your brew
  • 104 pt [Link / HN] Monad Tutorials Timeline
  • 84 pt [Link / HN] Spirit Airlines canceled all flights and is going out of business

>> You can subscribe to my mailing list here for a monthly update. <<