NEWS

GLM 5.2 and the AI margin collapse thesis: why open-weights models could squeeze frontier lab margins, where HN pushes back, and what it means for model choice.

Ilya Sutskever's rumored 30-paper ML reading list, now on 30papers.com. What the papers cover, the compression thesis behind them, and what HN thinks.

Small AI models run on handheld devices where networks fail, like an on-device pill scanner. HN debates offline emergency kits and domain-specific models.

Ternlight is a 7 MB ternary-quantized embedding model that runs semantic search in the browser via WASM, with no API calls. How it works and what HN thinks.

A Dartmouth study of an AI tutor in intro statistics links full engagement to a 0.71-1.30 SD exam gain. What it measured, and why selection bias still matters.

Anthropic's new research reveals LLMs have an internal 'workspace' for silent reasoning - and it could change how we build safer AI.

A controlled study of 660 Claude Code trials shows clean codebases reduce token usage by 7-8% and file revisitations by 34%, while pass rates stay the same. Traditional maintainability principles still matter in the age of AI coding.

Elm 0.19.2 brings faster builds on the road to 1.0, and Evan Czaplicki previews equatable and hashable types from the Acadia project. What HN makes of it.

GPT-5.6 Sol Ultra is coming to Codex with trained subagents that communicate during tasks, reportedly hitting 91.9% on Terminal-Bench 2.1. What HN thinks.

Comparing LLMs by token pricing alone can lead you to choose worse, more expensive models. Cost per task tells the real story.

A simple image rotation button reveals deep truths about responsive interface design - why buttons must always respond predictably, even during animations.

Flipper Zero firmware reaches stable 1.0 and moves to a community contribution model. What changes for contributors and how HN sees official vs custom firmware.

Douglas Thain's free compilers textbook walks you through building a C-like compiler that emits real X86 and ARM assembly, plus what HN recommends reading next.

Event sourcing for AI agents: a new paper makes the append-only log the source of truth for agent state. What it enables, and whether HN thinks it is just CQRS.

jamesob's local LLM guide: the $2k and $40k hardware paths, the BIOS and kernel settings that matter, and why multi-GPU setups fail at PCIe and IOMMU.

Leanstral 1.5 is Mistral's open Apache-2.0 model for Lean 4 theorem proving. The benchmark claims, the HN pushback on its bug-finding pitch, and how to use it.

Box3D is an open source 3D physics engine from the creator of Box2D, with cross-platform determinism, a wide SIMD contact solver and Valve Rubikon heritage.

Cloudflare's Monetization Gateway adds native x402 support, so developers can charge for API calls with stablecoin micropayments and no accounts or API keys.

Godot bans autonomous AI agents and substantial AI-generated code in contributions. What the policy allows, why maintainers want it, and how HN reacted.

Webernetes ports Kubernetes to TypeScript so a cluster runs in the browser. What ngrok's Sam Rose rebuilt, what is left out, and how it helps teach k8s.