2026-09-04
Top stories
- ARC Prize measures GPT-6 Astra at 62.7 percent on ARC-AGI-3 under its neutral harness and 99.9 percent under the provider adapter ARC Prize measured GPT-6 Astra at 62.7 percent on ARC-AGI-3 in its neutral harness and 99.9 percent via the provider adapter.
- GitHub changes the rate-limited unauthenticated Git response from 403 to 401, with intermittent anonymous clone failures reported GitHub now returns 401 instead of 403 to rate-limited unauthenticated Git requests, and clones fail intermittently.
- Artificial Analysis prices GPT-6 Astra at 2.5 times GPT-5.6 Sol and records evaluation regressions beside the coding gains Artificial Analysis prices GPT-6 Astra at 2.5 times GPT-5.6 Sol and records evaluation regressions beside its coding gains.
AI
ML research
- A preregistered audit reports LLM judge rankings on shared endpoints failing their own reliability gates A preregistered audit reports LLM judge rankings on hosted endpoints failing the reliability gates the authors set in advance.
- SWE-Gate finds a third of functionally passing agent repairs violate reviewer constraints A preprint reports 221 of 644 agent repairs that passed functional tests failed constraints derived from real review comments.
Agentic coding
- Armature publishes 5,292 valid sessions measuring which third-party services coding agents install Armature published 5,292 valid agent sessions showing repository language, not the stated requirement, decided which vendor won.
- Hugging Face releases funes, a local-first memory layer built from existing coding-agent session traces Hugging Face released funes, a local-first memory layer that recalls original agent turns from existing session traces.
- A pilot study reports coding agents route between grep and LSP by task, and that output shape moved success more than retrieval precision A pilot study reports coding agents route between grep and LSP by task, and that output shape moved success more than precision.
- Google's Antigravity terms make third-party clients grounds for account suspension Google's Antigravity terms name third-party clients a breach and grounds for suspending Antigravity and Gemini CLI accounts.
Languages and runtimes
- Daniel Lemire publishes a case where Python set and dict operations run in quadratic time Daniel Lemire published a three-line Python case where set and dictionary operations run in quadratic rather than linear time.
- Rust 1.98.1 fixes a vtable miscompilation that emitted undefined behaviour Rust 1.98.0 could emit a trait object vtable holding a null pointer where a function pointer belongs, and 1.98.1 fixes it.
Engineering posts
- The author of a 1993 Amiga game reports rebuilding its 68000 assembly byte-identically and porting it to Godot with a coding agent The author of a 1993 Amiga game rebuilt its 68000 assembly byte-identically, then ported it to Godot with a coding agent.
- Nolan Lawson argues coding agents are displacing frontend expertise Nolan Lawson argues frontend work is the cheapest code to hand an agent unsupervised, and that agent experience now drives stacks.
Hacker News
- Ask HN: Why were OpenAI, Claude, and Grok simultaneously down? An Ask HN thread with 541 comments worked the cross-provider outage correlation from the outside before any vendor named a cause.
- Ask HN: Who is using MCP in production? An Ask HN thread with 134 comments collects self-reported Model Context Protocol deployments rather than demos, unaudited.