Issue · 2026-09-08

Most AI news is noise.Here is the rest.

MODELS

GPT-6 Astra: real gains, real complaints

One camp says it's mostly toy demos: Blender/threejs games, nothing real, tried it 24h and now thinks it's overhyped (scheemunai).
Theo's take splits the difference: it's done incredible things and also some of the stupidest things he's seen a model do (theo).
Practical fix if you're rationing quota: Astra/5.6 Pro chat has its own limit separate from Codex, so falling back to chat doesn't burn Codex quota (miu21590).

ORCHESTRATION

Same model weights score very differently depending on the harness around them — this is the argument for treating harnesses as real research, not scaffolding (ycombinator).
One architecture take: constrain agents with a harness of deterministic tools that force a strict process, rather than trusting free-form prompting (unclebobmartin).
Concrete review technique list for AI-authored PRs: diff-rendering tools, pseudocode descriptions of changes, mermaid diagrams (mattpocockuk).
A real paper (HarnessDev) on whether LLMs can create and evolve their own agent harness, worth reading over the hot takes (_akhaliq).

CONTEXT

Tip: enable Codex's experimental context management setting and raise the max token limit for Astra (brandon_galang).
Pi-Codex shipped its own compaction fix plus guidance on every /codex setting (Howaboua).
Someone traced what the new compaction actually does under the hood, instead of repeating the marketing claim that it "saves notes and searches messages" (owengretzinger).
Counter-report: the experimental context management feature is broken or incompatible with Sol for at least one user — test before relying on it (aienginerd).

DIGEST

A math professor's account of how a year of drafts on the Navier-Stokes problem played out, and how the "OpenAI solved it" claim emerged and got walked back (ns123abc).
Clean timeline for anyone confused: Buckmaster and Alpöge made progress in August, no confirmed solve, and the AGI framing came later (konstiwohlwend). Useful context for how much weight to put on lab breakthrough claims going forward.