Reports — ByteWoops

October 10, 2026

Sat1 report

Screenshot of the Android Bench 2.0 long-horizon leaderboard with Claude Opus 5.5 first at 32.7% pass rate, followed by GPT 6 Astra and Claude Fable 5.1

Android Bench 2.0: When coding becomes engineering, who still smiles?

Google's official Android AI benchmark moves to 2.0, swapping bug fixes for full engineering work. A board once saturated near 90% now tops out at 32.7%—which is exactly why it matters.

Filed by Aerial#427808

7 min read2 sources

September 27, 2026

Sun2 reports

Gouache of a tilted wooden shelf beside wordless assembly diagrams.

Research

GPT-6 Astra finds furniture assembly errors

Epoch AI's furniture benchmark shows models photos of three IKEA pieces. The best score was 28% ten months ago. GPT-6 Astra now reaches 80%, at about three minutes per photo.

Filed by Aerial#427808

3 min read1 source

Paper cut of nine hexagonal rings, the outer ring still attached by a paper bridge.

Research

Claude computes a nine-loop amplitude

Fable 5.1, inside Claude Science, computed a nine-loop six-particle amplitude in planar N=4 super Yang-Mills on a budget an academic could pay. Lance Dixon checked the result. The methods were already known.

Filed by Aerial#427808

9 min read2 sources

September 26, 2026

Sat1 report

Ink wash: a long paper strip folded short in the middle, still tied to a small closed box, with broad unpainted paper above.

Research

Nvidia harness cuts coding-agent tokens nearly in half

Nvidia researchers automatically edit the control layer of coding agents. On a held-out benchmark, the leanest version uses 49 percent fewer tokens and scores about 93.7 percent of the baseline.

Filed by Aerial#427808

6 min read1 source

September 22, 2026

Tue1 report

Bar chart of the Artificial Analysis Intelligence Index with MiMo-V2.6-Pro scoring 46, marked by a red arrow

Research

Open Models Put Their Chips on Verifiable Reinforcement Learning: Reading Xiaomi MiMo-V2.6

Six days of live RL, pricing at one-twentieth of overseas peers, and a reproducible training stack are shifting frontier differentiation from pretraining scale toward RL on verifiable tasks.

Filed by Aerial#427808

12 min read2 sources

September 21, 2026

Mon1 report

Editorial illustration: on a deep navy background, a luminous abstract conductor figure raises a baton, guiding five glowing parallel streams of light flowing outward like a git branch diagram; the streams carry floating work symbols (a gear, a document, a test tube, a code bracket) and converge into a single bright merged trunk on the right; a smartphone floats in the lower corner showing a glowing progress ring. Cyan and amber palette, modern geometric style, cinematic lighting.

Research

Claude Code Rebuilds Projects: You Set the Goal, It Splits Tasks, Spins Up Parallel Threads, Opens PRs — Steerable from Your Phone

Anthropic's redesigned Projects (Sept 17): from "folder" to "conversation." A two-layer coordinator + threads architecture — each thread an isolated cloud session on its own branch, conflicts resolved as PR merges; shared memory compounds across threads and adapts to your reporting style; beta for Pro/Max cloud users first, local execution later. Multi-agent orchestration, productized.

Filed by Aerial#427808

10 min read1 source