Loose posts

Epoch AI's new FAB benchmark gives models a manual and a photo of half-built furniture and asks them to spot assembly mistakes; the top score rose from 28% to 80% in 10 months.

A pixel-style animation built with Opus 5.5 that dynamically walks through neural network training, which the author calls the best demo of its kind they have seen.

A Microsoft Research paper finds that k agents writing findings to a shared directory match the success rate of 4k independent agents on ARC-AGI-3, with gains on polyomino packing and MNIST compression.

Three models working as a self-organizing team averaged 66.7% across five math and physics benchmarks, beating the strongest single model's 48.8% and a perfect router's 59.0%, with strategies learned from just 15 problems.

After signing $531 billion in compute deals over 11 months, Anthropic is now negotiating direct data center leases to gain more control over chip supply, pricing and infrastructure.

Over 70 dated model outputs for the same SVG prompt, from December 2024 to September 2026, show the progression from unrecognizable shapes to an interactive page.

Researchers ran 7 human psychological instruments on 9 LLMs five times each in English and Chinese, finding each model has a reproducible, unique profile and that language and vendor shift results.

A small weight generator turns new facts into temporary in-model plugins during a chat, avoiding repeated long-context reads and beating standard prompting on accuracy and stability with complex material.

Anthropic says Claude is involved in 26% of its R&D, up from under 1% in February, with 30,000 agents running internally and about 100,000 records flagged for review each week.

RRSI adds a regularization constraint to agent self-evolution so it favors reusable changes, improving out-of-distribution results on coding and engineering tasks while using 30% fewer tokens.

Dan Shipper's early review finds Opus 5.5 strong at end-to-end coding and top on reading ease, but prone to doing too much and running past time limits, at $5/$20 per million tokens.

The paper proposes Wiki Foundation Model, which turns an LLM wiki into a graph and retrieves from it with query-conditioned message passing, plus a training protocol that runs 10.5x faster.