LocalLLaMA user prefers Qwen4-27B at 1M context over big MoE at 128k

A typical Devin Review now costs $0.25–$0.75 per PR, down 82% since August via a better model mix and harness optimizations. Bugs found per PR rose from 1.22 to 1.32, and each fixed bug costs customers about $2.25 on average.

METR agrees July 15 models posed low catastrophic risk but says it's plausible Anthropic had crossed its early warning threshold for automated R&D, and that the public report understates acceleration. Anthropic redacted most of the evidence METR found most informative.

Apple's ICML 2026 paper amortizes maximum inner product search by training small neural networks once, offline, for a fixed key database, learning how to route queries instead of comparing against millions of vectors. The method exploits the gap between short user queries and long curated keys that query-agnostic indexing methods ignore.

Personal Finance connects bank and card accounts via Plaid with read-only access, rolling out in beta this week to Pro, Max, and Ultra subscribers in the US. Claude can see balances, credit limits, and up to two years of transaction history, but can't move, manage, or invest money.

Amazon's A+ Content quality evaluation runs on Amazon SageMaker AI, per an AWS Machine Learning blog post. Selling partners using high-quality A+ Content see 8–20% higher sales than those without.