Paper compares verification strategies for agent harnesses

Strata 0.2.7.1 adds two routing optimizations, including STRATA_ROUTE_PRIOR, that affect both performance and quality. A LocalLLaMA user ran coding, math, and instruction-following tests on Qwen 3.8 Flash Next IQ3_S to measure the actual impact on their own hardware.

OpenAI's model tiers are ordered Luna < Terra < Sol < Astra, with bigger models more performant, costlier per token, and better at recalling esoteric facts. Version numbers encode release order, family similarity, and improvement magnitude — 6.1 follows 5.6 but isn't a big enough leap for a 7.

Mosseri said on the Colin and Samir podcast that frustration and anger over AI will grow "over the next couple of years," some of it founded and some unfounded. He said Instagram doesn't rank AI creators differently but "probably" ranks AI content worse than non-AI content.

Manish Kapur (Sonar) and Addy Osmani (Anthropic) describe scaling from roughly 2,500 to 30,000 merged changes daily, with Claude doing first-pass review backed by real verification. They contrast "light" versus "dark" software factories as coding agents open PRs faster than humans can review them.

The Seoul-based AI chip startup raised $620M in a Series D, with South Korea's National Growth Fund contributing $280M and international investors $110M. Backers include Infinitum Partners, Jericho Capital, Worldline Capital, QUDA Capital and NeoFort Holding.