Analysts debate whether AI compute costs are rising or falling

A Reddit user documented a real quality downgrade in GPT-6 Astra, which OpenAI acknowledged and partially fixed. The poster argues downgrades happen more often than users notice and are anti-consumer.

Matthew Berman's video examines DeepSeek's performance on a Rubik's Cube test, where the model fails the task.

ARC Prize says ARC-AGI-4 will benchmark autonomous open-ended innovation, keeping the effort open-source as a shared research target. The announcement notes humans still significantly outperform AI at open-ended tasks.

A Reddit user released smolbenchmark, a leaderboard for models that fit in 8GB of memory, ranked by decode speed, tokens per joule, and heat. It targets tablets and other low-power local hardware rather than GPU servers.

Altman told Fortune's Alyson Shontell that going public now would be "ill-advised" given safety concerns, saying "not 2026, yeah. We've got a lot of stuff to do." OpenAI has already filed confidentially for an IPO.

Dwarkesh Patel's video explores the convergence of AI model outputs, arguing that models from different labs increasingly produce similar responses. No specific models, benchmarks, or figures are cited in the available source material.

Quartermaster installs codebase-mapper, builds the project map, adds starter live-rules, and picks stack plugins and permissions, with every install, file edit, and settings change awaiting user approval. It then watches sessions, proposes one change at a time, and rolls back changes that didn't help.