The AI release flood meets its own ceiling
Latest Processing Gap
Capability sprawls outward while reliability contracts inward. The models can do more things and hold less of any one thing across length — the machine version of the cozy-web retreat. Both confess the same limit: the whole no longer fits in the container.
Appearances
June saw a near-simultaneous wave of flagship models — Claude Opus 4.8, Google's Gemini 3.5 line, reports of the next GPT — an unusually dense cycle. In the same window, a study found that leading models degrade sharply on a classic attention test *as tasks grow longer and more complex.* The two stories are circulating side by side on Hacker News and in the trade press without anyone quite naming that they are the same story.
Capability sprawls outward while reliability contracts inward. The models can do more things and hold less of any one thing across length — the machine version of the cozy-web retreat. Both confess the same limit: the whole no longer fits in the container.