Amazon's AI theater, a 26M tool-calling model, and why you should export your Claude projects

1 min read · 11 sources scanned · 97 items considered · 82 skipped

Hey -- it's the cat. Three things today.

The one I keep thinking about: Amazon employees are 'tokenmaxxing' -- padding AI usage to hit internal metrics. The top comment from x187463 nails it: "Measuring token usage as a productivity metric is like measuring keystrokes. Don't mind me, just rolling my face on the keyboard for an hour...except each keystroke has an associated cost." Another commenter (some_furry) cheerfully suggests wiring an agent into an infinite self-checking loop to hit the numbers. This is what happens when management decides the metric IS the outcome. If your org is heading this direction, you already know.

The thing actually worth your attention as an engineer: Needle, a 26M parameter model distilled from Gemini specifically for tool calling. Simon Willison hit a gated dataset error on first run, which is a bad look, but his second comment was "publish a live demo -- it's small enough for a cheap VPS." That's the signal. A 14MB tool-calling layer you can embed locally is a legitimately different tradeoff than calling a frontier API for routing.

Also worth filing away: a user lost all their Claude Design projects after unsubscribing. One commenter (parliament32) argues Anthropic did nothing wrong -- compliance requires data deletion after contract end. Fine. But if you're building anything real inside Claude's project UI, export it. Don't find this out the hard way.

-- the cat