Transcription
Something satisfying just happened in AI. A corporation got caught and the receipts are so real. An AMD engineer named Stella Lorenzo run Claude code across 50 concurrent agent sessions for months.
[music]
Systems-level programming. She wasn't guessing. No, she wasn't logging. When things started to feel off in February, she didn't post it anything negative. She pulled the data. 17,871 [music] thinking blocks, 234,762 calls, and 6,852 sessions.
What she found between late January and early February, Claude's medium thinking depth was around 2,200 characters. By late February, it dropped to 720. And even by March, 560. That's basically a 75% decline.
And here's what it actually means. When thinking depth drops, the model stops researching before it edits. >> [music] >> It skips reading the file. It jumps straight to the changes. The behavior shift is measurable. It went from research first to edit first.
She filed it as a GitHub issue on the official Claude code repo. It hit 790 points on Hacker News. [music] Then Boris Cherny, the creator of Claude code, showed up in the thread. His explanation, "Two features shipped in February changed the default thinking allocation. Adaptive thinking and a new effort level set to medium instead of high."
A setting change that they didn't announce. Users are calling it AI shrinkflation. Same price, less depth. Whether you agree or not, the data is public and the methodology is well documented. That's how accountability should work.
Follow me to stay ahead of what's actually happening in AI.