AI of the morning
2025-02-25 19:24:47.589278+01 by Dan Lyke 0 comments
The Deep Research problem — Benedict Evans. On trying to get useful information out of ChatGPT, using OpenAI's examples:
We’re asking for a deterministic answer from a probabilistic question, and there it looks like the model really is failing on its own terms. In my opinion, or given my expertise, it shouldn’t be using Statcounter or Statistica, but even if it should, it hasn’t taken the correct number from them.
Every: I Created a Hacker News Simulator to Reverse-engineer Virality
Given a more detailed persona based on a two-hour interview with a human, ChatGPT can replicate their answers on surveys with as much as 85 percent accuracy.
Uh. Yay?
Back to Benedict Evans, in his summary he notes:
... OpenAI and all the other foundation model labs have no moat or defensibility except access to capital ...
Emphasis mine. With Microsoft dialing back on data center leases, maybe we're starting to see the bubble deflate.