Caveat: one user, one week, n=1. The residual gap never fully closed.
But the cheap model now asks "what's the cheapest test that could kill this claim โ and has it run?" before telling me something works.
That question was most of what I was paying for.
In the recent week with Claude Fable 5, instead of just using it, I had it run side-by-side comparisons vs Opus on all my real work โ same task to both, then a decision-level diff.
Goal: capture its JUDGMENT, not its answers, in a form Opus can load.
What I learned ๐งต
How to copy this:
โข Distill the diffs into a ~900-word skill file โ only rules that changed decisions
โข Never let a model describe itself (flattering fiction)
โข Demote rules that stop differentiating
โข Keep score both ways โ its falsified hypotheses are in my record too
@mreflow - your mission, should you decide to accept it, is to compare OpenAI and Google I/O's AI in an objective way. I/O seemed better than I expected/the past, but not as good as OpenAI, but, in a few ways unique where OpenAI can't. But I'm probably biased. Looking fwd to it.
@GitKraken Feels like you only figured out users were having problems way after it started. Please inform users that you are working on a solution early on (assuming you were).
Yep, replied to email per caller recommendation. another email saying, no, you need to call, and another agent will help. All this to replace a "known motherboard issue" with Pixel 5a5g, paying a replacement cost, while paying and under warranty.
Hey @Google, your phone support staff really want to help, but keep giving uncoordinated or bad information. i.e. after authorizing overnight, then failed, asking customers to call @FedEx to make it faster isn't allowed. Also email saying call, then calling says email? #ucandobtr