Toots from 2026-04-09#
There’s a tension between a model’s bias towards more code-level engagement and the organizationa…#
There’s a tension between a model’s bias towards more code-level engagement and the organizational need to consider some code checkpoint “shippable”. This isn’t a new problem, but I suspect there will be a renewed push for “just one more change” given symbol production costs.
This could be mitigated by a platonic continuous delivery stream of releases (variants?), but that requires a level of engineering rigor and discipline than the models don’t typically encode. And even if that were true, could you automatically put a production mutex on a variant until some statistical threshold were reached? Imagine hooking up your CI pipeline gates to your nightly DataWarehouse reports.
It also would that your software be forward compatible and your customers be risk-tolerant.
Mastodon Source 🐘#
“All of these are examples of the second argument I want to make. The failures of these models ar…#
“All of these are examples of the second argument I want to make. The failures of these models are getting significantly more consequential. The model misbehaves less often than its predecessors. But when it does, the misbehavior is more sophisticated, harder to detect, and embedded in more capable actions.”
https://www.gettheleverage.com/p/the-training-wheels-are-coming-off?__readwiseLocation=
Mastodon Source 🐘#
My new work laptop is on the way! Really looking forward to starting on Monday. Great people, gr…#
My new work laptop is on the way!
Really looking forward to starting on Monday. Great people, great opportunities.