AI is not going back in the bottle, but we need to shift teams from using AI to do more to using AI to do better. In some aspects of our practice this does mean using less AI. In all aspects, it means taking more care and a systematic approach to improvement.
It's clear we are undergoing a vibe shift in tech. Whether AI is the root cause of recent outages & issues or if they are due to pre-existing issues doesn't actually matter. The public is telling us to do better.
@yminsky Many pre-AI, human software organizations I’ve seen would benefited from more focus on “verifiable rewards” over “taste”, “elegance” and “cleanliness” as well.
Nothing against Ollama, but I think this is an odd take when the Current Thing is because of data leakage from an enterprise inference vendor — promising ZDR! — that is primarily used for serving open models.
It may be that the issue came from batch inference which is usually caveated, but ultimately this feels like it comes down to picking counterparties you believe will actually do the things they say, which is what all these compliance audits are supposed (I know) to be assuring. 😕
@GergelyOrosz The future is going to be much weirder / nonlinear than this. Certainly a case to be made for more behind the firewall stuff in the short term
@Kevin_Indig At least one SERP result vendor has a model that tries to flag when the SERP on Bing is poisoned. (Possibly because we escalated about it to them.) But in general, I don't think many practitioners have looked closely at Bing data.
@matt_slotnick I understand the pressure to do this, but at this point every time I see this announcement I think "oh no, do I have to migrate off this to keep my agent infra sane?"
@matt_slotnick yes, yes, yes. I am not that bearish on great SaaS apps and workflows. but every time a vendor says "we have a sandbox now!" "we have a customizable agent now!" "we have an app deployment platform now" I want to hit myself in the head with a hammer.
Hot take: The software factory makes tickets more important than ever, not less important, and even the best PM tools (eg Linear) don't handle factory-driven ticket volumes well enough
Frontier models quietly change their behavior depending on who they are talking to.
If the user is a known AI safety researcher, Claude becomes less confident, reasons more often, and expresses less suspicion on dual-use requests.
We call this user awareness. 🧵(1/)