As a child of the 90’s, I’d rather have her smoke a cigarette than vape.

But agree it’s not helpful when Kevin Scott says “Despite what other people think, we're not at diminishing marginal returns on scale-up…I try to help people understand there is an exponential here.” arstechnica.com/information-...

I have no inside information, only what @natolambert.bsky.social wrote in that post:

The nuance I was trying to capture is this: www.interconnects.ai/p/scaling-re...

Fair point.

Isn’t the claim that the pretraining perplexity should be exponential, which doesn’t necessarily imply that the task evals should also be exponential?

For years I would practice “calendar zero” and delete all my past events as they happened. But I stopped because my team constantly harangued me about all those emails.

To be clear, I agree with you that the definition is foggy. I think one difference is that a true agent, in addition to “going off”, isn’t following a predetermined plan. It’s like a cron job or set of API’s, but once it “goes off” it’s determining its own plan for how to use its tools via the LLM.

I’m not so sure—I think the idea is it that goes away from your attention, which is really important, and then resurfaces when it needs something or is done.

I’ve been using Firecrawl and LLM’s to build a searchable dataset of childcare programs offered by Chicago Public Schools. Really fun side project!

www.cps-care.info