l'empathie mécanique @dpwiz@qoto.org

@tristanC The types are sine qua non

@marick It is even simpler than that: "Survival of the fit enough".

Whatever else is going on, you only need to meet the bar of being viable. No need to waste energy to reach the apex in all of the environments. You can put it to make a niche of your own instead. Or not, the evolution wouldn't judge.

@ericflo This.

Responses API are a bit cleaner and nicer to use. But everyone else is stuck on completions and we can't have nice things )=

@jedbrown @dpiponi This is true only for the specific setups of a model, training, and sampling. Unless you hit some idiosyncrasies the amount of follow-up questioning needed to distinguish understanding can be made arbitrary high. This method will "detect away" human understanding before the LLM-based system breaks.

@wilfredh Maybe we're just getting used to that. AutoTune in mass-music, compression artifacts in codecs and headphones...

@boilingsteam Made my heart miss a beat.

@nikitonsky No, it's okay. It works crappy too.

Expired: Ancestral simulations.
Transpired: Stakeholder simulations.

@reidrac I've been doing hours with Claude et al for about a year by now and not sure I can stand doing 100 hours more. I get physically sick from reading all that crap.

In that year I went from "nah, LLMs are useless" to "huh, they can do some JS for me", to "fucking christ on a stick! just leave my code alone ffs".

@drmakimber @BartoszMilewski Spiced with a "logarithm is just the number of zeroes" log-odds make quick takes rather easy.

@wilfredh On a border (=
But you'd have to play on it somehow.

Maybe it could be done in more ways, but rn the block design is almost locked down by the paper-and-shadow visual metaphor. Maybe you can find something more flexible so you can enjoy more decision freedom later.

@L29Ah ну или забить болт и генерить сразу жсон

@L29Ah что-то такое нужно: https://hackage-content.haskell.org/package/autodocodec-schema-0.2.0.1/docs/Autodocodec-Schema.html#t:JSONSchema

@L29Ah ты ж небось хочешь с генериков жсон-схему генерить. Да даже если и просто руками, то всю схему придётся затаскивать.

I've almost procrastinated writing this episode down because the most expected outcome of it would be "the post is flushed down the social stream with only a few eyes to see it ever". And it was kinda pre-disappointing.

But I couldn't stop thinking it through and it still kinda writes itself in my head.

Meanwhile I recall that I had quite a reach for one of my works this year, so this isn't just a numbers game. And the problem of doing *anything at all* is actually one of the challenges in that post-human adventures (thankfully explored in Permutation City, Diaspora, 31 laws of fun, etc).

So, challenge... not accepted. Not yet. Something is missing still. So I just browsing idly on the viability FP32 precision for modelling orbital maneuvers and...

...

Unfortunately I can't remember how I managed to actually start typing, sorry.

Perhaps having a Fediverse client on a muscle memory and the learned anticipation of randomized reward from reading it (keep up the good work, all y'all!) made me open it. And then a typing box helpfully focused in a client was presented in a prominent position. I'm not sure, but whomever designed that - I'm grateful to them.

This IMO brings some evidence against efficacy of using Skinner boxes to alleviate low-energy mood swings.