Search results

I’m surprised that we’re over 1000 comments in and nobody has linked this classic text yet: https://www.joelonsoftware.com/2000/04/06/things-you-should-never-do-part-i/ Tl;dr you should ~never rewrite from scratch, because your messy codebase is just a reflection of the messy business logic you...

Emphasis mine; do you actually not understand what “why” means?

Right but the training data doesn’t represent reality accurately enough. That’s the entire problem. Measures of how accurately the answer reflects the training data can’t solve that. I skipped the first paper, I’m still working through the second one. There’s some interesting insights but...

For certain very circumscribed use cases. If "just look at the output probabilities" was a general solution to output accuracy, the problem would've been generally solved already. The fundamental problem is that the model has no way of evaluating probability WRT the real world, only WRT its...

So, the precision of the training is actually pretty high, but the important thing that sets the actual context for this is that primary training is only measuring if the LLM can reproduce the training data. The standard approach leaves a chunk of that data to one side as "test data" to allow...

I mean, if you're going to start assessing recent Deep Learning terminology for accuracy of metaphor, you're going to find yourself having to rewrite the lexicon from scratch, it's misleading terms all the way down. Explaining "neuron": OK so an Artificial Neural Network is basically a graph...

Operating in a similar conceptual space to this, right?

Happy to be given a less misleading formulation; I'm trying to be precise there (although, OK, not trying that hard). It's also something that I heard like twenty years ago (albeit from someone who had the very specific credentials needed to be making that claim authoritatively), so I don't have...

I know it sounds like magical thinking, but I'm just trying to extrapolate reasonably based on the evidence we have. 10k years of development - and, again, this is likely an absurdly small increment in this context, we're likely talking 100k+ years if we're being remotely realistic, probably 1m+...

Yeah, I absolutely am assuming that, because it seems exceptionally likely that it will be the case. Look back to our capabilities today vs where they were in 8000BC. Could a person from 10k years ago kill a person today? Yeah, of course. Could a civilization from 10k years ago wipe out modern...

Right but if your species reached space travel just the tiniest bit sooner in astronomical timescales - say, ten thousand years - the chances of a “24th century” planetkiller being a meaningful threat to your “1024th century” planet is ~0.

Yeah but only if “similar tech level”, which is (pun intended) astronomically unlikely.

No, exactly, tech advancements happen way faster, so a given time gap is much more significant for technological species. A 10k tech gap is equivalent to hugely larger evolutionary spans. That is the usual human foolishness of believing that we live in a uniquely special time. A much more...

The problem with Dark Forest theory is the same problem with the Fermi paradox, namely it doesn’t account for cosmic timescales. The chances of an up and coming intelligent species posing a future threat to an established species with as little as 10k years of additional technological...

Regarding definitions, this is a generalized opinion that I have, not one specific to this discussion. We can discuss this or any other topic without resorting to words with unclear definitions; it makes it slightly more long-winded, but it aids in precision by forcing participants to explain...

OK, and if you, when summarizing what you thought I meant, had not inserted your own opinions into that summary, I might've had more inclination to be more forthcoming.

You replied to me six times, and not a single one of those posts included a question mark. Yes, a couple of times you asked me what I meant, but I ignored those because a) I'd just explained here and here in very short, simple posts what I meant, and b) I didn't feel obliged to answer your...

Why? Why assume things you don't need to assume? There's a lot of things we need to assume to go about our day successfully, but "the fundamental parameters of the universe are uncorrelated" is not one of them.

I'm deliberately trying to say as little as possible about whether or not LLMs "have understanding" for the reason I laid out in the post you're quoting, here, emphasis added: Specifically, that "do LLMs have understanding in some final, objective, definite sense?" depends on your precise...

If you don't understand my position, feel free to ask me what it is rather than just responding to what you assume I should be saying. "What else I could mean", in this situation, would be literally what I have stated multiple times. I'm beginning to think it's not bad faith, it's just poor...