solveit

The bright side is that it should eventually be technically feasible to create much more powerful and effective guardrails around neural nets. At the end of the day, we have full access to the machine running the code, whereas we can't exactly go around sticking electrodes into everyone's brains, and even "just" constant monitoring is prohibitively expensive for most human work. The bad news is that we might be decades away from an understanding of how to create useful guardrails around AI, and AI is doing stuff now.

I get the feeling that you're winging the specific numbers because they're spectacularly incoherent.

But anyway, the United States is extremely rich and has essentially no big problems that can be solved by a small amount (say, a few billion) of money. The problems are either so big that it would take trillions to solve (supporting aging population etc), or blocked by something other than money (politics, regulations, etc). The big problems that can be solved just by throwing a few billion at them are solved quite easily by either the government or by private entities like the Gates Foundation.

Even the US median household income is "only" $83k. Looking at stuff like this + the rest of the blog I'm not convinced the author is any less out of touch than the people this post is criticizing.

There's nothing wrong with being in debt per se, and in particular, (federal) student loans and mortgages are loans rigged in favor of the borrower.

Do you recall which paper it was? I would be interested in reading it.

It's frustrating how far you can go out of your way to avoid being associated with such superficially similar tropes and still fail miserably. Yudkowsky in particular hated that he couldn't get a discussion without being typecast as the guy worried about Terminator. He hated it to the point he wrote a whole article on why he thought Terminator tropes were bad (https://www.lesswrong.com/posts/rHBdcHGLJ7KvLJQPk/the-logica...).

As a side note:

> any serious attempt at discussion gets bogged down by [...] without taking a single shower in the same span of time.

This is unnecessary and (somewhat ironically) undermines your own point. I would like to see less of this on HN.

> I'd argue that the biggest reason machines are black boxes are because no one is bothering to look inside of them.

People do look, but it's extremely hard. Take a look at how hard the mechanistic interpretability people have to work for even small insights. Neel Nanda[1] has some very nice writeups if you haven't already seen them.

[1]: https://www.neelnanda.io/mechanistic-interpretability

Error correction is possible even if the error correction is itself noisy. The error does not need to accumulate, it can be made as small as you like at the cost of some efficiency. This is not a new problem, the relevant theorems are incredibly robust and have been known for decades.

The previous poster was actually inviting you to expound on the other reasons.

> Math fans tend to discount and dismiss applied statistics as being not math, in a way that they don't do for physics, for some reason I don't fully grasp.

It's because we're secretly afraid that the physicists are smarter than us.

Less facetiously, physicists keep discovering things that lead to new mathematics we would never have dreamed of ourselves, so we have a healthy respect for how insightful they can be.

Might be the other ingredients reducing inflammation, widening your clogged pipes and letting stuff drain.

I think mathematicians looking at that expression would curl their lips in mild disgust, but also evaluate it to True.

Depends how the flow goes. It's reasonable to only check the diagonal when one player has actually played on the diagonal, in which case this is fine.

I like good handwriting, but good, fast handwriting takes thousands of hours of practice. We used to spend a good chunk of school drilling it into kids, but now it's really hard to justify everyone spend that kind of time when we can technology our way around it and there are other valuable skills to learn.

We really have no way to know this. But I would be very surprised if modern chess engines didn't regularly blunder into losing (from the perspective of a hypothetical 32-piece tablebase) positions, and very very surprised if modern chess engines perfectly converted tablebase-winning positions.

Practicality of reproducing training runs that cost tens of millions aside, it's hopeless. Determinism is hard enough with a single GPU, fixing a seed isn't going to be much help when training is distributed across hundreds of GPUs.

It's relatively trivial to get an LLM that does that and every big lab has one, even if they're not selling them.

ChatGPT 4o as of right now just runs python code, which I guess is "Let me get my calculator", see https://chatgpt.com/share/670df313-9f88-8004-a137-22c302f8bf...).

Claude 3.5 just... does the multiplication correctly by independently deciding to go step-by-step (don't see a convenient way to share conversations, but the prompt was just "What is 1682671* 168363?").

Or you're using a log scale because your effects span many orders of magnitude, or whatever. The specifics are not particularly central to my point beyond that it isn't too hard to wrangle up some scenario where the equation says something completely different.

If at the end of the day you have to compare wildly different things, you have to compare wildly different things. Usually this will involve making a judgment call about how to weigh these things, or in other words, selecting units for the purpose of adding quantities of wildly different units.

Only if they are measured on a scale allowing negative values for outcomes and sufficiently high values for service. It would be unsurprising to have outcomes measured from zero and service topping out at 100%. The equation doesn't say anything remotely precise, it's just a bundle of vibes.

I'm not quite sure how exactly to interpret your question, but untreated type 1 diabetes usually kills you within five years of diagnosis.

> It's not perfect, but the friction helps prevent drive-by bad behaviour.

> No it doesn't.

There's really no use in continuing this discussion when one party is unable/unwilling to use precise language to discuss marginal effects. Obviously I presume what you mean is that the marginal effect is too small to be relevant, but discussions with people who round that off to "No it doesn't" rarely go anywhere productive.

Most projects with an active user-created mods community are heavily influenced by them.

Stack the vectors up so it's a matrix and take the determinant. The sign tells you which one it is.

Rest in peace.

> Pedantically, it's that the bumps grow exponentially narrower.

Has to be smaller in both directions, right? It has to be vertically smaller to allow f to be computable, but horizontally smaller to keep f' large.

> IIUC, a simpler (but more nitpickable) version would be...

Here I think your f as defined is identically zero, and so you don't get f' by differentiating.

> A much simpler version is f'(x) = {-1 if x<0; +1 if x>0; Chaitin's constant if x=0}. f(x) = abs(x).

Similarly, differentiating abs(x) doesn't get you that function.

But I see where you're coming from, and I agree this doesn't tell us anything profound about computability. That said, it is a cute result that differentiable functions are flexible enough to admit this kind of construction, and that's straightforward but not obviously trivial, as your attempts somewhat ironically show. I think Myhill's proof is pretty close to a minimal working example and trying to fix your examples would result in something that looked very similar.

> how is it possible that there can be a continuous computable function whose derivative is uncomputable?

For others: The original paper is short and very readable - https://projecteuclid.org/journalArticle/Download?urlId=10.1...

The intuition seems to be that computability of a function is actually approximability, and derivatives can be large even when the function is very small, by having the large slope confined to a very small area. So by constructing a function f that contains an uncomputable (but recursively enumerable!) set encoded in the derivative by strategically placing little bumps, but having the bumps grow exponentially smaller, Myhill was able to construct a function that was easily approximable but whose derivative wasn't.

The key detail here is that the bumps grow exponentially smaller in the order the uncomputable set appears in its enumeration. This allows Myhill to construct a sequence of computable functions that (uniformly) approximate f to within 2^(-n). If the bumps weren't ordered in this way, we wouldn't know how far down the sequence to go to approximate f to within 2^(-n). The ordering lets us just look at the first n elements of the uncomputable set.

Why doesn't this allow us to approximate the derivative? Because at any x, we don't know how well we have to approximate f in order to approximate f'(x), so we need knowledge of the entire uncomputable set, which we cannot have.

If safety is cheap but overcoming culture is expensive, at some point it becomes misleading (wrt ethics of participants, not correct course of action) to say the problem is that management doesn't care enough to spend money on safety, even if management is the only lever we have to fix the issue.

Sorry to hear that. Thanks for sharing.

What went wrong?