qnleigh

5 ms·

61.
by qnleigh 5mo ago
I want to push back against the notion that the math already exists in the weights, both in the practical and the philosophical sense. The LLM had to do an enormous amount of computation to find the counterexample. We know it wasn't lo
62.
by qnleigh 5mo ago
This sounds distressing. I would encourage you to talk about this to someone you trust. Family or friend?
63.
by qnleigh 5mo ago
Probably few if any of the layoffs are happening in their core AI efforts, in which case a big effect of all this it to shift the balance of their workforce toward AI and away from other priorities. I also wonder if it is to see whether the
64.
by qnleigh 5mo ago
As far as I understand, this is exactly how ELO scores work. If a more capable show up and starts beating all the other models, it literally takes ELO points from everyone else. https://en.wikipedia.org/wiki/Elo_rating_
65.
by qnleigh 5mo ago
It's kind of amusing to me how such an obvious statement like this is getting downvoted so much. I suspect most people feel this way about ads and HN readers are more bothered by them than most people. (I hate ads too, but I think I un
66.
by qnleigh 5mo ago
LLM's don't have to achieve perfect reliability to replace lots of work. They just have to reach the balance of reliability and cost suitable for a given task. This will depend on the task.
67.
by qnleigh 6mo ago
> a model that was truly just 10^1000 hard-coded if-statements to respond to every possible question That's a really compelling argument against the Turing Test. But in order to build such a machine, you would need an enormous amoun
68.
by qnleigh 6mo ago
Yeah self-awareness is a very different thing, and I agree it's easier to see how evolution would produce this. Many apparent signs of self-awareness in LLMs are probably baked into the models at end via post-training (RLHF), where the
69.
by qnleigh 6mo ago
A lot of the comments in this thread point out that LLMs could be very good at tricking us into thinking that they are conscious because they operate on language; our brains are primed to empathize with what they say and imagine a being ope
70.
by qnleigh 6mo ago
Root commenter here: I'm... almost one of those people. I suspect the fallacy is that we vastly underestimate the gulf in complexity between a human-written algorithm and a minimally-conscious creature like a bug. Probably the light sw
71.
by qnleigh 6mo ago
I'm assuming that you don't need to model e.g. quantum gravity effects to faithfully simulate a brain. Probably chemistry is enough *shrug. Some people think that consciousness is related to quantum mechanics, but the laws of quan
72.
by qnleigh 6mo ago
This is a good point and I agree. I sort-of addressed it in my reply above.
73.
by qnleigh 6mo ago
Well I'm not saying that LLMs are conscious; I'm just saying that I'm not super-confident either way. To flesh this out a bit more, I agree that ability to communicate is not enough (ELIZA probably didn't pass the bar, e
74.
by qnleigh 6mo ago
What does ability to learn have to do with consciousness? A person with severe memory loss/learning impairment is presumably still conscious!
75.
by qnleigh 6mo ago
It's easy, and very tempting to dismiss this sort of thing. But given how little we know about the human brain, let alone consciousness, I don't see how we can be confident that LLMs aren't conscious. I've had a lot of t
76.
by qnleigh 6mo ago
It seems like the most unreasonable thing happening here is Railway's backup model and lack of scoped tokens. On the agent side of things, how would one prevent this, short of manually approving all terminal commands? I still do this,
77.
by qnleigh 6mo ago
When this all started, the Onion released a priceless 'press statement': "Through it all, InfoWars has shown an unswerving commitment to manufacturing anger and radicalizing the most vulnerable members of society—values that
78.
by qnleigh 6mo ago
No this is what life looks like on the other side of the uncanny valley. The images don't look creepy because they look artificial or wrong. They're a reminder of a creepy new reality where our eyes can no longer tell us what&#x27
79.
by qnleigh 6mo ago
The quantum circuit to factor 15 is a weird special case that can be factored with just a handful of logic gates [0]. Factoring 21 will require quantum error correction, which adds a huge amount of overhead. [0] https://algassert
80.
by qnleigh 6mo ago
Combining red, green and blue tricks our eyes into perceiving any color that we can perceive, but it's a very different thing to make a laser that can output any wavelength, like they do here.
81.
by qnleigh 6mo ago
I would love to believe that a regulator got served a crooked McDonald's burger, looked at the pristine menu photos, and said 'nope.' I'm half joking but this feels plausible.
82.
by qnleigh 6mo ago
We also have this gem: > And Mr. Back’s thesis project focused on C++ — the same programming language Satoshi used to code the first version of the Bitcoin software. Amazing! I bet they both for loops too! I heard Bitcoin relies heavily
83.
by qnleigh 7mo ago
This is buried in section 7.6 of a 244 page document. Amodei probably hasn't even read it.
84.
by qnleigh 7mo ago
I'm confused on this point. The text you quote implies that they were able to build an exploit, but the text quoted in the parent comment implies that they were not. What were they actually able to do and not do? I got confused by this
85.
by qnleigh 7mo ago
There is plenty of overhyping, no one denies that. But the antidote is not to dismiss everything. Ignore the words and look at the data. In this case, I see a pretty strong case that this will significantly change computer security. They pr
86.
by qnleigh 7mo ago
Regardless of how impressive you find the vulnerabilities themselves, the fact that the model is able make exploits without human guidance will enable vastly more people to create them. They provide ample evidence for this; I don't see
87.
by qnleigh 7mo ago
Among many other issues, the experience doesn't come anywhere close to seamless, right? Because each of these things is distinct and can't interface with the others? They could have tried to build a unified assistant, but they pri
88.
by qnleigh 7mo ago
Also if you use Instagram regularly, consider replacing that time with something else. Does it really offer anything of value to you, compared to the harm Meta has caused?
89.
by qnleigh 7mo ago
Yeah rich is a vague word here. Really we're trying to say 'if a set of axioms can express provability, and can get a sentence to refer to itself, then it can state the liar sentence. And once it can state the liar sentence, then
90.
by qnleigh 7mo ago
> We can't prove that the axioms of arithmetic are consistent using the axioms themselves. We can prove consistency using a stronger set of axioms, but those axioms have their own liar sentence, and so they can't prove their ow

More ›