> Before approving construction, I would want communities of humans to understand why the design works and what justifies confidence in its safety. I would hope that we all would.
Until very recently, I pored over every single line of code Claude generated with razor sharp scrutiny. I would usually catch issues with every response. I'm catching fewer problems these days. Maybe the model is just getting better, and maybe I'm being less careful while under pressure to ship more and more often. But model capability is obviously growing. Even back in March, you could tell it "give me a function that adds two numbers" and you could be 100% confident that it would write the correct function. There was almost no point in looking at the code. Since then, the complexity floor of problems in the category "this is so simple that the model couldn't possibly get it wrong" is rising, and with it, my cognitive surrender to the model is increasing too. Why check it? It's obviously going to be correct.
If AI designs a terawatt fusion plant, then of course we're going to meticulously pore over every detail to ensure safety, reliability, efficiency, whatever. If we find no flaws in the design whatsoever, will we be less careful about the second one? The third one? What about the ten thousandth one? Will "a nuclear fusion plant" become something that models couldn't possibly get wrong?
Terence Tao is arguing that the human involvement in research is crucial, but doesn't convincingly justify why, in my opinion. He says that "human agency is a value of fundamental importance" and that we will need to build "thriving human communities that can understand [AI ideas] together" - not for the sake of correctness, which AI may surpass us on, but for, I guess, the possibility of reclaiming human meaning and purpose. I don't disagree with this at all, but it's not an argument, it's a statement of values. Unfortunately, the stark reality is that if AI does surpass humans, it will become the economically dominant strategy to not verify them and not double check them, but to just do whatever they say. This seems like a great way to raise p(doom). But as the models get better and better, and as I'm scrutinizing Claude's output less and less... I just hope that there are more Terence Taos out there than people like me.
Imagine the first time electric calculators calculated the square root of 5. I'm sure people would verify again and again if what the circuits calculated was right.
Then in the 80s, you presses 2 buttons and there you had it in your classroom without thinking twice if the electricity arrived correctly at the transistors.
How crazy will the world be once our [current gen] ANN are like that!
Calculators and computers are deterministic, they give the same output to the same output every time. Language models specifically are not. So it might give you are function that is correct, or a function that is not, or worse yet a function that behaves correctly, but introduces some god-awful bug down the line that can cause serious havoc. It is obvious that they need supervision, not only for output, but also sandboxing and various harnesses for them to not do any “oops, I deleted your codebase sry” kind of nonsense people post to Reddit.
So I think the problem is to determine which problems under what instructions we can safely give to a model application to solve and how we test the output for safety and functionality. This would create more usable and safe, albeit a bit more boring, AI-based applications alin to a calculator or general computer. Whether this is posswith current model architecture is another thing.
Memory bits flip randomly. It's not a super rare thing either. You and me have experienced that many times without knowing. The only reason that computers feel deterministic is that we have error-correcting code to fix that. But in the most extreme cases, when multiple bits flip together, once "deterministic" program can generate unexpected output.
So why do you trust computers? Because statistically the case is just very unlikely. Therefore if AI is statistically unlikely to make mistakes there is no reason to not trust them.
You are confusing hardware noise and whether the algorithm itself is verifiable as if parity bits were not a thing. We don’t trust computers because the medium itself is infallible, but because the logic is sound.
With statistical models - such as LLM’s - there is no logic as such, but statistical assumptions based on given data. The output can ge very good or very bad, but you are fool to trust it blindly. Therefore you need a deterministic way to verify, whether meat- or software-based.
A program can only be as deterministic as the hardware it runs on. If you have a turing machine whose symbols on the tape just (truly) randomly change every step, then none of programs on this machine is "deterministic."
> meat-
There is no meat-based deterministic way to do anything yet yet. You're talking sci-fi.
pyridines · · focus · HN ↗
Until very recently, I pored over every single line of code Claude generated with razor sharp scrutiny. I would usually catch issues with every response. I'm catching fewer problems these days. Maybe the model is just getting better, and maybe I'm being less careful while under pressure to ship more and more often. But model capability is obviously growing. Even back in March, you could tell it "give me a function that adds two numbers" and you could be 100% confident that it would write the correct function. There was almost no point in looking at the code. Since then, the complexity floor of problems in the category "this is so simple that the model couldn't possibly get it wrong" is rising, and with it, my cognitive surrender to the model is increasing too. Why check it? It's obviously going to be correct.
If AI designs a terawatt fusion plant, then of course we're going to meticulously pore over every detail to ensure safety, reliability, efficiency, whatever. If we find no flaws in the design whatsoever, will we be less careful about the second one? The third one? What about the ten thousandth one? Will "a nuclear fusion plant" become something that models couldn't possibly get wrong?
Terence Tao is arguing that the human involvement in research is crucial, but doesn't convincingly justify why, in my opinion. He says that "human agency is a value of fundamental importance" and that we will need to build "thriving human communities that can understand [AI ideas] together" - not for the sake of correctness, which AI may surpass us on, but for, I guess, the possibility of reclaiming human meaning and purpose. I don't disagree with this at all, but it's not an argument, it's a statement of values. Unfortunately, the stark reality is that if AI does surpass humans, it will become the economically dominant strategy to not verify them and not double check them, but to just do whatever they say. This seems like a great way to raise p(doom). But as the models get better and better, and as I'm scrutinizing Claude's output less and less... I just hope that there are more Terence Taos out there than people like me.
xtracto · · focus · HN ↗
Then in the 80s, you presses 2 buttons and there you had it in your classroom without thinking twice if the electricity arrived correctly at the transistors.
How crazy will the world be once our [current gen] ANN are like that!
What an amazing thought.
delis-thumbs-7e · · focus · HN ↗
So I think the problem is to determine which problems under what instructions we can safely give to a model application to solve and how we test the output for safety and functionality. This would create more usable and safe, albeit a bit more boring, AI-based applications alin to a calculator or general computer. Whether this is posswith current model architecture is another thing.
raincole · · focus · HN ↗
How do you know?
Memory bits flip randomly. It's not a super rare thing either. You and me have experienced that many times without knowing. The only reason that computers feel deterministic is that we have error-correcting code to fix that. But in the most extreme cases, when multiple bits flip together, once "deterministic" program can generate unexpected output.
So why do you trust computers? Because statistically the case is just very unlikely. Therefore if AI is statistically unlikely to make mistakes there is no reason to not trust them.
delis-thumbs-7e · · focus · HN ↗
With statistical models - such as LLM’s - there is no logic as such, but statistical assumptions based on given data. The output can ge very good or very bad, but you are fool to trust it blindly. Therefore you need a deterministic way to verify, whether meat- or software-based.
raincole · · focus · HN ↗
> meat-
There is no meat-based deterministic way to do anything yet yet. You're talking sci-fi.