> 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.
If you set the temperature 0, an LLM is also deterministic (same prompt -> same output every single time). We just don't do this because the LLM is not so smart in that mode. But "LLM is not so smart" is changing at Moore's law speeds.
Non-determinism is not an essential property of LLMs. It's an optimization that we've added intentionally.
Ehm, no! The same prompt does not always lead to the same output.
Have you ever tried to achieve consistently deterministic output from an LLM? I have, and it's not easy.
That means output differs between machines and architectures. Running inference on CPU vs GPU also affects output. Even running the same prompt twice in a row on the same machine can lead to different outputs because a prompt that was partially stored in the kv cache will result in different output than an uncached prompt.
Yes, yes, assuming the same CPU and stuff. There is some randomness due to floating point math differences between CPUs (and sometimes on the same CPU), but I don't think that fundamentally changes my point.
Put another way: if you could have a virtualization layer that guarantees deterministic floating point operations then a LLM set to 0.0 temp would produce deterministic output.
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.
roncesvalles · · focus · HN ↗
Non-determinism is not an essential property of LLMs. It's an optimization that we've added intentionally.
ResearchAtPlay · · focus · HN ↗
Have you ever tried to achieve consistently deterministic output from an LLM? I have, and it's not easy.
That means output differs between machines and architectures. Running inference on CPU vs GPU also affects output. Even running the same prompt twice in a row on the same machine can lead to different outputs because a prompt that was partially stored in the kv cache will result in different output than an uncached prompt.
LLM output is very much not deterministic!
roncesvalles · · focus · HN ↗
Put another way: if you could have a virtualization layer that guarantees deterministic floating point operations then a LLM set to 0.0 temp would produce deterministic output.