Hacker News Reader

Comment by codingwagie | Hacker News Reader

“This may turn out to be the most consequential fact about all of history so far. It is possible that we will have superintelligence in a few thousand days (!); it may take longer, but I’m confident we’ll get there.”

I am a believer that people like sam are not lying. Anyone using these models daily probably believes the same. The o1 model, if prompted correctly, can architect a code base in a way that my decade+ of premium software experience cannot. Prompted incorrectly, it looks incompetent. The abilities of the future are already here, you just need to know how to use the models.

View on HN

I'm using these models daily, and I don't believe that they're a direct path to superintelligence (unless you'd consider something like the printing press to have been a direct path to, say, the integrated circuit or the Internet).

> Prompted incorrectly, it looks incompetent. The abilities of the future are already here, you just need to know how to use the models.

Something purportedly intelligent shouldn't need "correct usage", as it should arguably be able to infer and clarify all ambiguities itself, no?

The model's arent there yet. I am confused how people cannot extrapolate into the future and understand that the models will improve

And I'm surprised how many people think they can confidently tell whether they're looking at an s-curve or an exponential function based on very limited data points. I don't even doubt that superintelligence is a very real possibility! But it might or might not happen, and if it does, it might or might not be based on deep learning.

As a counterexample: The maximum speed of travel for the average person for millenia used to be as fast as they could run, then it was as fast as the fastest horse can run, and then within a century it has accelerated to almost the speed of sound – at which it has plateaued.

Looking purely at the decades of acceleration, you might have very well concluded from the data that we'd be making significant headway towards getting within double-digit percentages of the speed of light at this point.

Maybe don't think of it as curves or functions. Just go through an LLM and think about all the things that could be improved. It's a long list once you get into the details. By and large...sota models are:

a) trained on crappy data, including questionable RLHF feedback. b) trained with questionable embedding layers. c) trained with questionable loss functions d) trained with questionable optimizers e) trained at questionable precision (somewhat related to d) f) are very big which stops fast iteration around all the above.

It's kinda like semiconductors. You don't have to think of it as a curve - just ask people who are really close to them and they'll have a laundry list of stupid stuff which is currently done and will likely be improved upon over time.

I don't doubt that there's tons of work ahead to even just integrate current-day capabilities of LLMs into our society and economy.

But when talking about future growth potential, I don't think you can get around making assumptions about the shape of the growth function.

You can't write a nice article about it. But you can talk about it with a simple "things are going to get a loooot better". The problem with thinking about progress in functions is it suggests some underlying law exists, and there isn't one.

Even if someone could point to a function and say "10x better" by 2030 - what does that even mean in the context of an LLM for example?

https://xkcd.com/605

Interesting observation. My experience with o1 has been much more mundane. Sometimes I get the response I wanted, sometime it hallucinates, often times it writes buggy code. I've been experiencing this since ChatGPT was first released.

I actually originally wrote off the o1 model. Another thing I have found it's good at is finding bugs in a ton of code. Give it ten coding files, and a stack trace, it can find the bug.

Reply on news.ycombinator.com