tmnvdb

It seems easy to worry about somewhat far-fetched scenarios like this when one is not paralyzed and thus not making a trade-off between risks and a fully paralyzed life.

This seems like a bizarre claim on the surface, see also my other message above.

https://epoch.ai/data/ai-benchmarking-dashboard

Really now. I think that deserves a bit more explaination, given the cost per token has dropped by several orders of magnitude, we have seen large changes on all benchmarks (including entirely new capabilities), multimodality is now a fact since 4o, test time compute with reasoning models is making big strides since o1.... It seems on the surface a lot is happening. In fact, I wanted to share one of the benchmark overviews, but none include ChatGPT 4 anymore since it is totally not competitive anymore..

Nobody has certainty about the future. We can only look at what seems most likely given the data.

Why do you think this is a fundamental hurdle, rather than just one more problem that can be solved? I dont have strong evidence either way, but I've seen a lot of 'fundamental unsurmountable problems' fall by the wayside over the past few years. So I'm not sure we can be that confident that a problem like this, for which we have very good classic algorithms, is a fundamental issue.

Compute power increases and algorithmic efficiency improvements have been rapid and regular. I'm not sure why you thought that Back to the Future was a documentary film.

People somehow have expectations that are both too high and too low at the same time. They expect (demand) current language models completely replace a human engineer in any field without making mistakes (this is obviously way too optimistic) while at the same time they are ignoring how rapid the progress has been and how much these models can now do that seemed impossible just 2 years ago, delivering huge value when used well, and they assume no further progress (this seems too pessimistic, even if progres is not guaranteed to continue at the same rate).

Isn't this Musk offer just for the nonprofit part, i.e. not a comparable number?

From scrolling though his articles he is a guy who likes to write polemic articles declaring companies dead

Interesting stuff. As the authors note, using latent reasoning seems to be a way to sink more compute into the model and get better performance without increasing the model size, good news for those on a steady diet of 'scale pills'

Another sure sign of this is custom internal tooling for things that are not business specific like time tracking and report generation.

A company I worked at had an IT budget of 5000 euro for 120 people (90% SWE). The logic was: if we buy a commercial solution for time tracking our IT budget will double. Instead super senior engineer Y is maintaining his "free" in-house solution with half the feature set.

All of these problems are about not taking the cost of expensive engineering time and energy into account in a rational way.

Clearly generative AI can currently only be used when verification is easy. A good example is software. Not sure why you think that I claimed otherwise.

You are correct to note 'real' costs of the leading labs are not public. It is surely true that the labs are operating at below cost (we are definitely not paying for the full R&D), but it seems unlikely that this fully explains the reduction in inference costs over the last years. We also know from open models like deepseek that the cost per inference token at a fixed performance level is going down very quickly matching the curve of leading labs inference cost decreases. You can even test it yourself on your own pc if you want.

I would add that inference cost decreases is what we should expect, it stands to reason there will be algorithmic improvements in inference, and compute cost is still going down just because of (a somewhat sloped) Moore's law.

Maybe you could also be a bit friendlier and forthcoming in your responses.

There was a paper a while back on AI usage at work among engineers and it was very strongly correlated to age. This is not surprising, technology adoption is always very dependent on age. (None of this tells you if the technology is a net good)

There are different criteria in use for that. But sycophantic behavior is not the goal. It's something model builders actively try to prevent.

Regarding the apple paper: https://andrewmayne.com/2024/10/18/can-you-dramatically-impr...

Craving "authenticity" is somewhere very high up the hierarchy of needs, i.e. a luxury. That a lot of people do not care much for it is not a sign of moral failing but of having bigger fish to fry.

> So, a living parrot might be more intelligent than these stochastic parrots.

Or the other way around.

The real costs are dropping because of real efficiency gains and compute cost reductions.

The inference costs are dropping at a similar rate.

This is incorrect, there are real effiency gains.

Slightly old by the standards of this field, but a good overview: https://arxiv.org/abs/2403.05812

The costs are a problem. We don't have hard evidence that this will be solved, but with algorithmic efficiency and raw compute costs both changing rapidly, the cost per token has gone down by about a factor of 10 per year for the last 3 years, i.e. 1000x over 3 years.

It's clearly true that the LLMS are 'stochastic parrots', but for all we know that might be the key to intelligence. It is in itself not a deep observation any more than calling your fellow humans 'microbial meatbags'.

Saying that LLMs are stochastic machines does not establish an upper bound for success.

> And if there are facts or code involved, both require manual confirmation.

The hidden assumption here seems to be that the model needs to be perfect before it has utility.

Having a ground truth doesn't mean it does not make huge and glaring mistakes.

I think the first part of that statement requires more evidence or argumentation, especially since models have shown the ability to practice deception. (you are right that they don't _always_ know what they know)

Death, taxes, and insoluble and exhausting debates around machine thinking and its value.

Is it good for one person (the writer) to ask a loaded question just to save some time on making their reasoning explicit, ony for lots of other people (the readers) to have to do extra work to understand what the argument is?

This is patently false. They are trained to generate correct responses.

People have stopped using LLMs? I wasn't aware of that. Can you share a source for that?