The scaling laws are also power laws, meaning that most of the big gains happen early in the curve, and improvements become more expensive the further you go along.
LeCun has already proved himself and made his mark and is now in a lucky position where he can focus on very long term goals that won't pay off for a long time (or ever). I feel like that is the best path someone like him could take.
Basically what we have done the last few years is notice neural scaling laws and drive them to their logical conclusion. Those laws are power laws, which are not quite as bad as logarithmic laws, but you would still expect most of the big gains early on and then see diminishing returns.
Barring a kind of grey swan event of groundbreaking algorithmic innovation, I don't see how we get out of this. I suppose it could be that some of those diminishing returns are still big enough to bridge the gap to create an AI that can meaningfully recursively improve itself, but I personally don't see it.
At the moment, I would say everything is progressing exactly as expected and will continue to do so until it doesn't. If or when that happens is not predictable.
do you consider gpt itself and reasoning models to be two grey swan events? I expect another one of similar magnitude within two years for sure. I think we are searching more efficiently for such ideas than before w/ more compute & funding.
I would say GPT itself is less an event and more the culmination of decades of research and development in algorithms, hardware, and software. Of course, to some degree, this is true for any novel development. In this case, the convergence of development in GPUs, software to utilize them well while being able to work in very high levels of abstractions, and algorithms that can scale is something I'm not sure we will see again so quickly. All this preexisting research is kind of a resource that will be completely exploited at some point. And then the only thing that can drive you forward are truly novel ideas. Reasoning models were a fairly obvious next step too as the concepts of System 1 and 2 have been around for a while.
You are completely right that the compute and funding right now are unprecedented. I don't feel confident making any predictions.
I suppose in the beginning, it was about finding ways to measure how effective different altruistic approaches actually are and focusing your efforts on the most effective ones. Effective then essentially means how much impact you are achieving per dollar spent. One of the more convincing ways of doing this is looking at different charitable foundations and determining how much of each dollar you donate to them actually ends up being used to fix some problem and how much ends up being absorbed by the charitable foundation itself (salaries etc.) with nothing to show for it.
They might have lost the plot somewhere along the line, but the effective altruism movement had some good ideas.
> One of the more convincing ways of doing this is looking at different charitable foundations and determining how much of each dollar you donate to them actually ends up being used to fix some problem and how much ends up being absorbed by the charitable foundation itself (salaries etc.) with nothing to show for it.
Color me unconvinced. This will work for some situations. At this point, it's well known enough that it's a target that has ceased to be a good measure (Goodhart's Law).
The usual way to look at this is to look at the percentage of donations spent on administrative costs. This makes two large assumptions: (1) administrative costs have zero benefit, and (2) non-administrative costs have 100% benefit. Both are wildly wrong.
A simple counterexample: you're going to solve hunger. So you take donations, skim 0.0000001% off the top for your time because "I'm maximizing benefit, baby!", and use the rest to purchase bananas. You dump those bananas in a pile in the middle of a homeless encampment.
There are so many problems with this, but I'll stick with the simplest: in 2 weeks, you have a pile of rotten bananas and everyone is starving again. It would have been better to store some of the bananas and give them out over time, which requires space and maybe even cooling to hold inventory, which cost money, and that's money that is not directly fixing the problem.
There are so many examples of feel-good world saving that end up destroying communities and cultures, fostering dependence, promoting corruption, propping up the institutions that causing the problem, etc.
Another analogy: you make a billion dollars and put it in a trust for your grandchild to inherit the full sum when they turn 16. Your efficiency measure is at 100%! What could possibly go wrong? Could someone improve the outcome by, you know, administering the trust for you?
Smart administration can (but does not have to) increase effectiveness. Using this magical "how much of each dollar... ends up being used to fix some problem" metric is going to encourage ineffective charities and deceptive accounting.
That's fair enough, there are problems with this way of thinking. I suppose you could say the take-away should be "Don't donate to charities where close to your whole donation will be absorbed as administrative costs". There definitely are black sheep that act this way and they probably served as the original motivation for EA. It's a logical next step to come up with a way to systematically identify these black sheep. That is probably the point where this approach should have stopped.
> It seems that you're saying that a therapist will be "rubbish unless they use basic Cognitive Behavioural Therapy concepts" ? i.e. that this is the only valid approach to therapy?
I believe the parent poster is saying that CBT is the only form of therapy you can trust an LLM to pull off because it's straightforward to administer.
Computerised CBT is even already being delivered and by quite a bit less sophisticated systems than LLMs. Resourcing constraints have made it very popular in the UK.
With many drugs that are used both therapeutically and recreationally, it is the case that the average recreational dosage is much larger. Ketamine is an exception here, as therapeutic doses are actually quite high. The common mistake here is to equate intravenous dosage with intranasal dosage, when the bioavailability differs significantly between these routes of administration. And that's not even considering that most reported recreational dosages are wrong due to cutting agents.
There certainly is recreational abuse with very large dosages, but I don't think it's fair to claim that the majority of users fall into this category.
I think we have seen a general trend towards centralized platforms on the internet. Where you had many individual niche sites before, now you have a few all-encompassing platforms. There are some exceptions, but I generally find that many of those platforms want to maximize your time on the platform itself. As a consequence, they do what they can to keep you from leaving the platform via a link to some other website.
Yes, there is additional context that is not explicitly stated in the question. It is clear that you are looking for a job to earn money and live your life and everyone already knows this, so there is no need to talk about. The question is: Why did you apply (here out of all the places you could have applied to)?
It's not like the US doesn't have a problem with affordable housing, so I don't see how this plays any role in the divide.
Germany has plenty of applied research organizations, from universities (e.g. RWTH) to things like Fraunhofer. The funding schemes behind these organizations are horrible and I would argue that in many ways, they are machines to burn up potential. Even with all this, Germany has been doing okay on the publicly funded AI research front, but that is irrelevant. The US isn't leading because of publicly funded AI effort, but because of privately funded AI effort.
This seems like the classic shifting of goalposts to determine when AI has actually become intelligent. Is the ability to communicate not a form of intelligence? We don't have to pretend like these models are super intelligent, but to deny them any intelligence seems too far for me.
My intent was not to claim communication isn’t a sign of intelligence, but that the appearance of communication and our tendency to anthropomorphize behaviors that are similar to ours can result in misunderstandings as to the current capabilities of LLMs.
glenstein made a good point that I was commingling concepts of intelligence and consciousness. I think his commentary is really insightful here: https://news.ycombinator.com/item?id=42912765
AI certainly won't be intelligent while it has episodic responses to queries with no ability to learn from or even remember the conversation without it being fed back through as context. This is the current case for LLM models. Token prediction != Intelligence no matter how intelligent it may seem. I would say adaptability is a fundamental requirement of intelligence.
>AI certainly won't be intelligent while it has episodic responses to queries with no ability to learn from or even remember the conversation without it being fed back through as context.
Thank God no one at the AI labs is working to remove that limitation!
The guy in memento is clearly still an intelligent human despite having no memory. These arguments always strike me as coming from a "humans are just special okay!" place. Why are you so determined to find some way in which LLMs aren't intelligent? Why gatekeep so much?
