Isn't it also possible that Sam is being genuine about his concerns and his statements because he wants what is best for humanity?
"Fancy RNN" is a pretty ridiculous assertion.
And no, researchers didn't expect what GPT-3/4 has been shown to do to be around the corner at all.
GPT's aren't chatbots. That's just a neat natural consequence that's happened. They're machines that reason, understand and follow instructions in plain language. And their abilities go far beyond what any expected language modelling to provide.
They are not machines that reason, they are approximators. It’s all just token matching based on data we’ve fed it. Further it didn’t happen over night but through successive improvements and at no point was the next improvement considered some infeasible thing.
Seems to me maybe you’ve bought into the hype here, and are confusing that with reality.
And meaningless distinction of the year award goes to..
"It's not real [insert property]" is not an intelligent argument. By all means, divine the way to distinguish results of the two. After all, what kind of important distinction can't be distinguished by results?
>Further it didn’t happen over night but through successive improvements
The only difference between GPT-3 and GPT-2 was scale. They didn't even change the tokenizer until 4. There were no "successful improvements" to smoothen the massive gap in capabilities between the two. So to say that was expected just shows more lack of knowledge here.
"Other than that, how was the play, Mrs. Lincoln?"
GPT-2 is two orders of magnitude smaller. In terms of forebrain neuron count this is the difference between a human and an elephant shrew or budgerigar. It was absolutely expected by reasonable people that 2 OOMs of scaling will provide a qualitative jump.
When GPT-3 was released, it was by far the largest artificial neural network ever trained. And I mean by far. Now there wasn't any big jump in hardware capabilities to spur this sort of gulf. It wasn't a case of "Oh now we can train a very large model"
So Want to know why there was such a gap ? It's because most researchers assumed the models would overfit the data or display diminishing returns long before 175b.
Brain neurons are not comparable to ann parameters.
They approximate a function that performs reasoning...
> Conclusion: computer power is unlikely to be the issue anymore in terms of AGI being possible. The main question is whether we can find the right algorithms.
> One of the big things influencing me this year has been learning about how much we understand about how the brain works, in particular, how much we know that should be of interest to AGI designers. I won’t get into it all here, but suffice to say that just a brief outline of all this information would be a 20 page journal paper (there is currently a suggestion that I write such a paper next year with some Gatsby Unit neuroscientists, but for the time being I’ve got too many other things to attend to). At a high level what we are seeing in the brain is a fairly sensible looking AGI design. You’ve got hierarchical temporal abstraction formed for perception and action combined with more precise timing motor control, with an underlying system for reinforcement learning. The reinforcement learning system is essentially a type of temporal difference learning though unfortunately at the moment there is evidence in favour of actor-critic, Q-learning and also Sarsa type mechanisms — this picture should clear up in the next year or so. The system contains a long list of features that you might expect to see in a sophisticated reinforcement learner such as pseudo rewards for informative queues, inverse reward computations, uncertainty and environmental change modelling, dual model based and model free modes of operation, things to monitor context, it even seems to have mechanisms that reward the development of conceptual knowledge. When I ask leading experts in the field whether we will understand reinforcement learning in the human brain within ten years, the answer I get back is “yes, in fact we already have a pretty good idea how it works and our knowledge is developing rapidly.”
> I suspect that for the next 5 years, and probably longer, neuroscientists working on understanding cortex aren’t going to be of much use to AGI efforts. My guess is that sometime in the next 10 years developments in deep belief networks, temporal graphical models, liquid computation models, slow feature analysis etc. will produce sufficiently powerful hierarchical temporal generative models to essentially fill the role of cortex within an AGI.
> Right, so my prediction for the last 10 years has been for roughly human level AGI in the year 2025 (though I also predict that sceptics will deny that it’s happened when it does!) This year I’ve tried to come up with something a bit more precise. In doing so what I’ve found is that while my mode is about 2025, my expected value is actually a bit higher at 2028. This is not because I’ve become more pessimistic during the year, rather it’s because this time I’ve tried to quantify my beliefs more systematically and found that the probability I assign between 2030 and 2040 drags the expectation up. Perhaps more useful is my 90% credibility region, which from my current belief distribution comes out at 2018 to 2036.
And here's Rich Sutton's famous Bitter Lesson, a month after GPT-2 [2]:
> We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
Surprise indicates the mismatch of your mental model and reality, not the inherent weirdness of the latter. Both the surprised/alarmed people and people still in denial about the power of LLMs have to revisit their assumptions and ask if they were founded on any credible understanding to begin with.
1. http://www.vetta.org/2009/12/tick-tock-tick-tock-bing/
2. http://www.incompleteideas.net/IncIdeas/BitterLesson.html
When GPT-3 was released, it was by far by the largest artificial neural network ever trained and not because of any big jump in hardware technology. That's not the usual state of affairs for technology everyone or even most expect to pan out the way it did.
I believe he disagreed with the direction the board was taking it and tried to take over, which when failed he stopped funding it, and they went to investors
It's nice to think that you can have so much money you stop caring about it but it doesn't really make sense. You don't get to personally keep billions of dollars without caring about money.
> It's nice to think that you can have so much money you stop caring about it but it doesn't really make sense. You don't get to personally keep billions of dollars without caring about money.
It is definitely not every billionaire that thinks like this but there are some that seem to not care too much about material possessions like Musk or Sam. Their endeavors seem to be the things that they care about the most and they don't seem to be doing it to attain the most amount of money possible.
I have been running ML on my personal data and modeling worlds for the last few years. There’s no reason to keep OpenAI around given open source. My own data does not need a river water cooled data center.
OpenAIs ONLY moat is a government one so governments can keep up the free market ruse but have OpenAI in its back pocket for military and intelligence applications
I don't know that I agree with this strategy, but I can respect that it's a consistent viewpoint.
So what I hear from him is "Yes, this is dangerous stuff, and we're trying our hardest to make it safe, but we can't control everyone, so we need regulation ASAP because market forces are going to push everyone to race and release unsafe AIs".
> I can respect that it's a consistent viewpoint.
I don't think I agree that there's a great deal of consistency between that stated perspective and what OpenAI is actually doing. It's also, conveniently, a point of view that allows him to keep doing what he wants to do, and making bank by doing it.
Understand, I'm not saying he's a bad actor. I'm saying that it's hard to rule that out. From his WorldCoin stuff to this, there is plenty of reason to be suspicious of his motives.
aw heck, let's go so far as to say anyone willing to testify if front of congress primarily does so because they want to protect the best interests of others.
Would be a fucking swell world to live in.
So.. you're saying that every person that testifies before congress has bad intentions just because they testify before congress?
Say he did want what was best for humanity, how should he have acted differently in front of congress?
To me, from every interview I've heard from him, he doesn't seem like a Zuckerberg.
Don't assume they're they're to help. Don't assume they're there to hurt.
They are there to cover their own asses and nothing more.
Why is the assumption that they are only there to cover their own asses ok to make but other assumptions are not?
We don't know what he is actually thinking. It is entirely plausible that he is not a sociopath and actually wants to minimize negative impacts on society through regulation because he thinks that is the best path forward.