I think some of these 13B models may only be pretrained and may not have benefited from RLHF. So it’s not clear how much of the difference is due to parameter count.
The pixels may be 7.5 microns but you’re forgetting that they are viewed through a lens. The point stands: 4K pixels for the full field of view, which is a lower density than 4K for a small screen.
I’m surprised that these instructions are making such heavy use of negation, explicit (“not”) and implicit (“avoid”). It’s well known [0] that LLMs, like humans, struggle with negation. Some instructions, of course, cannot be easily framed without negation, but in other cases, there are obvious positive alternatives.
Clarification question: When you press C-c C-c on the B block, what is sent to ChatGPT, only B or the whole previous conversation? If the latter, this would break Org Babel semantics. Not necessarily a dealbreaker, but something to be aware of.
I know. But the idea that they'd pop the Champagne now suggests that the closure of the remaining plants was a reason for Russia to celebrate. However, if the phasing out of nuclear in Germany was ever a reason for them to celebrate, they would have celebrated long ago. Also, Russia is a major source of Uranium, so it's not at all clear that this is actually good for them. Further, even 20 years ago, nuclear covered only a tiny part of Germany's primary energy needs (at no time more than 12%) and could have easily been compensated for with coal. So again, it's not clear at all that ending nuclear has pushed the geopolitical needle.
No, Germany does not go the furthest. Other countries have quit nuclear long ago, e.g. Italy closed all plants in 1990 following a 1987 referendum, i.e. 33 years before Germany. Similarly Lithuania. Denmark already decided in the 1980ies not to build new plants.
Plus, given that sampling is probabilistic, one would have to prompt the model many times and evaluate all the responses. When I see examples as in the post, I always assume that the result was completely cherry-picked from a larger samples that contained all kinds of stuff.
I only tested GPT-3 and it started predicting illegal moves after 8-10 turns or so. When I played unusual moves, it went off rails much sooner and often immediately. If you know of a formal evaluation of GPT-4 chess skills, please post the link. Would love to see it. If it really makes illegal moves only rarely, that would be indeed impressive, but I believe it when I see it :)
I learned a lot from DPReview about photography, but I also learned a surprising amount from them about what excellent writing looks like. Sometimes I read long reviews about cameras that I wasn't even interested in just because the writing was so good.
Yeah, 63ms/token (number from the post) is totally usable. Didn't the CPU version that was recently posted take 8min/token? But I think that was a larger model.
Typical chess programs are specifically designed to play chess. A lot of human knowledge about chess is hard-coded in these programs. These programs are therefore not AI. They play chess and only chess. Bing AI on the other hand was never designed to play chess. It was merely designed to predict the next word in a text, a task that is quite different from chess. If such a program can play chess, that's much closer to what one might call AI, because it learned the rules of chess and chess strategy and tactics on its own. However, let's not get carried away, GPT models do play decent opening moves, but then quickly start predicting illegal moves. What they're doing is best described of having memorized common opening moves and they don't really have any solid understanding of the game.
We don't need faster trains in Germany. We need trains that are more or less on time. Punctuality is a complete disaster and probably the main reason why people hate taking the train in Germany. If you have a connection that requires changing to another train, the chance that you miss your connection is too damn high.
I don't think that people use cars because it's cheaper. Owning and using a car in Germany is pretty expensive and public transport probably was already cheaper for the vast majority of use cases.
"Potentially" is doing a lot of work here. I'd say that the trained model is worth approximately as much as the electricity that was required to train it.
The training data contains tons of false information and the training objective is simply to reproduce that information. It's not at all surprising that these models fail to distinguish truth from falsehood, and no incremental change will change that. The problem is paradigmatic. And calling people cynics for pointing out the obvious and serious shortcomings of these models is poor form IMO.