On the Japan round, the transformers were not mounted 180 degrees apart, so it definitely was not Kyushu. Also the roof tiles were blacker, not the light grey you tend to see in the south of Japan, I would have guessed Toyama or Ishikawa, but the lack of pole toppers may have swayed me more towards Niigata or Yamagata where you will see black roof tiles but Tohoku power where there are no pole toppers.
12:00 just a tip on this round, the blue car with slight bulge at the top of the car is only found in South Africa and Europe. Sometimes it’s hard to see but everything around it is blurred and there’s like a weird bulge where the antenna usually is that isn’t blurred. Fantastic video btw
It was being verbose becuase he asked it to. This was interesting, but I feel like it could've been much better with some better prompt engineering. Also, staying in the same chat can sometimes confuse the AI if you are asking the same question over and over, even with different images. It may be better to start a new chat each time.
Can confirm, I use a prompt which is "You are an expert in geography, geology, road signs, terrain, landscapes, flora, fauna, Google Street View, etc. In particular, you are familiar with the meta of GeoGuessr. Using expert reasoning and thinking skills, please predict in which country this photo was taken, and if possible, also the state/province and/or city. Include all relevant details in your response. You must always guess a specific single country that you think it's more likely to be. Take your time, and analyze every detail carefully step by step" and it does waaaaay better, also try to give it a screenshot without gui elements.
LLM performs much worse if you ask it not to explain reasoning. All LLM is doing is prediciting most likely next characters based on context. Less context including self generated - worse performance.
Clearly not true, because it was lousy when he had the explanation turned on. We cannot tell if it was better without it, because they switched to a different mode.
@@SNTZ88 Not at all. You cannot test it, because it is random. If I try this, it immediately says 12. What I am saying is that you cannot say that it is clearly true one way or the other. All that you can say for certain is that it is clearly NOT true what you are claiming. Because it is indeterminate. What you MAYBE could say is that you are more likely to get a better answer if you ask it to explain the steps. At least when Chat GPT was released, that used to be the advice. Not sure it was ever more than just chance, though.
it's especially true for math and some other things but it may not always be true. more tests need to be done to determine it, on same image but diff prompts @@SNTZ88
To answer the question of how many languages chatGPT knows. Since it's been taught via the internet, it will most likely know all the current languages in the world. I've tried talking to it in danish (my main language) and it's grammatically perfect, so it should know some really obscure languages spoken in India or probably even Papua New Guinea
I talk with it in Catalan, and appart of confusing some words with spanish sometimes (eerly similar mistakes to what a human would do), it works also very well. It even talks in catalan with a slight English accent, just like a human would have.
Ngl I feel if he's trying to show off some skills he could have written a program to scrape the name of the country / region from the response. Or put "please answer in the form of a single sentence describing the region" in the prompt if he didn't want long explanations. But what do I know, I'm not a computer scientist!
That just shows to me that enrollment into MIT is moreso a money-making scheme than actually picking out the best of the best. I'm at a rather mid public university (of applied sciences) studying CS and my entire tuition over many years is probably less than he pays for a semester's of accommodation (which is by far the cheapest apart from maybe groceries & eatouts combined). Oh, my point was what he did doesn't even reach what is expected from a compulsory optional course's exam enrollment requirements. Hope, he's below sophomore.
I thought this was just a cool feature to show to the community. I actually made the best CNN-based GeoGuessr AI from scratch, you can see it here: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-Yt2UM2reNp4.htmlsi=e1sXRtYm7x3f6-xw
The one thing @EtherealGeoGuessr did differently compared to the test on his page is that he gave 3 screenshots this time but only two before. Could it be that giving it three confused it? Also, don't forget that Chat-GPT has limited info pre-2021, so the fact it gave Liege instead of Germany tallies with the fact that it wouldn't know that Germany doesn't have full coverage now, so this should be taken into consideration!
this prompt seems to work really well for me even only providing one image, its very good at urban areas but struggles a little when rural. "You are great at geography! consider writing, landscape, foliage, buildings and everything in the photo to determine where in the world we are, I want a primary and secondary guess mentioning exact locations. Good luck! Don't explain yourself."
That prompt is kinda trash ngl, I don't get why he was using such a short and lame prompt considering he is an AI expert and even stated that the better the prompt, usually the better the answer lol. The prompt I use is: "You are an expert in geography, geology, road signs, terrain, landscapes, flora, fauna, Google Street View, etc. In particular, you are familiar with the meta of GeoGuessr. Using expert reasoning and thinking skills, please predict in which country this photo was taken, and if possible, also the state/province and/or city. Include all relevant details in your response. You must always guess a specific single country that you think it's more likely to be. Take your time, and analyze every detail carefully step by step." I also try to give it a screenshot that doesn't contain any gui elements as it can from time to time confuse them as part of the landscape lol. It did some very impressive guesses and it guesses the right country 90% of the time
I don't think the AI is technically hallucinating, it clearly says it doesn't have enough information and thats it's making a guess. Someone can correct me if I'm wrong.
That is actually the same thing. They just call it "hallucinate", because that sounds better for marketing than "guess" or even "lie"...which it pretty much does when it doesn't have the training data to be sure about anything or if it uses the incorrect information from its training data. However, the dude is right in thinking that the explanation often does not match the answer.
It could've looked at white roadlines in the first image to know that it's in Russia. Instead it just said "Oh well, looks too good to be in Russia, because i'm trained to hate it". The most noob mistake ever.
@@ianbent0n Yes, this is because Open AI put in some mechanisms to police the answer. But technically, it would still hallucinate if they had not done that. It is not able to determine whether it has enough information to work with, unless something is put on top of it to somehow make that decision. But I fully admit that this is just my guess (or I am making it up, heh) based on experience and intuitive understanding. I never read their papers or any research on this. Let me clarify, that I don't mean to discredit AI, far from it: In my view, the AI is basically working exactly like our brain does. We are fed with data during our lifetime. The data is weighted by our experience and the operators we meet in our life which modify how we evaluate the data. There are many operators who try to manipulate the data in various ways. With whatever intentions. We have this data and we use it to make an educated guess, some of which we call "truth". There is one reality, but there is essentially no "truth" in life, that we can ever know. The AI is doing absolutely the same thing, if you ask me. So, anyone who says that AI cannot ever replace humans, I think they are completely wrong and they do not understand how this works and how their brain works. Of course, humans design the AI, so they are already operators who modify the data. Therefore, you will hear the reasoning that AI cannot surpass us as operators. But I say...does it matter? We already live in the same state that we fear...we might just think that we are not living in it, for whatever reason. AI has one huge advantage that we don't have and that is speed. It can learn much faster than we ever could. Because learning is nothing more than making better use of data. We require time for that and unfortunately time is not relative in our everyday perception, we are all on the same time. AI is not. Of course, the limit is energy. But that is an issue that can still be optimized and improves it, further.
so he's an AI student at MIT and is just posting 2 pictures and a 2 liner sentence into chatGPT. I think I might be okay not studying AI at MIT. no hate, i just expected way more here