Today I saw (but refuse to honor with a link) a YouTube video using ChatGPT to predict Trump's first 100 day in office "and it is worse than you imagined". No, actually that ChatGPT prediction is exactly as bad as people imagined. Because all that ChatGPT does is take all the fantasies from people who wrote about that subject on the internet and regurgitates them. You get both the horror visions from the left, and the crazy power fantasies from the right, mix them together and get an "AI prediction". Which is complete nonsense.
I was struck by that AI prediction approach, because it is diametrically opposed to my previous post. In that I had looked at various election promises and fears, and subjected them to a reality check. It is exactly that reality check that AI is unable to provide. Large language models repeat language they heard, they can't reason, and thus can't contradict. If I created an Tobold.AI to write my blog posts, it could only tell you what everybody is saying. It couldn't tell you why that might possibly be wrong.
I'm sure you know that's not how LLMs work. In fact, no-one knows how they work, not even the people who created them. There's constant research taking place trying to figure out how it is they do what they do. I recommend this article for a basic overview of the complexity involved.
ReplyDeletehttps://www.understandingai.org/p/large-language-models-explained-with
Whatever they're doing, though, they're certainly not just "repeating language they heard" and they most definitely can reason.
This paper, linked in the article Bhagpuss linked, gets a bit more directly to whether LLMs actually think or only appear to think:
ReplyDeletehttps://onlinelibrary.wiley.com/doi/full/10.1111/cogs.13309
If you skip down to the discussion, my reading of it is that there are roughly three classes of explanations for the performance of LLMs at "reasoning" tasks that are considered roughly equally viable by the authors of the paper. Two out the three hold that they only appear to think. My a priori bias is certainly towards this interpretaion.
In my mind the "if it looks and acts like a duck, it's a duck" interpretation only makes sense if you didn't build the duck yourself, and know full well that it's not really a duck. It would take much stronger evidence than I have ever been presented with to convince me that we have managed to create a duck, starting with completely different components than those ducks are made of.
I will say that regardless, it's amazing to be living in a time when "do these computer programs actually reason" is even a viable question that is up for real debate.
Well the difference may be that (like chatGPT) it may be able to apologize for its errors. Like now it seems to me that after losing the elections, democrats behaved very differently from your predictions (both recent and at the time of the 6th jan uprising), still you don't seem to have commented on how your predictions were off the mark and you don't have apologized to democrats for treating them "like republicans".....
ReplyDeleteThe absence of political violence a few days after the election doesn't guarantee an absence of political violence for the next 4 years. Political violence from the left did exist during the BLM riots and pro-Palestinian protests, will happen again, and Trump might very well still trigger some. It just doesn't follow the same playbook as political violence from the right.
ReplyDeleteI don't think you are going to see political "violence" as you call it from the left until Trump starts enacting policies that hurt people.
DeleteThe nature of Trumps win is different. Millions of white democrats didn't even vote. Trump won the popular vote. Trump won in what is effectively a landslide and will also have majorities in congress. I don't forsee any January 6th style event happening in the context of how he won.
I do believe, or at least I hope, that once his proposed 1 million person mass deportation plan starts in January that we do see protests and backlash.
The cynic in me says we likely won't even see protests in regards to mass deportations either and likely won't unless Trump fumbles the economy hard as apparently my fellow Americans are okay with just about anything as long as they think their rent and grocery prices will go down.
@Tobold: ahhh yes, the good old "not right now, but it'll be happening"..... why do you sound more and more like QAnon? You say you have a background in science (which I am honestly starting to question....) and still you cannot tell the difference between reality (i.e. what happens) and your projections (i.e. what you think happens)? The reality is that democrats never started anything like "The Steal" and their candidate conceded, something that Trump never did. But you can continue to live in your fairyland where "left is bad". Young people are progressive, old people are conservative, thanks for reminding us.
ReplyDelete@Helistar You are falling for a primitive form of partisanship, where because one side is bad, you can’t admit that the other side might be bad too. Yes, right wing political violence is bad, but left wing political violence is bad too, e.g. the antisemitic violence caused by the Israel-Hamas conflict. You can’t just say the other side is “more bad”, and that excuses everything. I never excused right-wing political violence, while you deny left-wing political violence. So, the only QAnon-like behavior is yours, just the other way round. Kudos to Biden and the Democrats for their dedication to a peaceful transfer of power, but that doesn’t erase past left-wing political violence, nor does it make future left-wing political violence impossible. America is deeply polarized, and 27% of Americans in a pre-election poll feared “civil war”. That wouldn’t be possible if only one side was able of violence. It is like the Israel-Hamas conflict: One side uses human shields, the other side shoots human shields. There are no good guys.
ReplyDelete@Tobold I hope you realize the IDF also uses "human shields" in Gaza as well... https://www.cnn.com/2024/10/24/middleeast/palestinians-human-shields-israel-military-gaza-intl/index.html
DeleteThere are indeed no good sides to the Gaza/Israel conflict. Both sides of that conflict engage in what are war crimes. Innocent Civilians in Gaza are being killed by both the IDF and Hamas. The main difference is Israel is killing far more then Hamas is capable of.
@Bhagpuss
ReplyDeletethank you, the article is enlightening indeed.
I will try to read the article but at least one thing strike me as near enough to the reasoning to be probable : a LLM trained on Reverso game move ( eg : White put in A15) and result was able to 1) build an internal représentation of the game board 2) become better at playing than 90% of people and 3) has built a représentation of its own performance level, that can be changed to make it better or worse.
ReplyDeleteSeveral people managed with specific prompts to get ChatGPT to spout nonsense, e.g. by asking how many rocks one should eat per day, or how to get cheese to stick better to pizza and getting a reply based on some satirical post somewhere on the internet. That shows that the AI can't understand a concept like "edible" and doesn't "reason".
ReplyDeleteIt is slightly more complex than that. ChatGPT or other LLM are very sensitive to the prompt used. Bad prompt can lead to horrible result, while good prompt can give far better than expected result. Chess is an example : until recently, everyone that tried ChatGPT find it was a very very bad player, with often illegal moves. But using the right prompt, and directly GPT without the 'chatGPT' filter, GPT is able to win over 99% of the population. And he has not been trained to be good at chess.
DeleteI am not saying GPT or LLM are smart, or should be trusted. They can make some very stupid, but very convincing argument. On the other hand, there is a good chance they are more than the 'stochastic parrot'.
The scientific consensus is not established here with the scientific community divided, some thinking he is a stochastic parrot, some saying it is superior, and some not able to conclude.
So I guess my take is : there is no clear answer yet of what are the current limit of LLM, and if those limits are inherent or just a matter of time. So, except if you are an expert in LLM (and I am clearly not), the right approach here is to avoid strong opinion, and be more nuanced.
I think it is possible to have a strong opinion that LLM AI should not be trusted, especially not in the domain of prediction. Financial advice on the internet is generally dubious, and “AI stock picks” are probably complete folly.
ReplyDelete