This is the technology worth trillions of dollars huh
With enough duct tape and chewed up bubble gum, surely this will lead to artificial general intelligence and the singularity! Any day now.
I don’t think this gets nearly enough visibility: https://www.academ-ai.info/
Papers in peer-reviewed journals with (extremely strong) evidence of AI shenanigans.
I would estimate that Google’s AI is helpful and correct about 7% of the time, for actual questions I’d like the answer to.
✅ Colorado
✅ Connedicut
✅ Delaware
❌ District of Columbia (on a technicality)
✅ Florida
But not
❌ I’aho
❌ Iniana
❌ Marylan
❌ Nevaa
❌ North Akota
❌ Rhoe Islan
❌ South Akota
Gosh tier comment.
You just described most of my post history.
Everyone knows it’s properly spelled “I, the ho” not Idaho. That’s why it didn’t make the list.
They took money away from cancer research programs to fund this.
only cancer patients benefit from cancer research, CEOs benefit from AI
Tbf cancer patients benefit from AI too tho a completely different type that’s not really related to LLM chatbot AI girlfriend technology used in these.
After we pump another hundred trillion dollars and half the electricity generated globally into AI you’re going to feel pretty foolish for this comment.
Connecdicud.
i rather manually search for info
Connedicut.
I wondered if this has been fixed. Not only has it not, the AI has added Nebraska.
You mean Connecdicud.
I would assume it uses a different random seed for every query. Probably fixed sometimes, not fixed other times.
What about Our Kansas? Cause according to Google Arkansas has one o in it. Refreshing the page changes the answer though.
Just checked, it sure does say that! AI spouting nonsense is nothing new, but it’s pretty ironic that a large language model can’t even parse what letters are in a word.
Well I mean it’s a statistics machine with a seed thrown in to get different results on different runs. So really, it models the structure of language, but not the meaning. Kinda useless.
Blows my mind people pay money for wrong answers.
Listen, we just have to boil the ocean five more times.
Then it will hallucinate slightly less.
Or more. There’s no way to be sure since it’s probabilistic.
If you want to get irate about energy usage, shut off your HVAC and open the windows.
sounds reasonable… i’ll just go tell large parts of australia where it’s a workplace health and safety issue to be out of AC for more than 15min during the day that they should do their bit for climate change and suck it up… only a few people will die
maybe people shouldn’t live there then?
of course you’re right! we should just shut down some of the largest mines in the world
i foresee no consequences from this
(related note: south australia where one of the largest underground mines in the world is, largely gets its power from renewables)
people should probably move from canada and most of the north of the USA too: far too cold up there during winter
Worthless comment.
Even more worthless than mine, somehow.
Well, for anyone who knows a bit about how LLMs work, it’s pretty obvious why LLMs struggle with identifying the letters in the words
Well go on…
They don’t look at it letter by letter but in tokens, which are automatically generated separately based on occurrence. So while ‘z’ could be it’s own token, ‘ne’ or even ‘the’ could be treated as a single token vector. of course, ‘e’ would still be a separate token when it occurs in isolation. You could even have ‘le’ and ‘let’ as separate tokens, afaik. And each token is just a vector of numbers, like 300 or 1000 numbers that represent that token in a vector space. So ‘de’ and ‘e’ could be completely different and dissimilar vectors.
so ‘delaware’ could look to an llm more like de-la-w-are or similar.
of course you could train it to figure out letter counts based on those tokens with a lot of training data, though that could lower performance on other tasks and counting letters just isn’t that important, i guess, compared to other stuff
Of course, when the question asks “contains the letter _” you might think an intelligent algorithm would get off its tokens and do a little letter by letter analysis. Related: ChatGPT is really bad at chess, but there are plenty of algorithms that are super-human good at it.
Good read. Thank you
Con-ned-di-cut
Wouldn’t that only explain errors by omission? If you ask for a letter, let’s say D, it would omit words containing that same letter when in a token in conjunction with more letters, like Da, De, etc, but how would it return something where the letter D isn’t even in the word?
Well each token has a vector. So ‘co’ might be [0.8,0.3,0.7] just instead of 3 numbers it’s like 100-1000 long. And each token has a different such vector. Initially, those are just randomly generated. But the training algorithm is allowed to slowly modify them during training, pulling them this way and that, whichever way yields better results during training. So while for us, ‘th’ and ‘the’ are obviously related, for a model no such relation is given. It just sees random vectors and the training reorganizes them tho slowly have some structure. So who’s to say if for the model ‘d’, ‘da’ and ‘co’ are in the same general area (similar vectors) whereas ‘de’ could be in the opposite direction. Here’s an example of what this actually looks like. Tokens can be quite long, depending how common they are, here it’s ones related to disease-y terms ending up close together, as similar things tend to cluster at this step. You might have an place where it’s just common town name suffixes clustered close to each other.
and all of this is just what gets input into the llm, essentially a preprocessing step. So imagine someone gave you a picture like the above, but instead of each dot having some label, it just had a unique color. And then they give you lists of different colored dots and ask you what color the next dot should be. You need to figure out the rules yourself, come up with more and more intricate rules that are correct the most. That’s kinda what an LLM does. To it, ‘da’ and ‘de’ could be identical dots in the same location or completely differents
plus of course that’s before the llm not actually knowing what a letter or a word or counting is. But it does know that 5.6.1.5.4.3 is most likely followed by 7.7.2.9.7(simplilied representation), which when translating back, that maps to ‘there are 3 r’s in strawberry’. it’s actually quite amazing that they can get it halfway right given how they work, just based on ‘learning’ how text structure works.
but so in this example, us state-y tokens are probably close together, ‘d’ is somewhere else, the relation between ‘d’ and different state-y tokens is not at all clear, plus other tokens making up the full state names could be who knows where. And tien there’s whatever the model does on top of that with the data.
for a human it’s easy, just split by letters and count. For an llm it’s trying to correlate lots of different and somewhat unrelated things to their ‘d-ness’, so to speak
Thank you very much for taking your time to explain this. if you don’t mind, do you recommend some reference for further reading on how llms work internally?
“This is the technology worth trillions of dollars”
You can make anything fly high in the sky with enough helium, just not for long.
(Welcome to the present day Tech Stock Market)
Bubbles and crashes aren’t a bug in the financial markets, they’re a feature. There are whole legions of investors and analysts who depend on them. Also, they have been a feature of financial markets since anything resembling a financial market was invented.
Just another trillion, bro.
Just another 1.21 jigawatts of electricity, bro. If we get this new coal plant up and running, it’ll be enough.
Behold the most expensive money burner!
We’re turfing out students by the tens on academic misconduct. They are handing in papers with references that clearly state “generated by Chat GPT”. Lazy idiots.
This is why invisible watermarking of AI-generated content is likely to be so effective. Even primitive watermarks like file metadata. It’s not hard for anyone with technical knowledge to remove, but the thing with AI-generated content is that anyone who dishonestly uses it when they are not supposed to is probably also too lazy to go through the motions of removing the watermarking.
Couldn’t students just generate a paper with ChatGPT, open two windows wide by side and then type it out in a word document?
I think I’d at least use an OCR program to do the bulk of the typing for me…
but that’s work.
Students view doing that as basically the same amount of work as writing the paper yourself
Depends on the watermark method used. Some people talk about watermarking by subtly adjusting the words used. Like if there’s 5 synonyms and you pick the 1st synonym, next word you pick the 3rd synonym. To check the watermark you have to access to the model and probabilities to see if it matches that. The tricky part about this is that the model can change and so can the probabilities and other things I don’t fully understand.
if you are going to do all that, just do the research and learn something.
Aye that’s exactly the same thing that I said
Huh that actually does sound like a good use-case of LLMs. Making it easier to weed out cheaters.
Hey look the markov chain showed its biggest weakness (the markov chain)!
In the training data, it could be assumed by output that Connecticut usually follows Colorado in lists of two or more states containing Colorado. There is no other reason for this to occur as far as I know.
Markov Chain based LLMs (I think thats all of them?) are dice-roll systems constrained to probability maps.
Edit: just to add because I don’t want anyone crawling up my butt about the oversimplification. Yes. I know. That’s not how they work. But when simplified to words so simple a child could understand them, its pretty close.
I was wondering if you’d get similar results for states with the letter R, since there’s lots of prior art mentioning these states as either “D” or “R” during elections.
Oh l I was thinking it’s because people pronounce it Connedicut
Awe cute!
