this post was submitted on 12 Aug 2025
18 points (90.9% liked)
Hacker News
2472 readers
368 users here now
Posts from the RSS Feed of HackerNews.
The feed sometimes contains ads and posts that have been removed by the mod team at HN.
founded 11 months ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
https://www.anthropic.com/research/tracing-thoughts-language-model
well that's provably false as the neural network can be traced to understand the exact process and calculations that it follows when doing math. See the above research under heading "Mental Math".
The reason LLMs struggle to count letters is because of tokenization. The text is first converted to tokens, numeric vectors which represents whole words or parts of words, before being given to the LLM, and the LLM has no concept of or access to the underlying letters that make up the words. The LLM outputs only tokens, which are converted back into text.
EDIT: you can downvote me, but facts are facts lol.
There is definitely more going on with LLMs than just direct parroting.
However, there is also an upper limit to what an LLM can logic/calculate. Since LLMs basically boil down to a series of linear operations, there is an upper bound on all of them as to how accurately they can calculate anything.
Most chat systems use python behind the scene for any actual math, but if you run a raw LLM you can see the errors grow faster as you look at higher orders of growth (addition, multiplication, powers, etc.).
Yes, exactly. It can do basic math and also doesn't really matter because it is really good at interfacing with tools/functions for calculation anyway
Also this is only true when LLMs are not using Chain of Thought.
So what you're saying is it can't count.
Got it.
Reading comprehension. The only thing it cannot count are specifically letters of tokenized words. If you separate the letters out into separate tokens ("B L U E B E R R Y", etc) it will get it correct 100% of the time.
Yeah, that's just not how neural networks work...
https://en.wikipedia.org/wiki/Regression_analysis
You tried.
https://en.wikipedia.org/wiki/Universal_approximation_theorem Actual math directly contradicts your beliefs. I know its trendy and you want to feel smarter than people who have spent literally decades researching NNs and staked billions of dollars developing it, but you're wrong.
Your claim is like claiming that boolean circuits cannot do math because all they do is "true" or "false".
And also, https://en.wikipedia.org/wiki/Recurrent_neural_network#Neural_Turing_machines recurrent neural networks are Turing complete when paired with memory, and therefore can be used to any calculations or computations that a conventional computer can do.
LOL
The funniest part of this is not the fact that an LLM just got 3 for 3 correct, and therefore has a 100% success rate, thus proving you wrong again, but the fact that your "favorite test" would be one that you incorrectly believe "no LLM will ever be able to" do because....
^ this you??? "My favorite test is to see if the screwdriver shoots laser beams" 🙃
~~I was using standard RGB hex codes, so I didn't really need to specify because its the assumed default. If it was something different, I would need to specify.~~ EDIT: oh I just realized you meant the LLM model, not the color model (RYB vs RGB). It was just from ChatGPT, thought the interface would be recognizable enough.
Huh? What do you mean? Go try it!
Yeah, so this is already a thing. 24-bit color (8 bits per color channel) already gives you 16,777,216 colors, which is pretty good, but if you want more precision, you can just use decimal (floating point) numbers for each channel, like sRGB(0.25, 0.5, 1.0) (https://en.wikipedia.org/wiki/SRGB) OR even better would be to use oklch (https://en.wikipedia.org/wiki/Oklab_color_space). This is a solved problem. Or you cold just define your range as 0 to 47204.
So... we've gone from "no LLM will ever be able to understand what complementary colors are" to "b-b-but what about arbitrary color models I make up??" And yeah, it will handle those too, you just have to tell it what it is when you prompt it.
Green is the correct answer in the RYB color model, which is traditionally used in art and most commonly taught in schools.
And... wait for it...
And an open-weight model (qwen3:32b)
So you're:
😂 multiple LLMs literally gave the exact answer that you claim they can't correctly give, on the very first try. Checkmate.
🤦 Oh... oh wow, I was giving you way more credit than what you actually meant. You do realize there is more than one color model? https://en.wikipedia.org/wiki/Complementary_colors#In_different_color_models You probably should read the explanation about complementary colors based on digital screens that they are providing to you (or just pay attention in elementary art class), because you might actually learn something new.
Red Yellow Blue and Cyan Yellow Magenta are both subtractive color models. RGB is an additive color model.
https://en.wikipedia.org/wiki/RYB_color_model
Try giving the LLM the hex color code and the color model you're using that code in, and it will give you the correct complementary color.
There is no one "correct" color circle. And your misguided beliefs about color theory do not have anything to do with LLMs.
By the way, they're called "additive colors", not "active colors". 🙃
😂 alright well, you've been corrected and proven wrong, with sources and screenshots. And clearly you're getting a teeny bit upset over it. Sorry! There's nothing wrong with learning something new, and its okay that you had made a mistake.
This was a fun rabbit hole to go down! I tend to agree with most of the takes here on Lemmy but the complete AI derision is pretty wild and unfounded in reality. I have plenty of concerns about the tech but to say it’s useless you’d have to really not even have tried it out to see for yourself. I appreciate your patience, dedication to the truth, willingness to explain, and experimental attitude here.
This was a super bizarre case of "the neutral networks that are a predictive model of all of the world's knowledge don't share my belief that only one specific color model is valid, therefore they suck" 😂