I don’t think you have read the relevant papers or are familiar with LRM (Large Reasoning Models). Which is basically all model AIs (GPT5, Claude, Gemini, DeepSeek). It’s new in the last ~18-24 months
In a nutshell, they include logical thinking and correct chains of logical thought to the LLM training data, along with tasks like recognizing dogs and predicting next words.
So yes, they are literally trained to reason the exact same way they are trained to write stories and summarize books.
You can say “it doesn’t really reason” but it has exactly the same value as the assertion “it doesn’t really write stories or summarize books”
… maybe not, but there will be a story or a summary (or a logical chain of thought) in front of you if you ask for one.
I will 100% admit to not reading papers and keeping up to date. I went ahead and spent about 30m looking up various explanations and summaries of LRMs. Ok, so you take an LLM and tell it to break the problem down first. It’s still not reasoning. It’s running a simulation of a natural language conversation, and giving you the center of mass of the statistical distribution for the intermediate steps. Does this kinda sorta replicate the sounds a human makes? Absolutely. But it’s irresponsible and unethical to make any claims that this is a human like entity you can chat with, or that it is doing any reasoning.
Birds string together words they hear when they can repeat them, and end up with the short phrases they seem to make. It’s extremely rare for them to actually understand meaning, most often it’s simply association which is why you often get nonsensical responses that still sort of make sense, or sounds out of them that sound like words but just…aren’t. The simalcrum of language without containing any. Often words can be linked by that, and our own brain wants to find words and sometimes we decipher ones, similar to seeing shapes in noise – we just tend to realize that it’s actually just recognition, not real.
What’s happening here is the equivalent of recording a bird and playing back it’s recording to itself to get a new response, as a chain. It’s predictive text feeding itself, in a simplistic but not inaccurate manner given how language models actually work at a technical level, tokenizing the input to train and create matrices of language vectors that contain word fragments, and often loop back on themselves or into yet more matrices of options. This is the “beam size” option some models have when run, selecting how many search routines should be created simultaneously for things that map to probability values that make sense.
Our own reasoning is far more complicated. Sometimes we think it’s just words, but our brain will seamlessly weave inner monologue into concepts and imagery or ideas without text and back again, sometimes into sounds or other things. We stitch together everything so seamlessly because it all actually has meaning for us.
LLMs having “reasoning” at all is operating by the Sapir-whorf hypothesis, which would imply there is no reasoning without language. And even animals can fucking reason without language. We absolutely did too. Sapir-whorf was an infantile thought experiment turned theory of language that’s been patently proven wrong even when it makes for great sci-fi (see Arrival).
This isn’t the difference between hearing a song live and played aloud, or midi/samples vs instruments. This is that part of our consciousness operates in some absolute wild ways that we can still only classify at a high level because the complexities are so far beyond what we can describe with models that, by comparison, are simplistic as hell.
Put another way, without transcriptions of “that’s right, the square hole”, if you showed two photos to a model and asked “where does this piece go” it’s just going to “see” the shape in both, recognize the image->word mapping and come up with a response fitting that, without ever being able to “realize” it can go into the square hole without being prompted, because it can’t invent.
we are talking about chat bots talking to people in this post, and how you can steer the simulated conversation towards whatever you want
it did not debug anything, a human debugged something and wrote about it. Then that human input and a ton of others were mapped into a huge probability map, and some computer simulated what people talking about this would most likely say. Is it useful? Sure, maybe. Why didn’t you debug it yourself?
Fair, we need to get terms straight; this is new and unstable territory.
Let’s say, LLMs specifically.
it did not debug anything, a human debugged something and wrote about it. Then that human input and a ton of others were mapped into a huge probability map, and some computer simulated what people talking about this would most likely say
Can you explain how that is different from what a human does? I read a lot about debugging, went to classes, worked examples…
Why didn’t you debug it yourself?
In my case this is enterprise software, many products and millions of lines of code. My test and bug-fixing teams are begging for automation. Bug fixing at scale
Ok, that’s easy. If I make an LRM model of your dead grandma, is that your grandma? Why not? What’s different?
Your bug fixing teams are begging for automation. That tells me you have an unsustainable setup. You are providing a bug fix suggestion tool, I don’t see how that fixes your problem. Seems like you need better coding practices and possibly more people.
Actually, they have been doing that for about a year now
https://arxiv.org/abs/2501.09686
No, just because they say they want it to reason, does not mean it does
I don’t think you have read the relevant papers or are familiar with LRM (Large Reasoning Models). Which is basically all model AIs (GPT5, Claude, Gemini, DeepSeek). It’s new in the last ~18-24 months
In a nutshell, they include logical thinking and correct chains of logical thought to the LLM training data, along with tasks like recognizing dogs and predicting next words.
So yes, they are literally trained to reason the exact same way they are trained to write stories and summarize books.
You can say “it doesn’t really reason” but it has exactly the same value as the assertion “it doesn’t really write stories or summarize books” … maybe not, but there will be a story or a summary (or a logical chain of thought) in front of you if you ask for one.
I will 100% admit to not reading papers and keeping up to date. I went ahead and spent about 30m looking up various explanations and summaries of LRMs. Ok, so you take an LLM and tell it to break the problem down first. It’s still not reasoning. It’s running a simulation of a natural language conversation, and giving you the center of mass of the statistical distribution for the intermediate steps. Does this kinda sorta replicate the sounds a human makes? Absolutely. But it’s irresponsible and unethical to make any claims that this is a human like entity you can chat with, or that it is doing any reasoning.
When I get some time I’ll check this paper out: https://ml-site.cdn-apple.com/papers/the-illusion-of-thinking.pdf
As Daniel Dennett once asked: “What is the difference between a simulated song, and a real song?”
You say it’s not reasoning, but I’ve seen it debug and fix a core dump
Birds string together words they hear when they can repeat them, and end up with the short phrases they seem to make. It’s extremely rare for them to actually understand meaning, most often it’s simply association which is why you often get nonsensical responses that still sort of make sense, or sounds out of them that sound like words but just…aren’t. The simalcrum of language without containing any. Often words can be linked by that, and our own brain wants to find words and sometimes we decipher ones, similar to seeing shapes in noise – we just tend to realize that it’s actually just recognition, not real.
What’s happening here is the equivalent of recording a bird and playing back it’s recording to itself to get a new response, as a chain. It’s predictive text feeding itself, in a simplistic but not inaccurate manner given how language models actually work at a technical level, tokenizing the input to train and create matrices of language vectors that contain word fragments, and often loop back on themselves or into yet more matrices of options. This is the “beam size” option some models have when run, selecting how many search routines should be created simultaneously for things that map to probability values that make sense.
Our own reasoning is far more complicated. Sometimes we think it’s just words, but our brain will seamlessly weave inner monologue into concepts and imagery or ideas without text and back again, sometimes into sounds or other things. We stitch together everything so seamlessly because it all actually has meaning for us.
LLMs having “reasoning” at all is operating by the Sapir-whorf hypothesis, which would imply there is no reasoning without language. And even animals can fucking reason without language. We absolutely did too. Sapir-whorf was an infantile thought experiment turned theory of language that’s been patently proven wrong even when it makes for great sci-fi (see Arrival).
This isn’t the difference between hearing a song live and played aloud, or midi/samples vs instruments. This is that part of our consciousness operates in some absolute wild ways that we can still only classify at a high level because the complexities are so far beyond what we can describe with models that, by comparison, are simplistic as hell.
Put another way, without transcriptions of “that’s right, the square hole”, if you showed two photos to a model and asked “where does this piece go” it’s just going to “see” the shape in both, recognize the image->word mapping and come up with a response fitting that, without ever being able to “realize” it can go into the square hole without being prompted, because it can’t invent.
Only parrot.
Are you familiar with latent space representation?
Because yes, that’s how LRM’s work, cycling tokens in latent space multiple times before sending to upper layers and decoding into human words
https://arxiv.org/abs/2412.06769
A couple of things:
Fair, we need to get terms straight; this is new and unstable territory. Let’s say, LLMs specifically.
Can you explain how that is different from what a human does? I read a lot about debugging, went to classes, worked examples…
In my case this is enterprise software, many products and millions of lines of code. My test and bug-fixing teams are begging for automation. Bug fixing at scale
Ok, that’s easy. If I make an LRM model of your dead grandma, is that your grandma? Why not? What’s different?
Your bug fixing teams are begging for automation. That tells me you have an unsustainable setup. You are providing a bug fix suggestion tool, I don’t see how that fixes your problem. Seems like you need better coding practices and possibly more people.