Camestros Felapton

2026-09-23 19:49:19

GenAi v “Thinking” & “Reasoning”

There is a tendency for deep arguments to rapidly wander into shallow arguments about word definitions. With lots of arguments continuing about the nature of generative Ai, the obvious problemsome¹ word is “intelligence”. I’d like to say that what are being called “AI” are not intelligent but I don’t have a clear meaning for what intelligence is. We all get that though and it is also clear, that whatever Chat-GPT is doing isn’t what goes on in people’s brains.

Of course, there is no good reason why it should be. The whole history of machines is one of taking things a human did and automating the functional result of what the human (or animal) did. Sometimes, the process the machine follows is close to what a human does, and in other cases not. A sewing machine, uses a special kind of stitch process that makes sense for a sewing machine. A car resembles a horse drawn carriage only vaguely and we don’t expect the engine to have the same parts as a horse.

Ai-boosterisms² likes to pretend that these services are doing the same thing or an equivalent (but even better!) thing as a human brain, but it is more semantics than substance with the word “intelligence”. We can, collectively, extend our use of the term to cover Claude (or god-forbid Grok) or we can choose not to. Intellectually, I am open to the idea of forms of intelligence that are very different to what humans have, but I find the current technology does not fit with my vibes on that word, and that’s all we’ve got.

However, there are words that orbit “intelligence” that are not synonyms but are closely related. How do I feel about those words?

“Think” and “thinking” is an interesting one. For me it presupposes a degree of intelligence, but I don’t find it odd to say that a dog or a bird is “thinking”, in that gap between perception and action. So, for me at least, I don’t regard “thinking” as necessarily being a human-level act. How about computers? Colloquially, I’ve often taken to describe some delay from a program while a data-intensive process is completed as the machine “thinking about it”. That’s an analogy though and a degree of personification that I know is false. That a big SQL query on a database takes time is not at all like a person puzzling over something. I’d say “think”, “thinking” and “thoughts” about generative Ai is a way of rhetorically assuming what still needs to be proven i.e. that the machine can be called “intelligent”.

But let’s take a step further out and consider “reasoning”. This word is connected to “reason” which might be taken as synonym for “intelligence”, but it has more specific connotations. “Reasoning” implies a process. When I reason, I am conscious of my reasoning but consciousness is not intrinsic to the process. For example, when I cook I am conscious that I am cooking, but if I could make a machine that cooks, it wouldn’t need to be conscious, just complex.

Machines that can do reasoning tasks (or if we prefer, task that would require a human to reason if we got a human to do the task) are not new. A lot of artificial intelligence research in the past has been orientated around reasoning-like task such as playing formal games like chess, or creating mathematical proofs. Following deductive rules is not quite a synonym for “reasoning” but it’s close in a way that the other words I’ve discussed are. Such rule following behaviour is something computers can do, and the idea that somehow logical thought can be mechanised pre-dates computers.

Yes, but Large Language Models aren’t “reasoning” in the logic/deductive sense though, are they? LLMs are using a kind of stochastic, word sequence generation process based on vast statistical patterns captured from human examples. It is still mathematics but not mathematics that is analogous to logical deduction based on premises about the subject under consideration³. That’s all true or rather that WAS all true. In 2024, if you asked Chat-GPT things it mainly was doing the one very-big trick of solving the equivalent of vast multi-dimensional set of simultaneous equations about associations of words without any premises about the significance of those words.

Here I think we do have a bit of a gap between anti-genAi rhetoric and the current state of genAi. In 2024 it was easy to get Chat-GPT et al to beclown⁴ themselves by asking it basic problems, from counting letters in a word or comparing a kilo of feathers versus 2 kilos of gold. Many of those comical tricks no longer work. How come? As I mentioned before reasoning-models, in the sense of models intended to follow chains of logic etc, predate large language models. One early fix was to link up an LLM to a service like Wolfram-Alpha, an older service for solving maths and algebra problems online.

This is a powerful trick but still far from flawless. The language capacity of an LLM, process the words a human inputs into a query for a reasoning model, which does its stuff and serves back an answer, which the LLM expresses in words. That’s not how a human actually solves, say, a maths word problem but it is an the outline of it⁵.

So “reasoning”, is a trickier word. Unlike “intelligence” and “thinking” what current genAi is doing is closer in terms of my vibe-check to “reasoning”. The underlying mechanics is not the same (I don’t think my brain is doing tensor maths to find the next word I am using⁶) but it is at least sort of functionally similar.

Anyway, here is my word scorecard:

• Is genAi “intelligent”: no
• Do genAis “think”: no, not really
• Can genAis “reason”: maybe a bit of stretch but maybe yes

Anyway, I think if I solve a online crossword, I might be able to navigate to Betelgeuse.

1. The spell checker is saying “problemsome” isn’t a word but I’m tired of “problematic”. To me problematic sounds like a 1950s machine for dispensing problems without the need of a shop assistant. ↩︎
2. “boosterism” also is not a word the spell checker recognises. ↩︎
3. There is still logic there but at deeper layer, just as there is logic in any computer program. ↩︎
4. Oh come on, “beclown” is definitely a word. ↩︎
5. You will see in problem solving literature that kind of word-to-maths-to-word structure for describing how we solve maths problems but it is intended as a simplification. Introspection about your own process will show that things can involve more back and forth. ↩︎
6. But maybe it is which means you could repurpose a human brain to calculate general relativity problems for near-light speed navigation. Or maybe you can even if it isn’t, if we can map “find the next plausible word in a sentence” as a problem to a huge maths calculation for a computer than maybe it works vice-versa with however a human brain finds a word. ↩︎