[windowshopping]

The part that eludes me is how you get from this to the capability to debug arbitrary coding problems. How does statistical inference become reasoning?

For a long time, it seemed the answer was it doesn't. But now, using Claude code daily, it seems it does.

[ferris-booler]

IMO your question is the largest unknown in the ML research field (neural net interpretability is a related area), but the most basic explanation is "if we can always accurately guess the next 'correct' word, then we will always answer questions correctly".

An enormous amount of research+eng work (most of the work of frontier labs) is being poured into making that 'correct' modifier happen, rather than just predicting the next token from 'the internet' (naive original training corpus). This work takes the form of improved training data (e.g. expert annotations), human-feedback finetuning (e.g. RLHF), and most recently reinforcement learning (e.g. RLVR, meaning RL with verifiable rewards), where the model is trained to find the correct answer to a problem without 'token-level guidance'. RL for LLMs is a very hot research area and very tricky to solve correctly.

[fc417fc802]

Because it's not statistical inference on words or characters but rather stacked layers of statistical inference on ~arbitrarily complex semantic concepts which is then performed recursively.

[love2read]

This answer makes sense if you know that LLMs have layers, if you don't this answer is not super informative.

If I were to describe this to a nontechnical person, I would say:

LLMs are big stacks of layers of "understanders" that each teach the next guy something.

Imagine you are making a large language model that has 4 layers. Each layer will talk to it's immediate neighbor.

The first layer will get the bare minimum, in the LLM's of today, that's groups of letters that are common to come up together, called "tokens". This layer will try to derive a bit of meaning to tell the next layer, such as grouping of letters into words.

The next layer may be a little bit more semantic, for example interpreting that the word "hot" immediately followed by the word "dog" maps to a phrase "hot dog".

The layer after that, becoming a bit more intelligent given it's predecessors have already had some chances at smaller interpretations may now try to group words into bigger blobs, such as "i want a hot dog" as one combined phrase rather than a set of separated concepts.

The final layer may do something even more intelligent afterward, like realize that this is a quote in a book.

The point is that each layer tries to add a little meaning for the next layer.

I want to stress this: the layers do not actually correspond to specific concepts the way I just expressed, the point is that each layer adds a bit more "semantic meaning" for the next layer.

[mike_hearn]

DNNs aren't really "statistical" inference in the way most people would understand the term statistics. The underlying maths owes much more to calculus than statistics. The model isn't just encoding statistics about the text it was trained on, it's attempting to optimize a solution to the problem of picking the next token with all the complexity that goes into that.

[antonvs]

One problem is that "statistical inference" is overly reductive. Sure, there's a statistical aspect to the computations in a neural network, but there's more to it than that. As there is in the human brain.