How LLMs work

See: https://openai.com/index/chain-of-thought-monitoring/

Quote below:

  Chain-of-thought (CoT) reasoning models “think” in natural language understandable by humans. Monitoring their “thinking” has allowed us to detect misbehavior such as subverting tests in coding tasks, deceiving users, or giving up when a problem is too hard.

  We believe that CoT monitoring may be one of few tools we will have to oversee superhuman models of the future.

  We have further found that directly optimizing the CoT to adhere to specific criteria (e.g. to not think about reward hacking) may boost performance in the short run; however, it does not eliminate all misbehavior and can cause a model to hide its intent. We hope future research will find ways to directly optimize CoTs without this drawback, but until then

  We recommend against applying strong optimization pressure directly to the CoTs of frontier reasoning models, leaving CoTs unrestricted for monitoring.

  We understand that leaving CoTs unrestricted may make them unfit to be shown to end-users, as they might violate some misuse policies. Still, if one wanted to show policy-compliant CoTs directly to users while avoiding putting strong supervision on them, one could use a separate model, such as a CoT summarizer or sanitizer, to accomplish that.

> Don’t be distracted by human knowledge, as AI has been historically.

> Instead focus on methods for creating knowledge that scale with computation, like search and learning.

so the lesson is choose methods that scale with computation, not just that blindly scaling up anything (data, params, people, whatever) works, it is choosing the right x axis and the right scaling laws consistently wins out in the long run despite short term wins from other methods.

1: https://x.com/RichardSSutton/status/2056419165502935198

Also would love to know if the same Legal team advised on Gemini...

There have been minor changes to the architecture over the years, but these are basically all efficiency tweaks such as various types of attention (some pioneered in the open by DeepSeek) that better scale to large context lengths, and the confusingly named "mixture of experts" architecture, but what's more notable really is how little the architecture has changed. The capability gains have been coming from better training and better data.

- V3 https://arxiv.org/abs/2412.19437

- V2 https://arxiv.org/abs/2405.04434

- R1 https://arxiv.org/abs/2501.12948 (RL applied to ML models was well-known beforehand, but they show it in the open, at scale, on big models)

Then, there's the incentive analysis. If you can see that these models empirically get better with scale, why would you swap the main architecture? Those events will be pretty rare. I'm not saying there's noone cooking a new architecture, just that it is a pretty rare event. And it would have to come from some researchers that would be happy to not publish their findings, which is not really what a sizable portion of elite researchers (obviously not all) are incentivized to do.

Of course, it's a bit of a verbal compression to claim simply 'scaled up'. They are recognisable scaled up transformers, but most new models come with a few tricks, but we're at the point where those usually are not an architectural rewrite and added to solve an explicit problem, like hallucination, not for big new capability gains.

It is the combination of LLM and agent-harnesses that make it look really smart. Agent-harness is a programmatic device that lets us tap into the vast knowledge in the LLM.

It is probabaly true that many TV-commentators fail to appreciate this fact and therefore think LLMs are super-intelligent. No, it is the combination of LLM and the programmatic agent-haness that is the breakthrough.

An interesting thought is that the LLM could in theory code the agent-harrness, start it running every time we interact with it. Currently the agent-harrness I think is pretty static I think. In theory it could be dynamically created for every task. Would that make it better don't know.

There is also the case for Markov chains being theoretically able to do these if tuned well. Or even SAT problem.

If we look beyond written languages which are late inventions of human civilization, oral languages are continuous and build with blocks not words.

Chomskyan school misled the entire field of linguistics for decades by ignoring spoken languages.

Like the best leaps in thinking, once it is made, is is immediately obvious and intuitive.

What about improving the efficiency of token consumption, etc., basically opportunities for improving cost/performance?

I keep thinking there has to be a better way to share context with models than dumping entire gigantic skill files of raw text or otherwise into them - I'm betting there's a bunch of low-hanging fruit there.

Which sums up HN these days.

MoE was also pretty straightforward, just a bit surprising how well it worked (that you can get away with just 1/32 active parameters), but most researchers would have come up with it on their own probably.

The true ground breaking papers are the first two you mentioned (transformers and gpt2), and InstructGPT was also very surprising that it worked so well.

If you're up for it I would love to know how and why positional encodings work

My next child took a completely different path to language, including skipping all the non-verbal imitations.

And then at some point, you just suddenly can two-way communicate with them when you couldn't before, and then after that, they can engage in reasoning.

The "bitter lesson" is that fake-it-till-you-make-it is a valid way of doing knowledge work.

(Or not make it, then people will just claim you're holding the LLM wrong and it's not the AI's fault.)

Then the next question becomes "HOW do they predict the next token?" There are many ways that can be done, why is this particular algorithm so GOOD?"

When people say "We don't understand how LLM works" isn't it really saying we don't understand how this specific algorithm used to predict the next token works? No, it is not, because "we" do understand how all those algorithms work there are many descriptions of them available.

So the question then really is "Why is the prediction this algorithm makes, so good, as compared to some other statistical algorithms?"

It's not about "Why does AI work so well?". It should be "Why does this particular XYZ algorithm work so well?"

The capability of the LLM is not to reason, it's to generate text that matches the patterns seen in the training corpus. It's possible that all you need to "reason" is plausible text generation. I'm not saying it's not. But nothing the LLM does fails to be explained by plausible-text-generation.

I contend that the best way to understand an LLM's capabilities is to understand the nature of the probability distribution that produced it. For instance, why does an "angry" prompt tend to produce more help than a "polite" one? Trying to explain that in terms of emotions or reasoning doesn't make sense, but it's readily possible to explain through the connections between text in the training corpus...

If you train the LLM on a corpus that shows people saying the sky is red, you get an LLM that is predisposed to say the sky is red. This is true even if it's also trained on all of the science that explains how and why the sky is blue.

If it were to "figure out" or "reason", it would not have such a predisposition to emit "red" after "the sky is" just because that matches the reward during training.

In other words, the token prediction is important because it both explains the successes AND the failures of the LLM. If there were situations in which a bird could fail to fly, then how it tried to fly would also be crucial knowledge.

But I think it will have difficulty in crossing paradigm boundaries, by simply using documentation.

The exact syntax does not matter, only the grammar. If you give it the grammar, and then the keywords, it can find something that has similar grammar and then use your keywords.

As a for instance, back in the day some academics wrote a paper that compared GPT 3.5 to a couple of inductive programming systems (including one of mine) on solving programming problems in a certain well-known esoteric language which I shall call "L". The task was to solve those programming problems one-shot. The authors asserted that the "L" problem sets were unlikely to be in 3.5's training set, but I found them without much search in a public github repo. I mean the entire dataset was right there. In this case the researchers are colleagues and friends and I know they weren't simply negligent or malicious, they just missed the fact that their "unlikely to be in the training set" data was on the web.

So I'd always assume that if an LLM can perform a task that's because it's seen examples of the task during its training.

Without forgetting that LLMs have this really shockingly powerful ability to interpolate between examples and they can improve their performance on say Task A by training on Task B, where A and B are different but similar.

e.g. they seem to get better at translating between language pairs of which they have few examples of parallel text by training on other pairs of languages for which they have more parallel text; they seem to learn something about language translation in general by training on more examples of translation. I haven't got a good reference on that handy but it's well-known (and of course over-hyped and exaggerated by tech CEOs).

So without wanting to diminish your work, I'd guess that your new language's syntax is different and novel but everything else about it is more ordinary and the similarities are such that an LLM can wing it and write you a lexer etc. After all, the whole point about parser generators and similar tools is that the task can be abstracted and separated from syntax in the first place.

In fact LLMs are very good at that sort of thing, filling in the blanks as it were. I'm old enough to remember the excitement about GPT 3.5 being able to form syntactically correct sentences with nonsensical words give to it.

For example, I just asked Chat [1]:

  Hey chat. The gostak distims the doshes. What happens to the doshes?

  The doshes get distimmed.

_________________

[1] https://chatgpt.com/c/6a242b65-e248-83ed-9a6e-f238a1e871b6

Now, take that for loop, and replace the implementation of decode(context, ++position) and pass it to a human who was bored enough to play along and use a notebook to organize their thoughts and translate them to/from this encoding (you might write a helper function to do this for the human in the front-end of the new decode() impl, but the data flow in and out of decode() will remain the same):

decode(context, position)

{

  cached string_answer = ask_human_question_via_context(context);

  return decode_human_answer_to_tokens(cached string_answer, position);

Is the output you get not thinking anymore because it passed through this harness? Did the human's mind somehow get reduced to mere interpolation?

The human mind is still a human mind. Putting a simple harness in front of a mind does not affect its fundamental properties.

In an LLM, decode() is calling into a trillion parameter connectome.

When you got a reply illustrating how incorrect that claim from your first argument was, you shifted to focusing on the other argument (the one I actually happen to agree with - the cost to society of hitching increasing dependency on big tech will make the social media harms look like childs play).

I think your argument will be better received if you focus on the very valid concerns of societal harms, and acknowledge the ways LLMs are tremendously capable, without downplaying that.

I'm with the person you replied to in seeing how capable LLMs can be when you spar with them appropriately. They confabulate, but that's your job to catch as a sparring partner. But they do bring useful knowledge of thousands of PhDs into conversations - and even if you're among the most erudite humans on the planet, this is still an asset in intellectual search for truth on many topics.

Back to the genuine problems, and they are many: this power, concentrated in the hands of big tech, is a multiplier on the power already concentrated there, with many new capabilities - especially scary being the capabilities for subtly influencing and manipulating both individual and group behavior - for profit or otherwise - by the companies or their customers, or governments, or... The possibility space of harm and abuse is large..

On net, I think we should all be pushing to educate everyone around us on the pros, the nuances, the risks, and the big cons, and working to try to build a future of offline models rather than subscription service dependency..