This is the thesis behind the "Information Theory, Inference, and Learning Algorithms" course that was taught at Cambridge University.
> Why unify information theory and machine learning? Because they are
two sides of the same coin. In the 1960s, a single field, cybernetics, was
populated by information theorists, computer scientists, and neuroscientists,
all studying common problems. Information theory and machine learning still
belong together. Brains are the ultimate compression and communication
systems. And the state-of-the-art algorithms for both data compression and
error-correcting codes use the same tools as machine learning.
I had a long ranting comment I deleted. I just don't like this trend of people presenting work in a way that makes you think some combo of 1) they discovered from scratch themselves 2) it's new 3) they didn't try to cite or acknowledge where they learned it/point to good sources 4) they don't really care about trying to teach something deeply, they want shiny stuff that makes them seem deep.
This post references specific parts/calculations, but you'd never know it was not news if you didn't know better.
It's basic information theory, which has been around since the end of WWII. It's a common topic today because some of its subtle insights are becoming increasingly relevant in our current era of AI, as we learn to understand these black boxes.
it was vaguely in my understanding of information & intelligence with compression; it was also brought up in several of the initial trials against AI companies where they discussed how the AI is akin to compression.
Nope; there is a bit more nuance and the distinction is important.
Compression is functionally equivalent to prediction when the data distribution is exactly representative of all future problems. The story changes drastically if you want generalization -- because the test distribution could be arbitrarily different, even if it had the same support! Eg: you observe a rare edge case in your training data and (lossy) compression could simply ignore it. But if you wanted generalization in that particular part of the space -- either because an adversary was testing you, or for design freedom where you choose to build in that specific corner -- then you don't just want data compression, but good prediction performance on a test distribution which peaks in that corner.
Assuming that the training data distribution is exactly the distribution you will ever care for is implicitly doing a lot of the heavy lifting in the claim that compression = prediction, and I'm peeved at how much this statement is unthinkingly repeated like a manifesto.
There is nothing natural about the training data distribution, especially if the data generation process is exploratory while the downstream usage will be exploitative.
I’m not exactly sure what distinction you are trying to make, but a compressor that generalizes from currently observed day to potential future data _is_ a better compressor, as it will have better predictions as it observes more of the source data distribution. Maybe you are making a point about whether compressors can “learn” after compressing the input data, but this also isn’t a fundamental property of compressors. Also compressors do have to deal with non-stationary data sources, so it also isn’t correct to say a compressor only models stationary distributions.
How does that invalidate the "compression is prediction"? If the future data is different and you failed to generalize, then the failure to predict means you got worse at compressing and have to spend more bits storing the new information. Conversely, if the future data is the same as that you've seen previously, you could predict it very well, and compress better as a result.
A maximally efficient compressor for the existing data distribution is not in general (and often will not be) maximally efficient for future data. The former may only be enabled by convenient local optima of the input distribution that a compressor accounting for the latter could not take advantage of.
For instance, consider the distribution of strings drawn from the language '0+'. Now consider the same for the language '[01]+'. A compressor looking at only the strings of the first language within those of the second can do a much better job if it does not have to account for future data.
This also relates distantly to the idea of overfitting in machine learning.
If your compression algrotihm is deep enough (think LLM), it will capture a lot of abstraction, making it compress well even in future cases that differ from the passed but fit the scheme in some other way.
But not cases that don't fit the scheme, and the future always has some of those. We are continually surprised by what comes next, and have been notoriously bad at predictions that aren't clear trends (until those trends change).
Compression is not prediction, it is recall. Can we make predictions based on compression? Absolutely. Is memory encoded into physical neurons technically compression? I would argue also yes.
However, going from compression to prediction is a large jump that is unsubstantiated by this article and based on the claim that probabilistic recall is also prediction.
Two perfect counterpoints to this are markets and weather patterns. One cannot predict future events based on past performance or behavior. Change is the only thing that's constant, and chaos/entropy is everywhere we look.
For simple problems like programming, sure predictive recall works amazingly well, but let's not pretend LLMs are actually predicting something. This is exactly why LLMs suck at doing anything novel; they lack imagination and creativity.
I know less about this than every other commenter here, but both weather patterns and market performance do seem predictable based on past behavior when modeled at the right level of abstraction. “Sunshine on Monday” does not imply “rain on Tuesday”, but “cold front moving in Monday night” does. (Likewise “stock up Monday” doesn’t imply “stock down Tuesday” but “CEO arrested for fraud on Monday” does.)
I think this is relevant to the discourse on LLMs/programming because for months, people said “they’re just regurgitating their training set,” but now I think people are seeing (I am seeing) that they do learn more abstract models of the world than that. I don’t really know how, but it’s why they can generalize from other codebases and tools and so on.
This perspective is a useful source of intuition against the “LLMs can’t have new ideas, they’re just next-token-predictors” style arguments. What if you shift your perspective to thinking of training as optimization over a vast parametrized family of compression algorithms? Well, it suddenly looks a lot more plausible that “new” “ideas” can emerge from that process!
How I see it, is if the human brain does lossy compression/prediction of the natural world that learns from its "training set" (sensory inputs) and we have been able to come up with new ideas, then it seems like AI would be able to as well.
Imagine you're curve-fitting a bunch of data points on, say, the orbital motion of planets and asteroids. You get tons and tons of data on these orbital motions, and then put it into a huge black-box optimization algorithm that compresses the heck out of it. It compresess and compresses until it can't find a more compact representation, no matter how much more effort it applies. The output is a function, where you can provide an initial condition, and it gives you the predicted future orbital position at any requested time.
Of course, one thing you get out of this is a great curve-fit for your existing data, which you can interpolate to find the position of any observed planet at any desired time.
But could this function also succeed in predicting the orbital motion of objects that aren't in the dataset? If I spot a new comet, and put it into the compressed function, would I get an accurate prediction of its orbital motion, even though that object wasn't in the training data?
The answer is "it depends, but probably yes". Newton's laws of orbital motion turn out to be simple compared to the size of the training data. So if the black-box compression has done a good job, it might output that function, or a close approximation of it. With a sufficient quantity of sufficiently accurate data, it might even improve on it; random errors can't be compressed, but where the deviations between observations and Newton's law turn out not to be random but rather the influence of an unobserved gravity source, or general relativity, the black-box algorithm will likely capture that as well.
A lot of people seem to think of the training process as curve-fitting data (the "stochastic parrot" model), but I think of it more as "solving an inverse problem to approximate the unknown source that generated the training data". Machine learning has proven to be quite good at solving inverse problems, and this is just a very abstract one of them.
(A forward-problem is something like solving for the electric fields from a set of charged particles; an inverse problem, https://en.wikipedia.org/wiki/Inverse_problem, is one where you have data on the electric fields at various points and want to reconstruct the arrangement of charged particles that produced it. Or more generally, you have sampled data on the output of an unknown process, and want to reconstruct the process that produced the data).
The inverse-problem-solving happens at the ML training step. The language model itself, that comes out of that, is solving the forward-problem: it has a generative-process baked in and now it's generating new data from it. But if the training process has done a good job of compression, it will certainly be able to generate valid new ideas that aren't in the training set, because the inverse model has solved for the underlying features of the real process that generated the training data, and those features can generate additional valid outputs that it wasn't trained on.
Once MP3s were invented, I had the idea for the Apple IPOD; but obviously I didn't have a giant manufacturing wing, the ability to make small hard drives, or anything else.
I don't think Apple invented the ipod anymore than I invented it; LLMs likely would have also come to the same conclusion about an ipod like device.
Original ideas either dont exist or have a functionally irrelevent definition in comparison with inputing tokens to LLMs to get novel ideas out.
There's another element to this that I almost never see discussed. Ideas are not facts. Neither LLMs nor humans can generate new knowledge, as opposed to ideas, by thinking alone. Physical investigation and experimentation is necessary.
The exception being pure mathematics since it exists solely in the realm of ideas. I'm willing to call that knowledge, but it's still a distinction, the old analytic/synthetic dichotomy of Kant.
This is a lot less surprising when you learn how non-LZ compressors work, that is, by modeling a probability distribution and using those probabilities to encode information in the minimum number of bits required to transmit the data. A less obvious conclusion is that LZ compressors do this to implicitly, the length of each symbol they could emit (literal or match, etc.) can be converted to the probability distribution the LZ compressor induces, since the number of bits to encode the symbol is related to its probability by the information content.
A common design in compressors is to use LZ as a first step, but to then represent the constant data and/or offset-length pairs from LZ using an entropy coder.
Deflate (as used in gzip) uses a Huffman coder. LZMA (as used by xz) uses a predictive range coder. Zstandard can use either Huffman or FSE. Some high-speed compressors like LZ4 skip the entropy coding stage entirely at the expense of compression ratio.
Bzip2 is an interesting aversion of this pattern - it uses the Burrows-Wheeler transform as a first pass instead of LZ. Unfortunately, this is one of the major reasons why it's so slow.
I stumbled across a connection between LLMs and compression when researching N-dim polytope emergence in neural networks. Toy Models of Superposition (Anthropic, 2022) suggests that gradient descent can independently discover efficient geometric packing arrangements for sparse features. LVQ compression uses regular lattice structures, including some based on 4D lattices.
I found this interesting and wonder whether LLMs have a higher density ceiling, since training and inference don't rely on a fixed lattice and can instead learn their own representational geometry.
Cool visuals and breakdown. I wrote something in early 2025 about how LLMs seem to be an emergent behavior of lossy compression, but did not have the knowledge or verbiage at the time to get this detailed. In retrospect my writing seems naive and I'm happy to have found this and the Google paper linked inside. To be a fly on the wall in some of the labs, man.
Another thought that came from the same post is that, insofar as we see LLMs as human-style intelligence, they're more like stream of consciousness devices. Essentially incessant talking and buying enough time until you get to a usable answer. I think I associate some subset of intelligence with what you don't say, which is impossible with the SOC-style outputs, so this is something I think about a fair bit.
What could maybe differentiate current gen models from next gen is the ability to call tools modeled within the layers themselves, not externally. I think as far as I understand it, model trainers expect the model to do this itself in a way we don't understand or control, like a version of the bitter lesson. But I posit we can model many determinate tools as NNs themselves and figure out how to get the internal states of the LLM to make use of them during inference, e.g. calculators, indexes, citations. Just an enthusiast though, so grain of salt and all.
I was thinking about the same topic and the conclusion can be wrong. LLMs are compressors, but compressors are not LLMs. Mixing this can let you believe that you can use a compressor to do the same thing as LLMs, which you cannot.
Specifically I was thinking about a way to inject knowledge into LLMs training by using statistical properties of text in such a way that you don't have to train the LLM to achieve some level of predictions. There are actually some papers that inject n-grams statistics as a part of the neural network weights.
Any compressor actually can be used, trivially, as an autoregressive language model.
Given a context (for LLMs, this would include the entire pretraining dataset, plus the prompt), you compress `context + next_token` for every possible next token. The tokens that co-compress best with the existing context are the 'least surprising' continuations. Choose one of them and iterate.
You can easily generate text with gzip this way. It won't be very good text, because gzip compression is not as sophisticated as a transformer + SGD, but the principle is the same.
>Mixing this can let you believe that you can use a compressor to do the same thing as LLMs, which you cannot.
You can, actually! Any compressor can be losslessly converted into a generator, and vice versa.
Traditional compressors like gzip are of course very simple and can only replicate rough patterns from the input. But they are technically doing the same thing.
I agree that technically they are doing the same thing but in practice LLMs are better compressors than PNGs (learned this while I was researching this topic). That was quite surprising to me.
That sounds like boostrapping the weights involved in early layers, to obviate the need for those layers to learn (optimize) for the distribution in the training set.
Makes me wonder idly,
- is this conceptually akin in some sense to a "universal grammar," and if so
- with a broad enough training set, is there a latent durable universal grammar that might be similarly recovered and injected to the benefit of all training,
- does that grammar go beyond morphological/syntactical/grammatical features, into e.g. semantics and pragmatics
That was my thinking as well mainly to increase the speed of training. But it may turn out that the simple statistics that you can capture like this may account for 1% of the training and are likely to be captured as the first thing during the training.
But actually these techniques are used but they are hidden as speculative decoding with increasing complexity of approximations. For example you can have a part of the network that predicts the next word based on the markov chain, the next approximation is more complex etc.
I keep on seeing this claim, especially from popular creators such as 3Blue1Brown. How is this not borderline vacuous?
I'm not a LessWrong^TM rationalist guy, but one really good thought experiment I always keep in the back of my mind from them is Solomonoff induction. AIT people take it as a framework to work with - it's pretty cool, I agree. But I (and some other people, such as certain AI execs at Amazon - according to my interpretation of their public interviews) think it just highlights the trap - given an arbitrarily powerful oracle, you can get compression down pat. Like, if you assume the source is generatable with a turing machine, and you write a function to brute force over all turing machines, then whoa, your compression works. You will necessarily find the optimal compression at some point because your search function is literally searching over all possible turing machines that could've generated the input sequence, anyways (because the input sequence was generated by a turing machine)
These are the kinds of results you can get if you don't have any actual constraints on what the compressor can do.
(Of course, again - this is not the point of solomonoff induction - it's to use this as a base truth, to then layer parsimony on top of that. There are infinite number of turing machines that could match your prefix, parsimony filters, throw some bayesian inference on top of that, and you get Solomonoff induction. They constrain it afterwards. But I think to that intuition as a base whenever people claim new results.).
But I see in casual conversation, people constantly making claims like, "LLM's are so good because they compress a model of the world". What is that model then? Scott Aaronson has made points like this before - your "model" could just be a massive lookup table, so you can't just claim "compression" and win - the compressor must be reasonably small, too.
I don't object to the notion that LLM's have some notion of world models more sophisticated than memorization. That's proven by actual interventional experiments, such as the ones that actual interperability researchers do. But mere compression is vacuously powerful. "Vacuous" not in the sense that "oh, you might be suboptimal and be a little more complex", vacuous as in "the philosophical point you were trying to make is vacuous because you make a vacuously powerful statement".
(I'm not a total fan of intervention either, as an end-all gospel as some people use, but it's far, far better than not having it).
If you want to best predict what state comes next from a space of possibilities, you have to figure out how probable each next state is and pick the most probable one.
If you want to compress something, you have to figure out how probable each next next state is and assign the smallest code to the most probable state.
These two processes are essentially the same, and the resulting structure of a system which regulates either process will be similar, approaching the same structure at high confidence.
I often wonder how would language fare if we didn't have redundancy in abstractions, why do things get different terms, and if there is such a smaller set that contains everything in a lossless way (english-wise)
Small world. I just did a podcast on this same topic, but coming at it from a different direction, ie. me and my neighbor trying to beat the hutter prize for compression.
Hutter Prize being where you are paid if you can compress wikipedia small enough. LLMs do very well at that, if, big if, you ignore the cost of initial weights.
A cool Claude Shannon story:
Shannon wanted to measure how much information is actually contained in ordinary
English text. His 1948 theory said such a number must exist, but he had no way to
calculate it, because the patterns in English reach across dozens of letters and no
equation or frequency table captures all of them at once.
So instead of calculating it, he ran an experiment on a person.
He took a passage from a novel that the subject had not read, and covered it with a
card so only the text already guessed was visible. He asked the subject to name
the first letter. If the guess was wrong, he asked again, and kept asking until the
subject named the correct letter. He wrote down how many guesses it had taken,
revealed the letter, and moved the card one position to the right. Then he repeated
the process for the next letter, and the next, through the whole passage.
What this produced was not a sequence of letters but a sequence of numbers — one
number per letter, recording how many guesses that letter required. Most of the
numbers were 1, because someone fluent in English, seeing the preceding text,
usually names the next letter correctly on the first attempt.
Shannon then argued that this sequence of numbers contains exactly as much
information as the original passage.
Unrelated to the content: I was really pleased to see that this site defaults to the bare minimum for cookie consent. I reflexively clicked "Reject all" only to see that it was already the default, which threw me off.
This immediately reminded me of the Hutter Prize (http://prize.hutter1.net/) - a contest that has run since 2005(?) based on the premise that compression is closely related to intelligence.
This is the thesis behind the "Information Theory, Inference, and Learning Algorithms" course that was taught at Cambridge University.
> Why unify information theory and machine learning? Because they are two sides of the same coin. In the 1960s, a single field, cybernetics, was populated by information theorists, computer scientists, and neuroscientists, all studying common problems. Information theory and machine learning still belong together. Brains are the ultimate compression and communication systems. And the state-of-the-art algorithms for both data compression and error-correcting codes use the same tools as machine learning.
Book (creative commons): https://www.inference.org.uk/mackay/itila/book.html
Lectures: https://m.youtube.com/playlist?list=PLruBu5BI5n4aFpG32iMbdWo...
I had a long ranting comment I deleted. I just don't like this trend of people presenting work in a way that makes you think some combo of 1) they discovered from scratch themselves 2) it's new 3) they didn't try to cite or acknowledge where they learned it/point to good sources 4) they don't really care about trying to teach something deeply, they want shiny stuff that makes them seem deep.
This post references specific parts/calculations, but you'd never know it was not news if you didn't know better.
This is basically a thesis supported by Shannon’s information theory.
Any rigorous CS program should cover this in depth.
Grant Sanderson has an excellent video on the same topic [0]. It's part of a series that is ongoing.
[0] Compression is Intelligence Part 1 - https://youtu.be/l6DKRf-fAAM?si=yyLWq8x4sSRkWd98
I wonder if the author of the article knew about the series, or do they both just independently came across this topic to talk about it.
Anybody working in the field will be very familiar with these concepts.
It's basic information theory, which has been around since the end of WWII. It's a common topic today because some of its subtle insights are becoming increasingly relevant in our current era of AI, as we learn to understand these black boxes.
it was vaguely in my understanding of information & intelligence with compression; it was also brought up in several of the initial trials against AI companies where they discussed how the AI is akin to compression.
So they're both sourcing a bit broader zeitgeist.
Nope; there is a bit more nuance and the distinction is important.
Compression is functionally equivalent to prediction when the data distribution is exactly representative of all future problems. The story changes drastically if you want generalization -- because the test distribution could be arbitrarily different, even if it had the same support! Eg: you observe a rare edge case in your training data and (lossy) compression could simply ignore it. But if you wanted generalization in that particular part of the space -- either because an adversary was testing you, or for design freedom where you choose to build in that specific corner -- then you don't just want data compression, but good prediction performance on a test distribution which peaks in that corner.
Assuming that the training data distribution is exactly the distribution you will ever care for is implicitly doing a lot of the heavy lifting in the claim that compression = prediction, and I'm peeved at how much this statement is unthinkingly repeated like a manifesto.
There is nothing natural about the training data distribution, especially if the data generation process is exploratory while the downstream usage will be exploitative.
I’m not exactly sure what distinction you are trying to make, but a compressor that generalizes from currently observed day to potential future data _is_ a better compressor, as it will have better predictions as it observes more of the source data distribution. Maybe you are making a point about whether compressors can “learn” after compressing the input data, but this also isn’t a fundamental property of compressors. Also compressors do have to deal with non-stationary data sources, so it also isn’t correct to say a compressor only models stationary distributions.
How does that invalidate the "compression is prediction"? If the future data is different and you failed to generalize, then the failure to predict means you got worse at compressing and have to spend more bits storing the new information. Conversely, if the future data is the same as that you've seen previously, you could predict it very well, and compress better as a result.
A maximally efficient compressor for the existing data distribution is not in general (and often will not be) maximally efficient for future data. The former may only be enabled by convenient local optima of the input distribution that a compressor accounting for the latter could not take advantage of.
For instance, consider the distribution of strings drawn from the language '0+'. Now consider the same for the language '[01]+'. A compressor looking at only the strings of the first language within those of the second can do a much better job if it does not have to account for future data.
This also relates distantly to the idea of overfitting in machine learning.
"Compression can be prediction" would be more accurate.
That’s interesting.
Also sparked the thought that the assumption only holds if the future looks like the present.
If your compression algrotihm is deep enough (think LLM), it will capture a lot of abstraction, making it compress well even in future cases that differ from the passed but fit the scheme in some other way.
But not cases that don't fit the scheme, and the future always has some of those. We are continually surprised by what comes next, and have been notoriously bad at predictions that aren't clear trends (until those trends change).
100% agreed.
Compression is not prediction, it is recall. Can we make predictions based on compression? Absolutely. Is memory encoded into physical neurons technically compression? I would argue also yes.
However, going from compression to prediction is a large jump that is unsubstantiated by this article and based on the claim that probabilistic recall is also prediction.
Two perfect counterpoints to this are markets and weather patterns. One cannot predict future events based on past performance or behavior. Change is the only thing that's constant, and chaos/entropy is everywhere we look.
For simple problems like programming, sure predictive recall works amazingly well, but let's not pretend LLMs are actually predicting something. This is exactly why LLMs suck at doing anything novel; they lack imagination and creativity.
I know less about this than every other commenter here, but both weather patterns and market performance do seem predictable based on past behavior when modeled at the right level of abstraction. “Sunshine on Monday” does not imply “rain on Tuesday”, but “cold front moving in Monday night” does. (Likewise “stock up Monday” doesn’t imply “stock down Tuesday” but “CEO arrested for fraud on Monday” does.)
I think this is relevant to the discourse on LLMs/programming because for months, people said “they’re just regurgitating their training set,” but now I think people are seeing (I am seeing) that they do learn more abstract models of the world than that. I don’t really know how, but it’s why they can generalize from other codebases and tools and so on.
This perspective is a useful source of intuition against the “LLMs can’t have new ideas, they’re just next-token-predictors” style arguments. What if you shift your perspective to thinking of training as optimization over a vast parametrized family of compression algorithms? Well, it suddenly looks a lot more plausible that “new” “ideas” can emerge from that process!
> it suddenly looks a lot more plausible that “new” “ideas” can emerge from that process
This is not intuitive to me. It seems like a "new idea" is something that (almost by definition) isn't in the training set. Can you elaborate a bit?
Edit: but perhaps a good model could arise from training, which would be a good idea in the sense that parsimonious ideas are good scientific ideas.
How I see it, is if the human brain does lossy compression/prediction of the natural world that learns from its "training set" (sensory inputs) and we have been able to come up with new ideas, then it seems like AI would be able to as well.
Wow, I didn’t even notice how badly I underexplained that! Yeah, thanks, your edit is what I had in mind. jbay808 explains it well, too.
Imagine you're curve-fitting a bunch of data points on, say, the orbital motion of planets and asteroids. You get tons and tons of data on these orbital motions, and then put it into a huge black-box optimization algorithm that compresses the heck out of it. It compresess and compresses until it can't find a more compact representation, no matter how much more effort it applies. The output is a function, where you can provide an initial condition, and it gives you the predicted future orbital position at any requested time.
Of course, one thing you get out of this is a great curve-fit for your existing data, which you can interpolate to find the position of any observed planet at any desired time.
But could this function also succeed in predicting the orbital motion of objects that aren't in the dataset? If I spot a new comet, and put it into the compressed function, would I get an accurate prediction of its orbital motion, even though that object wasn't in the training data?
The answer is "it depends, but probably yes". Newton's laws of orbital motion turn out to be simple compared to the size of the training data. So if the black-box compression has done a good job, it might output that function, or a close approximation of it. With a sufficient quantity of sufficiently accurate data, it might even improve on it; random errors can't be compressed, but where the deviations between observations and Newton's law turn out not to be random but rather the influence of an unobserved gravity source, or general relativity, the black-box algorithm will likely capture that as well.
A lot of people seem to think of the training process as curve-fitting data (the "stochastic parrot" model), but I think of it more as "solving an inverse problem to approximate the unknown source that generated the training data". Machine learning has proven to be quite good at solving inverse problems, and this is just a very abstract one of them.
(A forward-problem is something like solving for the electric fields from a set of charged particles; an inverse problem, https://en.wikipedia.org/wiki/Inverse_problem, is one where you have data on the electric fields at various points and want to reconstruct the arrangement of charged particles that produced it. Or more generally, you have sampled data on the output of an unknown process, and want to reconstruct the process that produced the data).
The inverse-problem-solving happens at the ML training step. The language model itself, that comes out of that, is solving the forward-problem: it has a generative-process baked in and now it's generating new data from it. But if the training process has done a good job of compression, it will certainly be able to generate valid new ideas that aren't in the training set, because the inverse model has solved for the underlying features of the real process that generated the training data, and those features can generate additional valid outputs that it wasn't trained on.
Once MP3s were invented, I had the idea for the Apple IPOD; but obviously I didn't have a giant manufacturing wing, the ability to make small hard drives, or anything else.
I don't think Apple invented the ipod anymore than I invented it; LLMs likely would have also come to the same conclusion about an ipod like device.
Original ideas either dont exist or have a functionally irrelevent definition in comparison with inputing tokens to LLMs to get novel ideas out.
Incidentally, the relationship is bidirectional. You can try it out just for fun. zstd is a pretty crappy language model :)
There's another element to this that I almost never see discussed. Ideas are not facts. Neither LLMs nor humans can generate new knowledge, as opposed to ideas, by thinking alone. Physical investigation and experimentation is necessary.
The exception being pure mathematics since it exists solely in the realm of ideas. I'm willing to call that knowledge, but it's still a distinction, the old analytic/synthetic dichotomy of Kant.
This is a lot less surprising when you learn how non-LZ compressors work, that is, by modeling a probability distribution and using those probabilities to encode information in the minimum number of bits required to transmit the data. A less obvious conclusion is that LZ compressors do this to implicitly, the length of each symbol they could emit (literal or match, etc.) can be converted to the probability distribution the LZ compressor induces, since the number of bits to encode the symbol is related to its probability by the information content.
A common design in compressors is to use LZ as a first step, but to then represent the constant data and/or offset-length pairs from LZ using an entropy coder.
Deflate (as used in gzip) uses a Huffman coder. LZMA (as used by xz) uses a predictive range coder. Zstandard can use either Huffman or FSE. Some high-speed compressors like LZ4 skip the entropy coding stage entirely at the expense of compression ratio.
Bzip2 is an interesting aversion of this pattern - it uses the Burrows-Wheeler transform as a first pass instead of LZ. Unfortunately, this is one of the major reasons why it's so slow.
I stumbled across a connection between LLMs and compression when researching N-dim polytope emergence in neural networks. Toy Models of Superposition (Anthropic, 2022) suggests that gradient descent can independently discover efficient geometric packing arrangements for sparse features. LVQ compression uses regular lattice structures, including some based on 4D lattices.
I found this interesting and wonder whether LLMs have a higher density ceiling, since training and inference don't rely on a fixed lattice and can instead learn their own representational geometry.
There is Compression done by Prediction by partial matching [0]
There is the Kolmogorov Complexity [1], Normalized Information Distance [2] and Normalized compression distance [3] that correlates those.
Finally, there's the Pre-Big Bang Informational Compression and the Delayed Release of Antimatter [4]
All big {rabbit/black} holes to lose some time, if you have any.
[0] https://en.wikipedia.org/wiki/Prediction_by_partial_matching
[1] https://en.wikipedia.org/wiki/Kolmogorov_complexity
[2] https://homepages.cwi.nl/~paulv/papers/chapter08.pdf
[3] https://en.wikipedia.org/wiki/Normalized_compression_distanc...
[4] https://philarchive.org/rec/GREPBI
Author's bio:
> Annie Sexton is a Developer Educator at ngrok with a passion for nerd-sniping developers.
Cool visuals and breakdown. I wrote something in early 2025 about how LLMs seem to be an emergent behavior of lossy compression, but did not have the knowledge or verbiage at the time to get this detailed. In retrospect my writing seems naive and I'm happy to have found this and the Google paper linked inside. To be a fly on the wall in some of the labs, man.
Another thought that came from the same post is that, insofar as we see LLMs as human-style intelligence, they're more like stream of consciousness devices. Essentially incessant talking and buying enough time until you get to a usable answer. I think I associate some subset of intelligence with what you don't say, which is impossible with the SOC-style outputs, so this is something I think about a fair bit.
What could maybe differentiate current gen models from next gen is the ability to call tools modeled within the layers themselves, not externally. I think as far as I understand it, model trainers expect the model to do this itself in a way we don't understand or control, like a version of the bitter lesson. But I posit we can model many determinate tools as NNs themselves and figure out how to get the internal states of the LLM to make use of them during inference, e.g. calculators, indexes, citations. Just an enthusiast though, so grain of salt and all.
I was thinking about the same topic and the conclusion can be wrong. LLMs are compressors, but compressors are not LLMs. Mixing this can let you believe that you can use a compressor to do the same thing as LLMs, which you cannot.
Specifically I was thinking about a way to inject knowledge into LLMs training by using statistical properties of text in such a way that you don't have to train the LLM to achieve some level of predictions. There are actually some papers that inject n-grams statistics as a part of the neural network weights.
Any compressor actually can be used, trivially, as an autoregressive language model.
Given a context (for LLMs, this would include the entire pretraining dataset, plus the prompt), you compress `context + next_token` for every possible next token. The tokens that co-compress best with the existing context are the 'least surprising' continuations. Choose one of them and iterate.
You can easily generate text with gzip this way. It won't be very good text, because gzip compression is not as sophisticated as a transformer + SGD, but the principle is the same.
>Mixing this can let you believe that you can use a compressor to do the same thing as LLMs, which you cannot.
You can, actually! Any compressor can be losslessly converted into a generator, and vice versa.
Traditional compressors like gzip are of course very simple and can only replicate rough patterns from the input. But they are technically doing the same thing.
I agree that technically they are doing the same thing but in practice LLMs are better compressors than PNGs (learned this while I was researching this topic). That was quite surprising to me.
That sounds like boostrapping the weights involved in early layers, to obviate the need for those layers to learn (optimize) for the distribution in the training set.
Makes me wonder idly, - is this conceptually akin in some sense to a "universal grammar," and if so - with a broad enough training set, is there a latent durable universal grammar that might be similarly recovered and injected to the benefit of all training, - does that grammar go beyond morphological/syntactical/grammatical features, into e.g. semantics and pragmatics
That was my thinking as well mainly to increase the speed of training. But it may turn out that the simple statistics that you can capture like this may account for 1% of the training and are likely to be captured as the first thing during the training.
But actually these techniques are used but they are hidden as speculative decoding with increasing complexity of approximations. For example you can have a part of the network that predicts the next word based on the markov chain, the next approximation is more complex etc.
This paper proposes something similar where you can inject memory without training https://arxiv.org/abs/2605.16893
I keep on seeing this claim, especially from popular creators such as 3Blue1Brown. How is this not borderline vacuous?
I'm not a LessWrong^TM rationalist guy, but one really good thought experiment I always keep in the back of my mind from them is Solomonoff induction. AIT people take it as a framework to work with - it's pretty cool, I agree. But I (and some other people, such as certain AI execs at Amazon - according to my interpretation of their public interviews) think it just highlights the trap - given an arbitrarily powerful oracle, you can get compression down pat. Like, if you assume the source is generatable with a turing machine, and you write a function to brute force over all turing machines, then whoa, your compression works. You will necessarily find the optimal compression at some point because your search function is literally searching over all possible turing machines that could've generated the input sequence, anyways (because the input sequence was generated by a turing machine)
These are the kinds of results you can get if you don't have any actual constraints on what the compressor can do.
(Of course, again - this is not the point of solomonoff induction - it's to use this as a base truth, to then layer parsimony on top of that. There are infinite number of turing machines that could match your prefix, parsimony filters, throw some bayesian inference on top of that, and you get Solomonoff induction. They constrain it afterwards. But I think to that intuition as a base whenever people claim new results.).
But I see in casual conversation, people constantly making claims like, "LLM's are so good because they compress a model of the world". What is that model then? Scott Aaronson has made points like this before - your "model" could just be a massive lookup table, so you can't just claim "compression" and win - the compressor must be reasonably small, too.
I don't object to the notion that LLM's have some notion of world models more sophisticated than memorization. That's proven by actual interventional experiments, such as the ones that actual interperability researchers do. But mere compression is vacuously powerful. "Vacuous" not in the sense that "oh, you might be suboptimal and be a little more complex", vacuous as in "the philosophical point you were trying to make is vacuous because you make a vacuously powerful statement".
(I'm not a total fan of intervention either, as an end-all gospel as some people use, but it's far, far better than not having it).
The key principle is simple.
If you want to best predict what state comes next from a space of possibilities, you have to figure out how probable each next state is and pick the most probable one.
If you want to compress something, you have to figure out how probable each next next state is and assign the smallest code to the most probable state.
These two processes are essentially the same, and the resulting structure of a system which regulates either process will be similar, approaching the same structure at high confidence.
I always feel like people leave out the third case of the analogy: indexing
The article itself has decision trees for the compression explanation, which is also a lookup index.
In each case you try to recognise (re)usable structure.
Self-indexing succinct data-structures are a good example of the third side of the coin.
So it's a trinity: compression, prediction, indexing
> compressors and LLMs
Why only LLMs? All statistical models are compressor. You can say "model" and "compressor" are synonyms.
Article does not mention "embeddings" at all, even though it's commonly viewed as a compression method. Also "encoder" part on "auto-encoders".
I often wonder how would language fare if we didn't have redundancy in abstractions, why do things get different terms, and if there is such a smaller set that contains everything in a lossless way (english-wise)
Small world. I just did a podcast on this same topic, but coming at it from a different direction, ie. me and my neighbor trying to beat the hutter prize for compression.
Hutter Prize being where you are paid if you can compress wikipedia small enough. LLMs do very well at that, if, big if, you ignore the cost of initial weights.
A cool Claude Shannon story:
Sounds a lot like next token prediction to me.https://corecursive.com/the-hutter-prize/
http://prize.hutter1.net/
https://github.com/hkust-nlp/llm-compression-intelligence
https://www.princeton.edu/~wbialek/rome/refs/shannon_51.pdf
Unrelated to the content: I was really pleased to see that this site defaults to the bare minimum for cookie consent. I reflexively clicked "Reject all" only to see that it was already the default, which threw me off.
This immediately reminded me of the Hutter Prize (http://prize.hutter1.net/) - a contest that has run since 2005(?) based on the premise that compression is closely related to intelligence.
But all modeling is compression of a dataset? This is how I learned it in stats for CS majors 101.
3Blue1Brown - "Compression is Intelligence": https://www.youtube.com/watch?v=l6DKRf-fAAM
See also: Bellard's Lossless Data Compression With Neural Networks
https://news.ycombinator.com/item?id=19589848
https://news.ycombinator.com/item?id=27244004
And also the LLM version, and LLMZip
https://bellard.org/ts_zip/
https://arxiv.org/abs/2306.04050
Perhaps a similar observation; https://news.ycombinator.com/item?id=48703636 :
> Compression, Predictive modeling, or Complexity?
Perhaps a bad example: https://news.ycombinator.com/item?id=38400380 :
> "78% MNIST accuracy using GZIP in under 10 lines of code" (2023) https://news.ycombinator.com/item?id=37583593
I was rushing to post this and then found out somebody had already
You snoze, you loze!
https://www.youtube.com/watch?v=B6u-FPskfAE