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.
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.
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.
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.
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.
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.
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.
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.
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.
This immediately reminded me of the Hutter Price (http://prize.hutter1.net/) - a contest that has run since 2005(?) based on the premise that compression is closely related to intelligence.
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
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.
That’s interesting.
Also sparked the thought that the assumption only holds if the future looks like the present.
100% agreed.
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.
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.
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.
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.
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 :)
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.
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 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
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
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.
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.
https://corecursive.com/the-hutter-prize/
http://prize.hutter1.net/
https://github.com/hkust-nlp/llm-compression-intelligence
> 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".
This immediately reminded me of the Hutter Price (http://prize.hutter1.net/) - a contest that has run since 2005(?) based on the premise that compression is closely related to intelligence.
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
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