[ljoshua]

Less a technical comment and more just a mind-blown comment, but I still can’t get over just how much data is compressed into and available in these downloadable models. Yesterday I was on a plane with no WiFi, but had gemma3:12b downloaded through Ollama. Was playing around with it and showing my kids, and we fired history questions at it, questions about recent video games, and some animal fact questions. It wasn’t perfect, but holy cow the breadth of information that is embedded in an 8.1 GB file is incredible! Lossy, sure, but a pretty amazing way of compressing all of human knowledge into something incredibly contained.

[rain1]

It's extremely interesting how powerful a language model is at compression.

When you train it to be an assistant model, it's better at compressing assistant transcripts than it is general text.

There is an eval which I have a lot of interested in and respect for https://huggingface.co/spaces/Jellyfish042/UncheatableEval called UncheatableEval, which tests how good of a language model an LLM is by applying it on a range of compression tasks.

This task is essentially impossible to 'cheat'. Compression is a benchmark you cannot game!

[soulofmischief]

Knowledge is learning relationships by decontextualizing information into generalized components. Application of knowledge is recontextualizing these components based on the problem at hand.

This is essentially just compression and decompression. It's just that with prior compression techniques, we never tried leveraging the inherent relationships encoded in a compressed data structure, because our compression schemes did not leverage semantic information in a generalized way and thus did not encode very meaningful relationships other than "this data uses the letter 'e' quite a lot".

A lot of that comes from the sheer amount of data we throw at these models, which provide enough substrate for semantic compression. Compare that to common compression schemes in the wild, where data is compressed in isolation without contributing its information to some model of the world. It turns out that because of this, we've been leaving a lot on the table with regards to compression. Another factor has been the speed/efficiency tradeoff. GPUs have allowed us to put a lot more into efficiency, and the expectations that many language models only need to produce text as fast as it can be read by a human means that we can even further optimize for efficiency over speed.

Also, shout out to Fabrice Bellard's ts_zip, which leverages LLMs to compress text files. https://bellard.org/ts_zip/

[TeMPOraL]

And of course, once we extended lossy compression to make use of the semantic space, we started getting compression artifacts in semantic space - aka "hallucinations".

[soulofmischief]

That seems worthy of a blog post!

[TeMPOraL]

I don't know, it's not that profound of an insight. You throw away color information, the image gets blocky. You throw away frequency information, the image gets blurry. You throw away semantic information, shit stops making sense :).

Still, if someone would turn that into a blog post, I'd happily read it.

[soulofmischief]

There's more to it than that. You can draw strong analogies and also discuss where the analogy suffers. For example, you can compare decreased performance with accurately recalling specific information with high-frequency attenuation in lossy codecs.

[MPSimmons]

Agreed. It's basically lossy compression for everything it's ever read. And the quantization impacts the lossiness, but since a lot of text is super fluffy, we tend not to notice as much as we would when we, say, listen to music that has been compressed in a lossy way.

[entropicdrifter]

It's a bit like if you trained a virtual band to play any song ever, then told it to do its own version of the songs. Then prompted it to play whatever specific thing you wanted. It won't be the same because it kinda remembers the right thing sorta, but it's also winging it.

[arcticbull]

I've been referring to LLMs as JPEG for all the world's data, and people have really started to come around to it. Initially most folks tended to outright reject this comparison.

[simonw]

Ted Chiang wrote a great piece about that: https://www.newyorker.com/tech/annals-of-technology/chatgpt-...

I think it's a solid description for a raw model, but it's less applicable once you start combining an LLM with better context and tools.

What's interesting to me isn't the stuff the LLM "knows" - it's how well an LLM system can serve me when combined with RAG and tools like web search and access to a compiler.

The most interesting developments right now are models like Gemma 3n which are designed to have as much capability as possible without needing a huge amount of "facts" baked into them.

[exe34]

Wikipedia is about 24GB, so if you're allowed to drop 1/3 of the details and make up the missing parts by splicing in random text, 8GB doesn't sound too bad.

To me the amazing thing is that you can tell the model to do something, even follow simple instructions in plain English, like make a list or write some python code to do $x, that's the really amazing part.

[Nevermark]

It blows my mind that I can ask for 50 synonyms, instantly get a great list with great meaning summaries.

Then ask for the same list sorted and get that nearly instantly,

These models have a short time context for now, but they already have a huge “working memory” relative to us.

It is very cool. And indicative that vastly smarter models are going to be achieved fairly easily, with new insight.

Our biology has had to ruthlessly work within our biological/ecosystem energy envelope, and with the limited value/effort returned by a pre-internet pre-vast economy.

So biology has never been able to scale. Just get marginally more efficient and effective within tight limits.

Suddenly, (in historical, biological terms), energy availability limits have been removed, and limits on the value of work have compounded and continue to do so. Unsurprising that those changes suddenly unlock easily achieved vast untapped room for cognitive upscaling.

[Wowfunhappy]

> These models [...] have a huge “working memory” relative to us. [This is] indicative that vastly smarter models are going to be achieved fairly easily, with new insight.

I don't think your second sentence logically follows from the first.

Relative to us, these models:

- Have a much larger working memory.

- Have much more limited logical reasoning skills.

To some extent, these models are able to use their superior working memories to compensate for their limited reasoning abilities. This can make them very useful tools! But there may well be a ceiling to how far that can go.

When you ask a model to "think about the problem step by step" to improve its reasoning, you are basically just giving it more opportunities to draw on its huge memory bank and try to put things together. But humans are able to reason with orders of magnitude less training data. And by the way, we are out of new training data to give the models.

[exe34]

> But humans are able to reason with orders of magnitude less training data.

Common belief, but false. You start learning from inside the womb. The data flow increases exponentially when you open your eyes and then again when you start manipulating things with your hands and mouth.

> When you ask a model to "think about the problem step by step" to improve its reasoning, you are basically just giving it more opportunities to draw on its huge memory bank and try to put things together.

We do the same with children. At least I did it to my classmates when they asked me for help. I'd give them a hint, and ask them to work it out step by step from there. It helped.

[Wowfunhappy]

> Common belief, but false. You start learning from inside the womb. The data flow increases exponentially when you open your eyes and then again when you start manipulating things with your hands and mouth.

But you don't get data equal to the entire internet as a child!

> We do the same with children. At least I did it to my classmates when they asked me for help. I'd give them a hint, and ask them to work it out step by step from there. It helped.

And I do it with my students. I still think there's a difference in kind between when I listen to my students (or other adults) reason through a problem, and when I look at the output of an AI's reasoning, but I admittedly couldn't tell you what that is, so point taken. I still think the AI is relying far more heavily on its knowledge base.

[jacobr1]

There seems to be lots of mixed data points on this, but to some extent there is knowledge encoded into the evolutionary base state of the new human brain. Probably not directly as knowledge, but "primed" to quickly to establish relevant world models and certain types of reasoning with a small number of examples.

[oceanplexian]

Your field of vision is equivalent to something like 500 Megapixels. And assume it’s uncompressed because it’s not like your eyeballs are doing H.264.

Given vision and the other senses, I’d argue that your average toddler has probably trained on more sensory information than the largest LLMs ever built long before they learn to talk.

[antonvs]

> Have much more limited logical reasoning skills.

Relative to the best humans, perhaps, but I seriously doubt this is true in general. Most people I work with couldn’t reason nearly as well through the questions I use LLMs to answer.

It’s also worth keeping in mind that having a different approach to reasoning is not necessarily equivalent to a worse approach. Watch out for cherry-picking the cons of its approach and ignoring the pros.

[exe34]

> Relative to the best humans,

For some reason, the bar for AI is always against the best possible human, right now.

[antonvs]

It seems that 90% of discussion about AI boils down to people who feel threatened by it in some way, and are lashing out in irrational ways as a result. (Source for 90% figure: Sturgeon's Law.)

[jacobr1]

> And by the way, we are out of new training data to give the models.

Only easily accessible text data. We haven't really started using video at scale yet for example. It looks like data for specific tasks goes really far too ... for example agentic coding interactions aren't something that has generally been captured on the internet. But capturing interactions with coding agents, in combination with the base-training of existing programming knowledge already captured is resulting in significant performance increases. The amount of specicialed data we might need to gather or synthetically generate is perhaps orders of magnitude less that presumed with pure supervised learning systems. And for other applications like industrial automation or robotics we've barely started capturing all the sensor data that lives in those systems.

[Nevermark]

My response completely acknowledged their current reasoning limits.

But in evolutionary time frames, clearly those limits are lifting extraordinarily quickly. By many orders of magnitude.

And the point I made, that our limits were imposed by harsh biological energy and reward limits, vs. todays models (and their successors) which have access to relatively unlimited energy, and via sharing value with unlimited customers, unlimited rewards, stands.

It is a much simpler problem to improve digital cognition in a global ecosystem of energy production, instant communication and global application, than it was for evolution to improve an individual animals cognition in the limited resources of local habitats and their inefficient communication of advances.

[b112]

Not to mention, Language Modeling is Compression https://arxiv.org/pdf/2309.10668

So text wikipedia at 24G would easily hit 8G with many standard forms of compression, I'd think. If not better. And it would be 100% accurate, full text and data. Far more usable.

It's so easy for people to not realise how massive 8GB really is, in terms of text. Especially if you use ascii instead of UTF.

[horsawlarway]

The 24G is the compressed number.

They host a pretty decent article here: https://en.wikipedia.org/wiki/Wikipedia:Size_of_Wikipedia

The relevant bit:

> As of 16 October 2024, the size of the current version including all articles compressed is about 24.05 GB without media.

[b112]

Nice link, thanks.

Well I'll fallback position, and say one is lossy, the other not.

[thecosas]

A neat project you (and others) might want to check out: https://kiwix.org/

Lots of various sources that you can download locally to have available offline. They're even providing some pre-loaded devices in areas where there may not be reliable or any internet access.

[nico]

For reference (according to Google):

> The English Wikipedia, as of June 26, 2025, contains over 7 million articles and 63 million pages. The text content alone is approximately 156 GB, according to Wikipedia's statistics page. When including all revisions, the total size of the database is roughly 26 terabytes (26,455 GB)

[sharkjacobs]

better point of reference might be pages-articles-multistream.xml.bz2 (current pages without edit/revision history, no talk pages, no user pages) which is 20GB

https://en.wikipedia.org/wiki/Wikipedia:Database_download#Wh...?

[inopinatus]

this is a much more deserving and reliable candidate for any labels regarding the breadth of human knowledge.

[wahnfrieden]

it barely touches the surface

[inopinatus]

regarding depth, not breadth, certainly

[pcrh]

Wikipedia itself describes its size as ~25GB without media [0]. And it's probably more accurate and with broader coverage in multiple languages compared to the LLM downloaded by the GP.

https://en.wikipedia.org/wiki/Wikipedia:Size_of_Wikipedia

[pessimizer]

Really? I'd assume that an LLM would deduplicate Wikipedia into something much smaller than 25GB. That's its only job.

[crazygringo]

> That's its only job.

The vast, vast majority of LLM knowledge is not found in Wikipedia. It is definitely not its only job.

[Tostino]

When trained on next word prediction with the standard loss function, by definition it is it's only job.

[mapt]

What happens if you ask this 8gb model "Compose a realistic Wikipedia-style page on the Pokemon named Charizard"?

How close does it come?

[tasuki]

8.1 GB is a lot!

It is 64,800,000,000 bits.

I can imagine 100 bits sure. And 1,000 bits why not. 10,000 you lose me. A million? That sounds like a lot. Now 64 million would be a number I can't well imagine. And this is a thousand times 64 million!

[swyx]

the study of language models from an information theory/compression POV is a small field but increasingly impt for efficiency/scaling - we did a discussion about this today https://www.youtube.com/watch?v=SWIKyLSUBIc&t=2269s

[divbzero]

The Encyclopædia Britannica has about 40,000,000 words [1] or about 0.25 GB if you assume 6 bytes per word. It’s impressive but not outlandish that an 8.1 GB file could encode a large swath of human information.

[1]: https://en.wikipedia.org/wiki/Encyclopædia_Britannica

[agumonkey]

Intelligence is compression some say

[Nevermark]

Very much so!

The more and faster a “mind” can infer, the less it needs to store.

Think how much fewer facts a symbolic system that can perform calculus needs to store, vs. an algebraic, or just arithmetic system, to cover the same numerical problem solving space. Many orders of magnitude less.

The same goes for higher orders of reasoning. General or specific subject related.

And higher order reasoning vastly increases capabilities extending into new novel problem spaces.

I think model sizes may temporarily drop significantly, after every major architecture or training advance.

In the long run, “A circa 2025 maxed M3 Ultra Mac Studio is all you need!” (/h? /s? Time will tell.)

[agumonkey]

I don't know who else took notes by diffing their own assumptions with lectures / talks. There was a notion of what's really new compared to previous conceptual state, what adds new information.

[tshaddox]

Some say that. But what I value even more than compression is the ability to create new ideas which do not in any way exist in the set of all previously-conceived ideas.

[benreesman]

I'm toying with the phrase "precedented originality" as a way to describe the optimal division of labor when I work with Opus 4 running hot (which is the first one where I consistently come out ahead by using it). That model at full flog seems to be very close to the asymptote for the LLM paradigm on coding: they've really pulled out all the stops (the temperature is so high it makes trivial typographical errors, it will discuss just about anything, it will churn for 10, 20, 30 seconds to first token via API).

Its good enough that it has changed my mind about the fundamental utility of LLMs for coding in non-Javascript complexity regimes.

But its still not an expert programmer, not by a million miles, there is no way I could delegate my job to it (and keep my job). So there's some interesting boundary that's different than I used to think.

I think its in the vicinity of "how much precedent exists for this thought or idea or approach". The things I bring to the table in that setting have precedent too, but much more tenuously connected to like one clear precedent on e.g. GitHub, because if the thing I need was on GitHub I would download it.

[goatlover]

How well does that apply to robotics or animal intelligence? Manipulating the real world is more fundamental to human intelligence than compressing text.

[ToValueFunfetti]

Under the predictive coding model (and I'm sure some others), animal intelligence is also compression. The idea is that the early layers of the brain minimize how surprising incoming sensory signals are, so the later layers only have to work with truly entropic signal. But it has non-compression-based intelligence within those more abstract layers.

[goatlover]

I just wonder if neuroscientists use that kind of model.

[ToValueFunfetti]

I doubt there's any consensus on one model, but it's certainly true that many neuroscientists are using the predictive coding at least some of the time

https://scholar.google.com/scholar?hl=en&as_sdt=0%2C36&q=pre...

[hamilyon2]

Crystallized intelligence is. I am not sure about fluid intelligence.

[antisthenes]

Fluid intelligence is just how quickly you acquire crystallized intelligence.

It's the first derivative.

[agumonkey]

Talking about that, people designed a memory game, dual n back, which allegedly improve fluid intelligence.

[penguin_booze]

I don't know why, but I was reminded of Douglas Hofstadter's talk: Analogy is cognition: https://www.youtube.com/watch?v=n8m7lFQ3njk&t=964s.

[dgrabla]

Back in the '90s, we joked about putting “the internet” on a floppy disk. It’s kind of possible now.

[Lu2025]

Yeah, those guys managed to steal the internet.

[Wowfunhappy]

How does this compare to, say, the compression ratio of a lossless 8K video and a 240p Youtube stream of the same video?

[mr_toad]

I will never tire of pointing out that machine learning models are compression algorithms, not compressed data.

[inopinatus]

I kinda made an argument the other day that they are high-dimensional lossy decompression algorithms, which might be the same difference but looking the other way through the lens.

[dcl]

ML algorithms are compression algorithms, the trained models are compressed data.

[ysofunny]

they're an upgraded version of self-executable zip files that compresses knowledge like mp3 compresses music, without knowing exactly wtf are either music nor knowledge

the self-execution is the interactive chat interface.

wikipedia gets "trained" (compiled+compressed+lossy) into an executable you can chat with, you can pass this through another pretrained A.I. than can talk out the text or transcribe it.

I think writing compilers is now an officially a defunct skill of historical and conservation purposes more than anything else; but I don't like saying "conservation", it's a bad framing, I rather say "legacy connectivity" which is a form of continuity or backwards compatibility

[Nevermark]

It is truly incredible.

One factor, is the huge redundancies pervasive in our communication.

(1) There are so many ways to say the same thing, that (2) we have to add even more words to be precise at all. Without a verbal indexing system we (3) spend many words just setting up context for what we really want to say. And finally, (4) we pervasively add a great deal of intentionally non-informational creative and novel variability, and mood inducing color, which all require even more redundancy to maintain reliable interpretation, in order to induce our minds to maintain attention.

Our minds are active resistors of plain information!

All four factors add so much redundancy, it’s probably fair to say most of our communication (by bits, characters, words, etc., may be 95%?, 98%? or more!) pure redundancy.

Another helpful compressor, is many facts are among a few “reasonably expected” alternative answers. So it takes just a little biasing information to encode the right option.

Finally, the way we reason seems to be highly common across everything that matters to us. Even though we have yet to identify and characterize this informal human logic. So once that is modeled, that itself must compress a lot of relations significantly.

Fuzzy Logic was a first approximation attempt at modeling human “logic”. But has not been very successful.

Models should eventually help us uncover that “human logic”, by analyzing how they model it. Doing so may let us create even more efficient architectures. Perhaps significantly more efficient, and even provide more direct non-gradient/data based “thinking” design.

Nevertheless, the level of compression is astounding!

We are far less complicated cognitive machines that we imagine! Scary, but inspiring too.

I personally believe that common PCs of today, maybe even high end smart phones circa 2025, will be large enough to run future super intelligence when we get it right, given internet access to look up information.

We have just begun to compress artificial minds.

[holoduke]

Yea. Same for a 8gb stable diffusion image generator. Sure not the best quality. But there is so much information inside.

[ljlolel]

How big is Wikipedia text? Within 3X that size with 100% accuracy

[phkahler]

Google AI response says this for compressed size of wikipedia:

"The English Wikipedia, when compressed, currently occupies approximately 24 GB of storage space without media files. This compressed size represents the current revisions of all articles, but excludes media files and previous revisions of pages, according to Wikipedia and Quora."

So 3x is correct but LLMs are lossy compression.

[stronglikedan]

I've been doing the AI course on Brilliant lately, and it's mindblowing the techniques that they come up with to compress the data.

[tomkaos]

Same thing with image model. 4 Go stable diffusion model can draw and represent anything humanity know.

[alternatex]

How about a full glass of wine? Filled to the brim.

[Workaccount2]

I don't like the term "compression" used with transformers because it gives the wrong idea about how they function. Like that they are a search tool glued onto a .zip file, your prompts are just fancy search queries, and hallucinations are just bugs in the recall algo.

Although strictly speaking they have lots of information in a small package, they are F-tier compression algorithms because the loss is bad, unpredictable, and undetectable (i.e. a human has to check it). You would almost never use a transformer in place of any other compression algorithm for typical data compression uses.

[Wowfunhappy]

A .zip is lossless compression. But we also have plenty of lossy compression algorithms. We've just never been able to use lossy compression on text.

[Workaccount2]

>We've just never been able to use lossy compression on text.

...and we still can't. If your lawyer sent you your case files in the form of an LLM trained on those files, would you be comfortable with that? Where is the situation you would compress text with an LLM over a standard compression algo? (Other than to make an LLM).

Other lossy compression targets known superfluous information. MP3 removes sounds we can't really hear, and JPEG works by grouping uniform color pixels into single chunks of color.

LLM's kind of do their own thing, and the data you get back out of them is correct, incorrect, or dangerously incorrect (i.e. is plausible enough to be taken as correct), with no algorithmic way to discern which is which.

So while yes, they do compress data and you can measure it, the output of this "compression algorithm" puts in it the same family as a "randomly delete words and thesaurus long words into short words" compression algorithms. Which I don't think anyone would consider to compress their documents.

[tshaddox]

> If your lawyer sent you your case files in the form of an LLM trained on those files, would you be comfortable with that?

If the LLM-based compression method was well-understood and demonstrated to be reliable, I wouldn't oppose it on principle. If my lawyer didn't know what they were doing and threw together some ChatGPT document transfer system, of course I wouldn't trust it, but I also wouldn't trust my lawyer if they developed their own DCT-based lossy image compression algorithm.

[antonvs]

> LLM's kind of do their own thing, and the data you get back out of them is correct, incorrect, or dangerously incorrect (i.e. is plausible enough to be taken as correct), with no algorithmic way to discern which is which.

Exactly like information from humans, then?

[esafak]

People summarize (compress) documents with LLMs all day. With legalese the application would be to summarize it in layman's terms, while retaining the original for legal purposes.

[Workaccount2]

Yes, and we all know (ask teachers) how reliable those summaries are. They are randomly lossy, which makes them unsuitable for any serious work.

I'm not arguing that LLMs don't compress data, I am arguing that they are technically compression tools, but not colloquially compression tools, and the overlap they have with colloquial compression tools is almost zero.

[menaerus]

At this moment LLMs are used for much of the serious work across the globe so perhaps you will need to readjust your line of thinking. There's nothing inherently better or more trustworthy to have a person compile some knowledge than, let's say, a computer algorithm in this case. I place my bets on the latter to have better output.

[Wowfunhappy]

But lossy compression algorithms for e.g. movies and music are also non-deterministic.

I'm not making an argument about whether the compression is good or useful, just like I don't find 144p bitrate starved videos particularly useful. But it doesn't seem so unlike other types of compression to me.

[esafak]

> They are randomly lossy, which makes them unsuitable for any serious work.

Ask ten people and they'll give ten different summaries. Are humans unsuitable too?

[1718627440]

SMS codes are kind of a lossy text-compression.

[angusturner]

There is an excellent talk by Jack Rae called “compression for AGI”, where he shows (what I believe to be) a little known connection between transformers and compression;

In one view, you can view LLMs as SOTA lossless compression algorithms, where the number of weights don’t count towards the description length. Sounds crazy but it’s true.

[swyx]

his talk here https://www.youtube.com/watch?v=dO4TPJkeaaU

and his last before departing for Meta Superintelligence https://www.youtube.com/live/U-fMsbY-kHY?si=_giVEZEF2NH3lgxI...

[Workaccount2]

A transformer that doesn't hallucinate (or knows what is a hallucination) would be the ultimate compression algorithm. But right now that isn't a solved problem, and it leaves the output of LLMs too untrustworthy to use over what are colloquially known as compression algorithms.

[Nevermark]

It is still task related.

Compressing a comprehensive command line reference via model might introduce errors and drop some options.

But for many people, especially new users, referencing commands, and getting examples, via a model would delivers many times the value.

Lossy vs. lossless are fundamentally different, but so are use cases.