So my guess is that Claude’s backend is doing the same — so this hack is probably more of a loophole in token accounting that might get closed if Claude is doing what Gemini does
But then there is a comment below talking about how DeepSeek was able to get a huge improvement in compression by using visual tokens, https://news.ycombinator.com/item?id=48777848. I don't fully understand all of the underlying technical details so I am still fundamentally baffled about how going the OCR route could actually result in overall electricity/computational savings.
Whether such lossy compression is acceptable for your use case is up to you.
Multi-modal models do actually natively tokenize images, though. So it doesn’t have to be converted to text for it to work. They may do it anyway for accuracy, but it’s not at all required.
Effectively an image is scaled to a standard size, rasterized / cut up, and each cut is assigned a separate token, much in the same way text is tokenized. Train the model on this as well and you’ll end up having a model that can understand images.
I asked Gemini how to save costs and it said just send in all the images of the pages instead. Instinctively, as a developer, it's hard to fathom how sending 200 images is cheaper than sending the text, but it definitely works.
It'd be weird if they were doing this, since it would mean the context window size was a lie and that the API would presumably reject requests whose expanded form went over the 1m limit. For someone using pxpipe with an effective context compression of 90% in some instances, it'd hit the limit at barely 100k.
Edit: didn’t realize this occurred on local models(!!),
this is smarter https://news.ycombinator.com/item?id=48779884
One option, when an image is fed into an LLM, is to divide it into tiles, then those tiles pass through a 'vision encoder' neural network to make 'vision tokens' which are then input into the LLM much like text tokens are. Obviously you train the vision encoder and LLM to understand one another. This is known as an 'end-to-end OCR model'.
And it turns out, once you've trained a model to do this, you can vary the number of 'vision tokens' used to represent a given text document by scaling an image of a document up or down, and see what happens. You also get a load of other parameters like patch size and vision encoder complexity and so on.
Turns out it works really well; in some tests they used 90% fewer input tokens, but still got 97% output performance.
Some random person discovered a 60% across the board gain in all LLMs, using an extremely simple trick that none of the labs noticed in all these years. That trick being to rasterize 8bit characters into 8x8 pixels in a big image. 60% in a market worth trillions of dollars.
or
Anthropic's marketing team arbitrarily prices tokens to drive growth, according to vibes and feelings, and didn't think they needed to price images on par with text in their rush to burn cash & drive growth. Some folks take advantage of the trick during the first few days of the model's availability before Anthopic corrects their pricing, to align more proportionally with actual compute costs.
A text encoding uses 8bits per character on average, tokenization further compresses that
An image font would be 25 bits if 5x5, and most fonts are 12 pixels high
Of course it isn't efficient, this is a pricing inefficiency and a hack to exploit it (even the author describes it as an exploit)
Educate me: what is an "optical token" when dealing with LLMs?
Since then I've been using images with very simply worded prompts whenever I'm informing an agent of what is happening. Sometimes no text in the prompt at all.
It has been very very effective.
That being said, this isn't really what Karpathy was talking about. But it got me thinking a bit, and that got me to a much nicer workflow.
Also, some models still do OCR and it's usually way more expensive that way.
input tokens are cheaper than output tokens. seems like it would maybe reduce input tokens at the expense of many more output tokens if you're actually triggering OCR via thinking?
Would that reduce the number of tokens used too?
Images tho are natively compatible with Multi-Modal LLMs, so theres no image->text translation layer in between. It's that the unit of cost is different (e.g. "visual token" vs text token)
sure it was pretty resource intensity a few years before, but with turbo quant, sparse attention and various techniques, plus the advancing of hardware (dedicated prefill machine, memory pool for kv caching) the cost should be drastically reduced, and yet they still keep the same cost formula.
I can't help but laugh whenever someone proudly share how many billion input tokens they spent in their code sections and how much they saved with the subscription, meanwhile it is pretty much just electricity cost for the providers.
The image is still getting run through OCR and turned back into text before being fed into the LLM. There is no efficiency gain here, rather we have learned that Anthropic is applying a discount to text fed in via OCR.
For example:
> Honest caveat, visible in the clip: the pxpipe arm answered the count first and needed one follow-up nudge to also print the ledger balance in the requested one-line format; the plain arm followed the format on the first try. Legibility is solved on Fable — single-reply format compliance is the remaining rough edge.
If I reread this four times, I can sort of interpolate what happened, but it’s mostly pointless and confusing information.
In my experience all models do this to an extent, but Claude seems to be the worst at this. GPT 5.5 is a bit more terse but seems to compress more valuable information.
My guess is that it's a known problem, which steered the frontier models into bullet point preference.
To be fair, as you can see in the clip, the two models handled the prompt slightly differently. The pxpipe variant gave the right count initially but needed a quick follow-up to output the ledger balance in a single line. The standard model, on the other hand, nailed the formatting on its first try. We've completely solved readability here on Fable; our only real hurdle left is getting the models to follow formatting constraints perfectly on the very first reply.
Of course, this was just rewritten by another LLM.
Most LLMs by default seem to write both text and code with low information density.
You can kind of get around it by prompting them to optimize for compactness, but most just let it run with a more generic prompt.
"Elegant prose instantiated through remarkably tailored execution of written word may allow an author's desired intention to flow in a certain way to achieve a precise effect whilst simultaneously allowing said author to sound of much higher mind and thought to the reader." - I probably butchered it but my point is that AI slop seems to be the average of the outputs.
Images that look similar to others because they're average of all the current outputs. Same with music and video. We're noticing when something is AI because it has this signature that's average to other outputs.
Original content is crafted even though inspired by other works.
We're at a weird point where AI is capable but constrained.
As compute increases and AI becomes more personalised I feel the current implosion will explode again into variety.
people can make some really useful stuff with AI especially when its a domain they're already an expert in, and it would go a long ways for them to just sit and explain that 1. they used AI to help 2. their own words to explain what the heck they put together, especially if they can speak to some of the limitations AI has working with it. just goes a long way to demonstrate this guys stuff is worth tinkering with because he has a good grasp on what was created
for 99% of the stuff out there now people are literally operating in domains they don't understand at all, i just close my tab when i see the damn vibe coded readmes
It kinda makes sense too. Because while people do read code word by word, we often "glance over" it and do roughly pattern recognition on it to know what it does. Only homing in on something when we need to answer a specific question. I think humans kinda naturally do this exploit anyway
Exact details of text to image compression ratios are of course extremely dependent on the model architecture, training data, training objectives, etc., so there's probably not too much justification for generalizing to all models
An image token I recall is something like 16x16, so you get 32 bytes of overhead per pixel. And a character is minimally like 20 pixels including the whitespace, so you've jumped from 4 characters per token to maybe 12.
So 3x savings... which actually maps pretty closely to 60% savings.
It's forced by the nature of how LLMs use vector embeddings for language.
Basically, a single token in a LLM is represented as a n-element vector, where n is the "hidden dimension", also known as model dimension. In order for the model to be smart, the hidden dimension needs to be large, on the order of 2^16 on top-tier models. Elements of this vector are typically quantized to 2-byte floats, or sometimes smaller. Every possible fact is embedded as a direction in this very many dimensional vector space, and a token is related to a fact if the vector representing that token points into a similar direction as that fact. You can do vector math about these things, famously for most trained models, if you find the vector embedding for king, man, woman and queen, and calculate king - man + woman, the result is very close to queen.
(Does that mean that there are 2^16 possible different kinds facts about things in this model? No, because high-dimensional geometry is very unintuitively powerful. The facts are not axis-aligned, and they don't need to be perfectly non-orthogonal. This matters, because the numbers of individual vectors you can fit into a single 2^16 dimensional space that are orthogonal with each other (all angles 90degrees) is of course 2^16. But, if you allow for almost orthogonal vectors, the number is larger than the amount of atoms in the universe. If this sounds wacky, for people with a CS background it can help to think it working a bit like a bloom filter, in that collisions are possible. Although in actuality they are theoretical, because 2^16 is a very large number.)
A human might have written a disclaimer like this:
> When not using Fable, pxpipe may require additional follow-ups to precisely follow your formatting instructions.
This kind of garbled information dump is very inconsiderate of the reader, and all good writing is considerate of the audience consuming it.
That human rewrite is excellent. It ruthlessly cuts out the "narrative" of the test case (the transaction counts, the video clip, the "first try vs second try" details) and extracts the only piece of information an actual user reading a README cares about: what to expect when using the software.
Which suggests some ideas that should have been included in the prompt, to get closer to your ideal rewrite.
Text tokens are high-dimensional vectors, not 8 bits per character. Every token has a deep embedding, e.g. 1024 float values per text token.
DeepSeek-OCR proved 10x+ compression from visual embedding of text, which was a groundbreaking result. [1]
Very cool to see OP's project hacking on this principle. It's still not lossless, as noted in the github, but is a promising research direction.
[1] https://github.com/deepseek-ai/DeepSeek-OCR/blob/main/DeepSe...
And we're talking about images of texts, not images that represent complex imagery such as a very detailed scene or what have you.
The paper notes two things:
1) While the compression ratio for visual text is better than it is for regular text, but the absolute space required is still higher for the images. OPs were talking about the space required, not the ratio.
2) The results of the OCR must still be fed into a text-based LLM for linguistic processing. Otherwise, all you have achieved is turning an image into a bunch of text.
You are conflating tokens with embeddings.
Tokens fit in a single word, modern gpt uses a vocabulary with 200k possible values, which would fit into 18 bits.
Have a good one
The top line can be the OCR-able instruction on how to decode the rest of the image, and the rest of the image would be random-looking colourful palette. It might not even need to use 8 bits per character, since ANSI is 7 bits/character.
You can achieve this by changing the extension of an image file from .bmp to .txt
Guys, not to be mean, but maybe chill with the state of the art research and go back to studying fundamentals.
It's not a 60% percent reduction in cost for 100% of the same output. If you have a model and input text A, and you fix the seed etc. and run Text A through the model as text tokens and as compressed image tokens, you will not get identical outputs. You're specifically reducing the number of tensors needed to represent your input, which saves you on raw compute, but also by definition gives you less room to represent the information in your input. It's lossy, in other words.
Put another way, if you're using a model like Fable because you need the absolute frontier of capability and cheaper models cannot solve your tasks, then there is a very real chance that a compression strategy like this drops Fable's accuracy such that it's no longer suitable for your task. Which defeats the point of you paying for the most expensive model in the first place.
So, it's cool research. Might be useful for some people. Probably isn't something that has incredible utility in real use cases.
To me compression implies smaller size? However new line chars seems to be removed in the pic so I guess it could be expressed in fewer bytes than the original text with further compression ...
You could also imagine models where text tokens cover many characters and image tokens just a few pixels, which would invert the relationship, but this is typically suboptimal for the applications people have in mind when they train a model.
DeepSeek published a pretty well circulated paper on exactly this many months ago. It just hasn’t been attempted and shared publicly, asa retrofit, AFAIK.
Also, it’s no free lunch, the readme indicates that this “use images” hack is lossy and reduces success rates alongside the reduced cost. Most labs would focus on success increases regardless of price.
This is a lossy process, it produces worse results. It might be worth it for some situations, but applying it to everything would just be making your SOTA model worse
The image trick reduces context because it’s lossy. The README says you can’t use it for anything needing exact recall. It produces a gist of the input.
You could achieve something similar by using a small, cheap model to pre-summarize information for the expensive LLM. This is what many people do already and it’s a much better way to do it for most situations.
Also I don't think you realize how much dumb stuff is still left on the table. That the market is worth trillions is quite irrelevant here given the dynamism of the field.
Download the .bmp file and open it with 'edit' on windows or 'nano' on linux. You should see text. The bytes that precede the text are padding and a header that roughly 'tells how to interpret the data', in the sense that it describes the dimensions of the 'image'.
There's python code that shows how the file was generated, note that it just writes a string to the file and then wraps it with a .bmp header so that it can be interpreted by image viewing programs.
If you open that image with an image viewer you will see an image with colorful pixels. Whether you send it as an image or as text, the only difference lies in what you declare the file to be. For example if you change the extension, operating systems will "open" it with the appropriate program by default, but the data is the same, there's no difference between ASCII text and the encoding you describe where each subpixel describes one character, it's the same thing.
Your idea about using 7 bits instead of 8bits is more space efficient, but you'd need to decompress it so that the text is byte aligned. In terms of costs, you are not charged by the bandwidth consumed in transit, but by the computational cost, which wouldn't change as the text would need to be expanded to 8bits per char, (and then converted to the same token anyways)
Again don't want to be mean but these are 1st year comp sci topics, not at all related to LLMs. Hence why I recommend studying fundamentals more than papers that look like bleeding edge LLM research.
Sending an image of text instead of text reduces the number of input tokens, but they're still being processed by the model at the same precision. This probably also hurts performance in some way – the question is by how much.
What makes the DeepSeek-OCR and related results exciting to some researchers is less about the fact that you could devise a tokenization scheme that has fewer tokens, and more about how well it works.