In Gemini at least, if you look at how they process PDFs, they do an OCR and then feed the text + image to the model, without charging you for the text tokens (I believe).
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
This is really fascinating to me. I was reading this article and originally agreed with you, "I mean, under the covers it's got to be converting to text tokens at some point, so there is no way it's actually cheaper for Claude itself to execute."
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.
LLMs have a very bloated in-memory representation for text, on the order of megabytes of KV cache per byte of text. Meanwhile, for images a lossy representation is considered acceptable and it only takes up maybe a kilobyte of KV cache per byte of image. So if you can render your text into a hundred bytes of image per byte of text and then lossily expand it into 100 kB of KV cache per byte of text, you come out ahead!
Whether such lossy compression is acceptable for your use case is up to you.
Then you are paying for the electricity. It's not physically possible to do more computation & not use more energy b/c every arithmetic operation requires a minimum amount of energy so more operations = more energy.
I tried the same thing last year (with openai models), back then it worked to reduce prompt tokens, but you needed way more completion tokens, ultimately more expensive (and slower)
https://pagewatch.ai/blog/post/llm-text-as-image-tokens/
It’s so painful to read the LLM-compressed explanations. I can’t exactly identify what it is, but it’s an immediate tell and literally requires twice the effort to comprehend.
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 thought on this is that LLMs probably mimic writing patterns and structures from quality resources. But they don't construct a plausible thought hierarchy like an average human does, so their train of thought turns into a rollercoaster of thought. So the order of information is for humans completely out of order.
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.
dead giveaway for me that something someone made and wants to share = they dont understand what they put together enough to like speak to it with some level of authority.
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
Step back and think about it another way - ask which scenario is more likely:
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.
Nah, optical compression is a thing. You see it in a lot of different areas in ML. In this case, the "trick" has been known for a while, and belongs to a whole world of compression research. But I think where you're maybe getting mixed up is in where that 60% gain is coming from.
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 ...
The size is indeed smaller, because text tokens and image tokens are embedded as vectors of the same size, but text tokens typically only cover a few characters, while image tokens typically cover many pixels, so many that you can fit more characters in there. So the same text takes up fewer tokens as an image, and hence requires less time and memory to process.
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.
> 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 of multi-trillion dollar growth
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.
If the trick were genuinely useful, and was well circulated months ago, the resource-starved inference providers would have squeezed this trick dry already, instead of wasting 60% of their tokens, waiting for users to implement it themselves in 5 minutes of effort.
That's like saying quantization isn't real because the frontier labs aren't using it in their production inference.
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
I think you missed the part where this is a lossy technique that reduces performance.
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.
Alternative 1 isn’t all that unlikely given Opus 4.8 couldn’t do this. So it’s a recently possible hack. Not something LLM corps have been blindsided by for years. I also strongly recommend RTFA in this case, namely ”The honest part, read before relying on it”
This has been known since VLMs were a thing, that more information can be encoded visually and token efficiency is increased. But it came with performance issues (more hallucinations, etc).
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.
A token is probably not a single char, and an image is probably decomposed into tokens as well (and god knows how many tokens an image is decomposed into) which probably map to similar float-hungry vectors. Your counterargument could use a bit more flesh.
And we're talking about images of texts, not images that represent complex imagery such as a very detailed scene or what have you.
I kinda wonder if it's extracting usable context from 2D proximity between lines? Normal text input wouldn't have that kind of information (though it could, and it's arguably just a lookahead/behind of N characters on average).
People really need to read their cites and not just the summaries.
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.
huh, what if the image encoding is 8 bits per R, G, B values of the pixel, then one can encode the same amount of text in less pixel dimensions (3 letters would need 1 pixel instead of three 12x12 pixels)
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.
Not really. They arent actually using more resources this way either. This might be a fundamental inefficiency thats being removed
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
Saw a Tweet a while ago from someone (maybe Carmack, maybe Geohot, maybe Karpathy?) wondering if images were just the better option.
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.
seems really dumb and like it would need to violate basic information theory to work?
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?
I'm sorry, but this is retarded. It works, and it's clever, but but it's clearly a workaround for a pricing failure. Much like the bounty on poisonous snakes leading to people taking up snake-breeding, this just exploits and promotes waste. I think ultimately blame falls on Anthropic for the poor pricing system the enables such arbitrage. But I'm also disgusted by the inevitable tide of people exploiting this until its fixed, and creating an entirely unnecessary extra tide of digital junk.
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.
Edit: didn’t realize this occurred on local models(!!),
this is smarter https://news.ycombinator.com/item?id=48779884
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.
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
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.
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
I miss when HN was mostly people that knew what they were talking about.
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.
Educate me: what is an "optical token" when dealing with LLMs?
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)
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 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
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.
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?