Chinese AI company DeepSeek has just put out something that has the AI world excited. DeepSeek-OCR, its newly open-sourced model, does something completely counterintuitive with information handling: instead of inputting text directly into language models like everyone else, they’re converting text to images first.
Text is converted to images, and somehow, it makes the AI systems work better.The process is insane-sounding at first, but the reward does the talking.
DeepSeek’s own studies show that their model compresses information ten times better than traditional text-based processing. 10 text tokens reduce to just 1 “vision token” and retains 97% accuracy, as they demonstrate in their technical report. Even compressing down to 20 times, accuracy is maintained at about 60%.
Visual Compression of DeepSeek and the Context Window Revolution
What exactly does this mean? Language models now have the capacity to store and process ten times more information in the same space. Think of moving from a miniature filing cabinet to an entire warehouse, but without occupying any additional physical space.
One of the biggest limitations in current AI is the context window basically, how much text can be held in a language model’s “working memory” at one time while generating responses. Current models can only do so much before they start to forget previous parts of the document or conversation.
DeepSeek’s visual compression method significantly expands these windows of context. Instead of slicing up documents into very small pieces and delivering them byte by byte, you may be able to dump the system with huge amounts of information all at once.
How DeepSeek’s AI Loads Your Firm’s Entire Knowledge Base?
For companies, it’s gigantic. Picture being able to load the entire body of knowledge for your firm, all internal documents, or an entire codebase into an AI system’s memory all at once. No more digging through files individually or painstakingly deciding which papers to include. Just load it all in and let the AI play with it all at once.
The research has caught the attention of a number of the industry’s leading players. Andrej Karpathy, a co-founder at OpenAI, questioned whether the consequences were potentially even more far-reaching than initially apparent. He raised an interesting question: maybe text tokens have been the incorrect solution all along.
“The more interesting part to me…is whether pixels are better inputs to LLMs than text,” Karpathy X-posted. “Maybe it is more sensible that all LLM inputs should always be images. Even if you do happen to have plain text input, maybe you’d prefer to render it and then input that in.”
It is an extreme idea that challenges fundamental assumptions about how we’ve been building these systems.
The beauty of DeepSeek’s platform is that nothing is converted manually on the user’s part. The model itself converts the input text to 2D images internally, applies it to its vision encoder, and then works with the compressed visual representation in the background.
DeepSeek’s Visual Token Approach and the Future of AI Memory
Former quantitative trader Jeffrey Emanuel sees gargantuan real-world potential. “You could just stuff all of a company’s most significant internal reports into a prompt introduction and save this with OpenAI and just add your specific question or prompt on top of that,” he explained. No search tools are needed, and it is still quick and affordable.
He also mentioned that the developers can pass an entire codebase to the model one time and then only refresh it with each new update. The model keeps track of the newest version without reloading in full each time.
Of course, there is not exactly a walk in the park. DeepSeek’s work primarily demonstrates good data storage and reconstruction. More questionable is whether language models can reason just as well over these visual tokens as they can over normal text tokens.
There are also real-world issues to sort out, such as various image resolutions to handle or color difference that might affect compression quality.
Since DeepSeek opened up the model, developers have begun experimenting with it already. The prospect of frontier models with 10 or 20 million token context windows has everyone anticipating possibilities that were unimaginable a few months ago.
The research also indicates fascinating links to human memory patterns. The use of visual images to structure and remember information is similar to the ancient “memory palace” technique in which people use spatial and visual landmarks in remembering massive amounts of information.
Whether or not this visual approach becomes the new standard or simply a genius alternative only time will tell. But DeepSeek made one thing clear: there’s still plenty of room to reinvent how we’re building AI systems. Sometimes the best solution is the one nobody would’ve even conceived of.RetryClaude can err. Double-check responses.
