DeepSeek Solved AI's Memory Problem

Author: Two Minute Papers · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek Solved AI’s Memory Problem · Channel: Two Minute Papers · 05:14 · Watch on YouTube

Topics: Data, databases and embeddings, Prompt and context engineering, AI infrastructure and hardware, Multimodal models, Open source and software ecosystems

Summary

The video presents DeepSeek OCR as a way to reduce the cost of AI memory by converting documents into images and compressing them into a smaller number of vision tokens. It reports that the system can recover more than 96% of text at 10x compression, while OCR accuracy falls to 60% at nearly 20x compression. Its encoder uses local window attention, a 16x convolutional compressor, and global visual processing to handle high-resolution documents more efficiently. The video also describes generating training data at more than 200,000 pages per day on a single A100 GPU, while noting that OCR results alone do not prove true text compression and the model may rely on language priors to guess missing text.

Key points

The central idea is to make AI memory more efficient through artificial forgetting: retaining a compressed representation rather than every detail.
DeepSeek OCR renders information, including text, as an image and compresses it into a smaller set of learned, token-based visual representations.
At 10x compression, the video says the system can recover more than 96% of the text.
At nearly 20x compression, OCR accuracy falls to 60%.
The approach is more than applying JPEG to an AI system: it compresses visual representations of memories.
The encoder combines local window attention, a 16x convolutional compressor, and global visual processing.
The system is described as capable of generating LLM and VLM training data at more than 200,000 pages per day on one A100 GPU.
A limitation is that OCR accuracy does not by itself establish true text compression; under heavy compression, language priors may help the model guess missing text.

Chapters

Start	Chapter	Summary
00:00	The idea of artificial forgetting	The video introduces the cost of retaining tokens and asks whether AI could benefit from forgetting details as humans do.
01:03	Compression and text recovery	DeepSeek OCR compresses documents represented as images, with the video reporting high text recovery at 10x compression and reduced accuracy near 20x.
02:08	How DeepSeek OCR works	The video distinguishes the system from JPEG and explains its encoder’s visual processing and compression components.
03:41	Training data and limitations	The video describes the system’s potential to generate training data at scale and cautions that language priors may help it guess text lost through compression.
04:43	Closing and sponsor	The speaker concludes with praise for open research and introduces Lambda’s GPU services.

Quotes

What if the best way for an AI to remember a book is to stop reading it? — 00:00

We can recover more than 96% of the text. — 01:35

JPEG compresses pixels. This compresses memories. — 02:40

It skims, it squeezes, and only then it reads. — 03:11

it may not truly see the missing text anymore, but it also guesses the most likely words from the context — 04:13

Concepts and entities

DeepSeek OCR, DeepSeek, OCR, JPEG, A100, LLM, VLM, Two Minute Papers, Karolis Jonaityte, Lambda

Full transcript (original language)

00:00 What if the best way for an AI to remember a book is to stop reading it? You see, every word an AI reads becomes tokens, and tokens are expensive. The longer the book, the more the AI chokes on it. But, we humans do something strange. We don’t remember everything perfectly. Some details fade. The older grandpa is, the better he was at school. And scientists at Deep Seek asked the question, “What if we make an AI do

00:30 that, too?” Asking artificial intelligence to do artificial forgetting. But, what a crazy proposition. Why would you do that? How does that even make sense? Well, let’s see together. As we keep talking to an AI system, the longer we talk, the more documents we attach, the more expensive it is to keep running because it has to remember everything. Now, Deep Seek OCR says not to worry. Let’s take this big soup of information, even text, and render it as an image,

01:03 and then compress that image, a bit like what JPEG compression does to images. But, this is learned and token-based. And now, this data is drastically smaller. But, this still doesn’t make too much sense. So, how is that good? I mean, we lose a lot of information, right? Well, hold on to your papers, fellow scholars, and let’s see how much information we lose at 10x compression level that is throwing away almost everything and using 90% less space.

01:35 That is a lot of compression. And what can that be? We can recover more than 96% of the text. That is insane. So, why not push it a bit further? Well, as we go up to nearly 20x compression, OCR accuracy falls to 60%. Wow. I expected much, much lower. Okay, now important question. Is this really new? I mean, at first I was a bit confused by this work because this is an

02:08 unassuming paper about OCR. Yes, this does OCR. So, you give it a document, maybe even a handwritten note, and it gives you the text back. Formally, it takes a small amount of compressed vision tokens and turns them into a much larger number of text tokens. But, is this just JPEG? I mean, that technology is more than 30 years old. The idea is good, but is this just JPEG plugged into an AI system? Well, not quite. There is tons of new stuff here. Dear fellow

02:40 scholars, this is Two Minute Papers with Dr. Karolis Jonaityte. You see, JPEG compresses pixels. This compresses memories. First, these memories are turned into a picture, and then we compress that picture. But, we are not done yet. No, no, no. Not even close. The key engineering trick is that they created an encoder that is built to survive very high-resolution documents without exploding in cost. To do that, one, it

03:11 uses local window attention first. Two, a 16x convolutional compressor. And three, it does global visual processing. Okay, and what does that mean? Well, in other words, it does not stare at every pixel with full attention. No, it skims, it squeezes, and only then it reads. And here is something brilliant. Once you can do this cheaply enough, suddenly this is not just about remembering old

03:41 data. Wait, does that mean that it can not only understand documents, but also help generate new data? Yes. The paper says it can generate training data for LLMs and VLMs at more than 200,000 pages per day on a single A100 GPU in sane speeds. Now, we are fellow scholars here, so we look at limitations, too. The paper itself says OCR alone is not enough to prove true text compression,

04:13 and a later stress test suggests the model may rely heavily on language priors, not just visual memory. In other words, when the image becomes too compressed, it may not truly see the missing text anymore, but it also guesses the most likely words from the context, kind of like auto complete, but better. Overall, this is absolutely fantastic. This is not only an open system that is free for all of us, but it also creates

04:43 training data to help other open models come into being. That is absolutely amazing. What a time to be alive. And don’t forget, every research paper pushes humanity forward. But, the future belongs to the ones who understand it. Perhaps, subscribe and hit the bell if you feel that you enjoyed this. I use Lambda to reproduce AI research papers often in minutes. It’s also great to train your own models or fine-tune an existing one. Run inference or text to

05:14 image or video, easy peasy. Running a deep sea chatbot or agent, super fast, super reliable. Lambda gives you powerful Nvidia GPUs to run your own experiments. I test ideas from the papers I cover, and moments later, results. Love it. Seriously, try it out now at lambda.ai/papers.