ChatGPT is a Blurry JPEG and We Need That

Blurry JPEG is a great piece by science fiction legend Ted Chang. He compares LLMs like ChatGPT to lossy compression, blurry JPEGs. Fun analogy. His take is that LLMs only give you a superficial understanding of a topic and can ultimately get in the way of real knowledge.

I disagree with the conclusion of this article. Even blurry images have value for learning. Sometimes we need them. Pulling data from the real world and processing it is slow, so sometimes we need a less detailed but more efficient representation to understand fast and move fast.

Squint a bit and learning is ultimately fetching data from the world, same as a computer fetching from storage. So how fast you can fetch from a storage medium matters a lot. I once ran a thought experiment with a CPU calendar. Say one CPU cycle takes one second. (That's 4 trillion times slower.) How long do everyday tasks take?

On this calendar, reading 1MB from disk takes a full 38 days. And 1MB of text is a huge amount of information. A person trying to study and comprehend it would probably need about that long.

Disk seek comes to 92 days. Eerily close to what real-world researchers may spend investigating, analyzing, and understanding complex subjects.

Now you're the CPU and you need a 130.8MB image, the released James Webb Space Telescope photo. Bringing in an image that large would take almost 14 CPU years to process. The compressed version under 2.5MB? Sure, a "blurry JPEG." Most people would barely notice.

Same with ChatGPT. The "blurry image" loses quality due to compression, but it's much smaller and tens of thousands of times faster to decode. The quality of the information is often compromised, but as long as we're careful about the accuracy, it can revolutionize how we learn and process information.

So we must embrace blurry JPEGs and use them wherever possible. That's what internet companies do to compress information and make it accessible. And just as we strive to develop better compression algorithms, we must also push LLMs like ChatGPT to make knowledge compression more efficient and affordable.

Learning still matters, though. A lot of people ask why we need knowledge if we have LLMs. We need education because storing information in our brains, like registers and caches, lets us get at it in CPU hours, not months or years. Leaning on LLMs may be faster than traditional learning, but it's still orders of magnitude slower than memorizing and processing information in our own minds. LLMs can be useful tools for building knowledge, but they should help our own capacity for learning and understanding, not replace it.