Compression is Caring

Ask someone to name a smart person, then ask why. The answers rarely involve how much that person knows. They involve how well that person explains. Feynman said that if he couldn’t reduce an idea to the freshman level, he didn’t really understand it. Aristotle put it more bluntly: it is the audience that determines the speech’s end and object.

That instinct turns out to be more than a teaching philosophy. Ilya Sutskever has argued that training a large language model approximates optimal compression. The Hutter Prize pays out half a million euros for compressing Wikipedia because compression is treated as a proxy for AI. Recent work shows compression ability tracks intelligence benchmarks linearly. If intelligence is compression, and machines now compress better than we do, what exactly is left for us?

The answer is the one thing the model cannot see: who is receiving the output. Search made every fact available. LLMs turned facts into answers. Only you know your reader, your team, your customer, and the specific thing they need right now.

This talk is a practical argument for a two-step habit. Decompress aggressively with AI, asking the questions you were embarrassed to ask, and then recompress deliberately for the person (or computer) in front of you. We’ll talk about why domain expertise is the durable differentiator. The model can generate the answer, but it cannot tell you which answer matters here.

Takeaways

• Why compression is a useful working definition of intelligence, and what that implies about your own value
• A two-step workflow for using generative AI: decompress to learn, recompress for your audience
• Concrete ways early-career professionals can stand out when everyone has access to the same tools
• Why investing in domain expertise beats investing in prompt tricks