Eskildsen: AI models cannot compress all human knowledge into model weights
Simon Eskildsen · He built a new database in his bedroom—now he powers Cursor, Notion and Anthropic. | Simon Eskild... · PMF Show · Oct 30, 2025 · at 16:51
Turbopuffer founder Simon Eskildsen argues why AI applications inherently require external search and database tools rather than relying solely on model training weights.
0:00 / 0:20exact quote · 21.0s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“When you train an AI model, you're trying to basically compress all of human knowledge into a few terabytes. That's impossible. You can't do that. At least not by any computer science or any science really known. It does not make sense that you can compress all of human knowledge into a few terabytes. What you can compress into a few terabytes is a way to reason About the world and a mental model of the world.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Simon Eskildsen
Eskildsen: IOI computer science competition contributes more to GDP than the Olympics
“The problem is that no one's heard about it, but it contributes more to GDP than the real Olympics.”
Eskildsen: Turbopuffer is 100x cheaper than incumbent vector databases
“The innovation of Turbo Puffer is that it is almost a hundred times cheaper than the incumbents at the time to store that coordinate system.”
Eskildsen: Sacrificing write latency to S3 reduces search database costs 100x
“So, okay, you might give up three orders of magnitude in write latency, but actually you don't have to give anything else up. And then you have this true two orders of magnitude decrease in the base cost of the system. That's a great compromise if you're build…”
Eskildsen: Turbopuffer told seed investors it would shut down without 2024 PMF
“If Justin and I could not find product market fit by the end of 2024, we were just going to close up shop. Justin and I were very clear about this. We even told the investors Who were invited for that round in 2024 that we were going to close up shop by the en…”
Eskildsen: Readwise vector recommendations would have cost $30k/month on incumbent databases
“They have hundreds of millions of articles, and it would have cost about 30 grand a month to put in the production on one of the incumbents at the time. And that was just way too much, right?”
Eskildsen: Generational database companies require fundamentally new workloads
“If you want to build a generational database company, you need two and a half things. The first thing that you need is that you need a new workload, because if you want to make it as a really large billion dollar plus database company, you need a reason for mo…”