Skip to content
Melo Podcasts Home
CategoriesLanguagesFollowing

Episode notes

How humans and AI models both share the weakness of deterioration without novel inputs. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

Transcript

Read the transcript · about 640 words, follows along as you listen

S1:I've had several thoughts on the Carpathian Dwarkesh conversation that took place in late October of 25, but the one that keeps haunting me is something Karpathy just kind of casually mentioned before moving on to another topic. I think it might be the biggest idea in the whole conversation. He was talking about human model similarities and he says humans collapse during the course of their lives. Children have an overfit, yet they will say stuff that will shock you because they're not yet collapsed. But we adults, we end up revisiting the same thoughts. We end up saying more and more the same stuff. The learning rate goes down, the collapse continues to get worse, and then everything deteriorates.

S1:End quote. Since my 20s, I've been terrified of this happening to me. It pierces my soul whenever my partner says things like, I knew you were going to say that. Ouch. Predictable humor or wit isn't another example. How many older people do you know who tell the same stories and jokes over and over? They watch the same shows. They listen to the same five bands, and then eventually, like 2 or 1, their aperture slowly shrinks until they die. Luckily, Karpathy gives a solution right after we have to find sources of entropy. When we were kids, everything was entropy because everything was new. So we were constantly changing our preferences, our behaviors, our language and everything. It made us fresh, unpredictable, which is highly related to the concept I'm obsessed with from Shannon's theory of information, which in his model defines information as the part of transmission that isn't repeated or noise.

S1:I think about this constantly when I'm giving talks or participating in panels or whatever, or when I'm watching someone else do so. The main thing I'm asking myself, especially for my own content, is how much of this is new? How often will I'm presenting this? Will the viewer be pleasantly surprised? If the answer is not very often I redo it or I start over. I'm actively doing a bunch of stuff in addition to pathological reading to maximize entropy in my life. I'm reading a lot of old books on writing, like rhetorical figures and stuff like that, to try to get fresh phrases into my mind. I regularly reread and listen to Christopher Hitchens books and debates.

S1:Just having exposure to that level of non cliché language. And I'm currently building in cloud code a skill called Increase Entropy that incorporates all of this old and fresh language like a particle accelerator. So I can point it at a thought or a piece of content and basically come up with novel ways of saying the same thing. So I give it the way that I would say it in a kind of like just breaks me out of my mold. I even went so far in 2024 to create an AI prompt in fabric that would rate talks, blogs, panels, or whatever for wows per minute, meaning how often a given piece of content surprised the audience. I mean, this was a problem before AI, and now many are delegating even more and more of their thinking to a system that learns by crunching mediocrity from the internet. I can see things getting way worse in this respect. I guess it's somewhat comforting that this happens to both AI models and to people. It makes the whole thing more human somehow.

S1:And hearing Karpathy say it so plainly was jarring to me in a pleasant way. At least for us humans. The solution seems something like recognize that this is a problem that starts for everyone in there, probably like mid to late 20s, and constantly seek and consume sources of novelty and freshness to maintain young mind.

Transcript supplied by the publisher with the episode.

Unsupervised Learning

by Daniel Miessler · English · Tech & Science

Unsupervised Learning is about ideas and trends in Cybersecurity, National Security, AI, Technology, and Culture—and how best to upgrade ourselves to be ready for what's coming.

More from Unsupervised Learning

  1. 28 Feb 2026 · 1 hr 24 min

    The Great Transition

    There are a bunch of different transitions happening right now—all at the same time, all (I think) heading in the same direction. Here is a long-form exploration of the various pieces. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  2. 30 Jan 2026 · 25 min

    Starting 2026

    A welcome back and early entry into 2026. Sponsored by: Knocknoc! Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  3. 22 Nov 2025 · 7 min

    Judge AI based on Output, Not Mechanism

    How we can use an output-based system to judge whether or not different kinds of technology achieve understanding or intelligence. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  4. 20 Oct 2025 · 10 min

    Why I Think Karpathy is Wrong on the AGI Timeline

    Karpathy is confusing LLM limitations with AI system limitations, and that makes all the difference. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  5. 15 Oct 2025 · 4 min

    Novelty Exploration vs. Pattern Exploitation

    How going from exploration to exploitation can help you as both a consumer and creator of everything. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  6. 14 Oct 2025 · 7 min

    Magnifying Time

    Some thoughts on how novelty and attention magnify the time that we have. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  7. 17 Sep 2026 · 24 min

    Evolution, Lit RPG and AI Harnesses

    I went on a walk and ended up connecting three things I've been thinking about: evolution, LitRPG, and personal AI harnesses. Why listening to The Primal Hunter keeps getting me excited to work on my own goals, and how a personal AI harness is your own LitRPG system. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  8. 31 Aug 2026 · 21 min

    The Missing Piece Is Ideal State

    📝 Theo put out a video saying you should turn off Claude Code's memory, and it triggered me enough to record a response. My argument: the real problem with all AI is the gap between what's in your head and what the AI knows you want. I walk through how LifeOS closes that gap with the Ideal State Artifact, one document that works as both the build spec and the testing harness, and show the skill system and blogging pipeline that run on it. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  9. 14 Aug 2026 · 4 min

    The Different Games OpenAI and Anthropic Are Playing

    Anthropic is going after the enterprise operations layer; OpenAI wants to be everyone's personal agent. Both are multi-trillion-dollar markets, and they barely overlap. Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

  10. 9 Jul 2026 · 1 hr 15 min

    Debating the Morality of Dario Amodei

    A debate with Zack Korman on the morality of Dario Amodei specifically, and Anthropic in general. Zack's version of the video: https://youtu.be/mQ9-YoE9ykE Become a Member: https://danielmiessler.com/upgrade See omnystudio.com/listener for privacy information.

Every episode of Unsupervised Learning →

Take it with you

The Melo app keeps playing with the screen off, works in the car and on your watch, wakes you to your station, and browses the whole catalogue offline. Free, no ads, no account.

Get it on Google Play