Why we dream, GPT-4 playing games, and more

Jun 3, 2023

Tim Ferriss' latest podcast episode with David Eagelman is fascinating. David is a neuroscientist, author, and founder of Neosensory, which make bracelets that help with hearing loss and tinnitus, and internal test versions have even offered a sense of the infrared spectrum.

The part I found most interesting is David's theory of why animals dream. Brains are so quick to adapt that, left to their own devices, they would start re-purposing the visual cortex during sleep. Dreams are spawned by periodic bursts of random noise into the visual cortex to keep it active. The theory is supported by a very tight correlation between how much neoroplasticity different species’ brains have and how much REM sleep they get. The only highly plastic species with little REM sleep are elephants, which only sleep 1-2 hours a night and have excellent night vision.

Elicit is an interesting-looking tool that uses LLMs to help with research workflows, finding papers that match your question, extracting the answers, summarising relevant takeaways, etc.

Tim urban posted a great metaphor of bias as two side-by-side trains. A stationary train is a person without bias. When both are unbiased, they seem unbiased to each other. When both are biased, they are both moving and look motionless relative to each other, and thus each considers the other unbiased. And when one is unbiased and the other biased, both see a train moving past them: to the biased person, the unbiased train seems biased, but in a direction (of opinion) opposite theirs.

Optimising for engagement and addictiveness comes to LLMs, increasing time spent chatting by 70%

GPT-4 can play full games of chess at a decent level, if it's prompted in a very specific way. This suggests the base model has learned many systems/skills that we may not know about because they don't get "activated" when we use them normally.

GPT-4 can learn to play Minecraft by writing code and improving it with feedback, all in text.