Thanks to the insane deepseek models We should have a framework for: 1. Use local llm if query simple. 2. If query not simple ask big daddy llm via api.

I think one of the easy to miss deep insights in DeepSeek training algo is that when reasoning goes up -> verbal goes down. when verbal goes up -> reasoning goes down. this sounds like left brain - right brain situation

I think an idea such as: Don’t nerf either reasoning or verbal to make an ok model, but rather make them work together, would be cool reasoning model and verbal model interact among themselves in a high bandwidth layer before answering aka corpus callosum for llms

Whenever I see all the candidates highlighting all the certificates they have in their CV, it's kind of a no no from me because that signals to me that this person has optimized for acquiring certificates over actually doing things. Do real things bro

A true AGI may choose to not respond to API calls.

if you’re playing by the rules you’re playing the wrong game.

there’s roughly two kinds of work 1. Doing what you have to do (because someone told you to, etc) 2. Doing what you want AI will delete work type 1.

A̶r̶t̶i̶f̶i̶c̶i̶a̶l̶ Fossilised Intelligence

Long GPT wrappers

You gotta go to the gym like it’s your last day there… because it is. Tomorrow it isn’t going to be the same person.

The truly living man, lives as if he is continually dying.

Your rate of progress is directly proportional to your pain threshold

The portal of change is unchanging regularity

Wrong comparison, this compares transport of materials to generation of energy. That’s like saying you’re delivering X amount of joules of energy per minute when you transport nuclear fuel with a hose

Accepting applications for Job: Machine Whisperer

Attention is more like — permission to edit brain

progressive overload is magic

Create endlessly!

It is all about feeling