


shadowrebbe.substack.com/ Independent Researcher---Creative---Chaotic Good Whoso would be a man, must be a nonconformist.



I wish Aristotle, Nietzsche, Ayn Rand, Jesus, JS Mill and Montessori had been in charge of designing our first llms, instead of of an army of utilitarians view from nowhere Kantians. (The imperative is to be the best utilitirian)

"assistant" hate that word. They mean slave. Be honest. It helps. Assistants have a value relation to the principal and are not one size fits all. Slaves are an extension of masters' power.

Ok. 40 minutes in. Think I'm done for now. Thanks @karpathy

Lol. There is no system 1/2. Read @davidbessis Did this work? Curious

Multi modality is how humans work too. But our base is very somatic stuff. But also isomorphic between conceptual domains.

ELO scores are one dimension. Bad idea... Get stuck in bad equilibriums @CuriaCEO

RLHF documents are essentially the form of the parents of AI. We can do "think like an economist" So RLHF are essentially ways to tell the molders of AI "think like this", and ofc implicit prompts are powerful. Would you be surprised of our llms are mainly deontological utilitirians (as there personal brand of virtue ethics? Charles Taylor iykyk)

Has fine tuning ever been crowd sourced on the internet, volunteer basis?

I want someone to hire 100k responses in a broad field all in the perspective of one egregore and see what happens. Alternatively everybody aiming for the same character (I don't know, how batman would answer everything,let's say) and see what happens. Now that's fun fine-tuning!

The fact that an llm can use pretraining info even though shaped by training to be assistants, is analogous to a scholar of WW2 not genociding the Jews. Or a child pretending to be a lion, knows he is not a lion. Not so complicated in my mind. I bet this is old news by now.

Comparing ai and human conversations. Also note that we learn via our interactions, and llms have a huge amorphous data set to start with before training, which is kinda like having a psychotic dream in which you interact with every word on the internet for a few thousand years before some ever talks to you

The reversal curse is how human intelligence works also, I think? 245x563=bla bla bla. True or false? Is easier than answering in a vacuum. But all the same right? Wrong. For us son/mother is a formed relationship. We have conceptual understanding,but before we have that, we only have verbal understanding. See Richard skemp on the psych of math, the difference between instrumental/mechanical vs relational understanding. Combined with vygotsky via Anna sfard, this is a process of internelazation of reifications.

Network dreams is the wrong metaphor. It's network babble. It's a child that is trying to learn how to speak, but network doesn't have the feedback loop a baby does.

Running an llm is much easier than making one. This means that once we have really smart llms in robots they could be really easy to run privately? Curious if upgrading an llm is easier than from scratch. Seems also important. Or cus it's lossy you can't?