// everything is paraphrased from Sam’s perspective unless otherwise noted
Base model is useful, but adding RLHF – take human feedback (eg, of two outputs, which is better) – works remarkably well with remarkably little data to make model more useful
Pre training dataset – lots of open source DBs, partnerships – a lot of work is building great dataset
“We should be in awe that we got to this level” (re GPT 4)
Eval = how to measure a model after you’ve trained it
Compressing all of the web into an organized box of human knowledge
“I suspect too much processing power is using model as database” (versus as a reasoning engine)
Every time we put out new model – outside world teaches us a lot – shape technology with us
ChatGPT bias – “not something I felt proud of”
Answer will be to give users more personalized, granular control
Hope these models bring more nuance to world
Important for progress on alignment to increase faster than progress on capabilities
GPT4 = most capable and most aligned model they’ve done
RLHF is important component of alignment
Better alignment > better capabilities and vice-versa
Tuned GPT4 to follow system message (prompt) closely
There are people who spend 12 hours/day, treat it like debugging software, get a feel for model, how prompts work together
Dialogue and iterating with AI / computer as a partner tool – that’s a really big deal
Dream scenario: have a US constitutional convention for AI, agree on rules and system, democratic process, builders have this baked in, each country and user can set own rules / boundaries
Doesn’t like being scolded by a computer — “has a visceral response”
At OpenAI, we’re good at finding lots of small wins, the detail and care applied — the multiplicative impact is large
People getting caught up in parameter count race, similar to gigahertz processor race
OpenAI focuses on just doing whatever works (eg, their focus on scaling LLMs)
We need to expand on GPT paradigm to discover novel new science
If we don’t build AGI but make humans super great — still a huge win
Most programmers think GPT is amazing, makes them 10x more productive
AI can deliver extraordinary increase in quality of life
People want status, drama, people want to create, AI won’t eliminate that
Eliezer Yudkowsky’s AI criticisms – wrote a good blog post on AI alignment, despite much of writing being hard to understand / having logical flaws
Need a tight feedback loop – continue to learn from what we learn
Surprised a bit by ChatGPT reception – thought it would be, eg, 10th fastest growing software product, not 1st
Knew GPT4 would be good – remarkable that we’re even debating whether it’s AGI or not
Re: AI takeoff, believes in slow takeoff, short timelines
Lex: believes GPT4 can fake consciousness
Ilya S said if you trained a model that had no data or training examples whatsoever related to consciousness, yet it could immediately understand when a user described what consciousness felt like
Lex on Ex Machina: consciousness is when you smile for no audience, experience for its own sake
Consciousness…something very strange is going on
// Stopped taking notes ~halfway