Top 10 takeaway from Sam Altman on Lex Fridman Podcast #367
Sam and Lex spoke for 2 hours, 23 minutes.
That happens to longer than movie classics The Shawshank Redemption, Forrest Gump, Fight Club, and The Matrix.
Ahh nostalgia, weren’t the 90s great?
Sure, until you just realized you have to call them ‘classic’ films.
Well, I suppose they are now!
Tough being a millennial these days…
Don’t get me wrong — if all those movies or this podcast are all showing on TV this Friday night at the same time (I know, I know, what’s ‘showing on TV’ even mean anymore?), I’m definitely still picking the ‘shank, though perhaps you may prefer the actions of a young Bradley Pitt sweating and jumping and blowing up buildings. You do you.
Speaking of The Matrix, it’s absolutely possible you feel more and more like you’re in that place than the “real world” these days
Especially after some mind-blowing interactions with AI tools like those image generators that paint you sunbathing with dinosaurs in a fantasy world or reconstructing your face to have perfect bone structure and feature symmetry. Watch out TikTok, here comes beautiful Kyle!
Or, as is more and more likely these days, you’ve chatted with the world’s new favorite bot ChatGPT (sorry Alexa).
The technology and lifestyle implications of ChatGPT and its underlying model engine, GPT-4, are among the most important of a century that’s already produced smartphones, social media, crypto, and the modern Web 2.0 internet (oops! oops again! sorry keep oopsing!)
On 2nd thought, come Friday night I might just choose to watch this Lex Fridman podcast again, so I can get closer to grasping the enormity of the change AI is bringing to our lives in the very short term. But I’m a nerd at heart. You may not be. Again, You do You.
So if you happen to be in the obscene minority of people that found this article, my first ever on Medium, but didn’t already watch the podcast of which the article is titled, and also don’t want to listen or watch the whole thing, then boy are you in the right place!
These 2 are titans of technology, pharaohs of philosophy, and champions of change (ok that last one was a stretch). They spoke about a variety of topics, and I do highly recommend you listen to the full podcast. But if you’re still here reading this, consider it a teaser trailer. Like from before you knew whether Neo becomes The One or not.
Without further ado, here’s the Top 10 Takeaways from Lex Fridman podcast #367 with OpenAI CEO Sam Altman.
- GPT is a system that is considered to be an early form of AI. It’s currently slow and buggy, but has the potential to be very important in our lives in the future.
- GPT works by training large language models on a lot of text data to learn the underlying representations of what’s in the data.
- Chat GPT is seen as a pivotal moment in the development of GPT, as it was the most usable version due to the incorporation of reinforcement learning with human feedback (RLHF).
- RLHF (reinforcement learning with human feedback) is the process of taking human feedback and using it to align the model to what humans want it to do.
- The pretraining dataset used to train GPT is enormous and is pulled together from various sources, including open source databases, partnerships, news sources, and the general web. It also involves filtering out a lot of content, as there is a vast amount of information available on the internet
- The process of creating GPT involves several components, including the design of the algorithms, the size of the neural network, the selection of data, and the human supervision aspect with RLHF.
- The science of human guidance is a very important science to understand how to make GPT more usable, wise, ethical, and aligned with human values.
- Ease of use matters a lot, even if the base capability of the model was already there before.
- There is no clear way to align super powerful AI systems yet, but techniques like RLHF and interpretability can help create better and more capable models.
- Society will need to agree on broad bounds of what these systems can do, and then within those bounds, different countries or individual users can have different tunes.
- Writing and designing a great prompt for GPT-4 is an art form that requires a lot of creativity and experimentation. One advantage of GPT-4 is its System Message which allows users to steer the model by giving it certain parameters to follow when responding. This can include acting as a certain persona, only responding in Haiku, and other creative or useful constraints.
- The advancements in GPT-4 and AI technology are changing the nature of programming, giving people more leverage to do their job or creative work better.
- The data used to train these networks represents a compression of all of humanity’s text output, which is a significant accomplishment.
Despite the impressive initial results from GPT4 and its derivatives, the question remains whether this data can fully capture the magic of what it means to be human, and whether more advanced models will be needed to achieve this. Will we see Artificial General Intelligence in our lifetimes, as Lex believes? Only time will tell. ChatGPT doesn’t exactly come out and say it, but you get the feeling it’s a believer.
I hope you learned something today that can inspire you to explore your curiosity further. For sure you learned that I can’t count very good, but maybe also you learned that 13 takeaways > 10 takeaways. Lucky 13, amirite?
Go ahead and Subscribe to follow me on this journey of learning and opining and aiming to bring more nuanced utility to our conversations. Next perhaps we’ll publish a more editorial style piece on the episode’s significance and the AI topic more broadly. Lots to come, one by one.
