Must read if you’re interested in AI and its implications; Tyler’s commentary on the recent explosion of AI into the popular consciousness (driven in large part by ChatGPT) has been, in my view, the most realistic+pragmatic:
“I don’t have a prediction for the rate of improvement, but most analogies from the normal economy do not apply. Cars get better by some modest amount each year, as do most other things I buy or use. LLMs, in contrast, can make leaps.”
“I’ve started dividing the people I know into three camps: those who are not yet aware of LLMs; those who complain about their current LLMs; and those who have some inkling of the startling future before us. The intriguing thing about LLMs is that they do not follow smooth, continuous rules of development. Rather they are like a larva due to sprout into a butterfly.”
This podcast made me feel very stupid, and very inspired.
—
Turing test – in 1950s, Turing didn’t mean it to be a rigorous formal test, more of a philosophy experiment, didn’t specify things like parameters of test, how long test should last, etc
More modalities than just language to express intelligence eg physical movement
Played chess at 4, earnings from winning a chess competition let him buy a computer
Bought programming books, started making games, felt they were a magical extension of your mind
“AI is ultimate expression of what a machine can do or learn”
At 12yo he got to chess masters level, the process of learning chess makes him think a lot about thinking and about brains
“Chess computer handbook” by David Levy – explained how chess programs were made
First AI program he built was on his Amiga, programmed Othello
Wrote game called “Theme Park” with a core AI, sandbox game, reacted to players, every game was unique
He designed and wrote AI for games in 90s – at the time, game industry was cutting edge of tech (GPUs for game graphics, AI, John Carmack)
“Black and white” game – train a pet, and depending on how you train it, it would be more or less kind to others, powerful example of reinforcement learning
DeepMind – core part of strategy from start was to use games to test how well AI is doing, if the ideas are working
Eg, Go – clear rules and win conditions, humans have played for thousands of years, easy to test how good is your system vs human players
Part of why their AI has progressed so quickly – by developing against games
“Chess is drosophila of intelligence” – Gary Kasparov
Many AI researchers have all wrote chess AI programs
DeepBlue beating Kasparov was a huge moment – he was in college at time – came away more impressed with Kasparov’s mind than with DeepBlue (because Kasparov could play almost at the AI level, but could also do all these other things as a human, while DeepBlue at that time couldn’t even play tic-tac-toe)
What makes chess compelling as a game? Creative tension between bishop and knight – leads to a lot of dynamism
Chess has evolved to balance those two more or less equally (worth 3 points each)
Balanced by humanity over hundreds of years
Different levels of creativity
1. Lowest level is interpolation – averaging everything you see (eg, “an average looking cat”)
2. Next is extrapolation – AI coming up with a new move in Go that no one’s seen
3. Out of the box innovation – coming up with a new game entirely – AI nowhere close to this yet
Currently AI can do 1 and 2 but not 3
For 3, if you were to instruct an AI to create a game, you’d say “come up with a game that takes 5 minutes to learn, but lifetimes to master, aesthetically beautiful, and can be completed in 3-4 hours”
We can’t abstract high level notions like that to AIs (yet)
AI could be used to make current games better by taking game system, playing millions of times, and then improving the balance of rules and parameters – give it a base set + Monte Carlo tree search – takes humans many years and thousands of testers to do it
His first big game was theme park, amusement park – then whole cities – and Will Wright’s made SimEarth simulating the whole earth
“Simulation theory” – doesn’t believe it, in sense that we’re in a computer simulation / game, but does think best way to understand physics and universe is from computation perspective – information as fundamental unit of reality instead of energy or matter
Understanding physics as information theory could be valuable
Roger Penrose – Emperor’s New Mind – he believes we need quantum, something more, to explain consciousness in the brain
Most neuroscientists / mainstream biologists haven’t found any evidence of this
While continually classic Turing machines keep improving – and DeepMind / Demis work is champion of this
Thinks universal Turing machines can eventually mimic human brain without Penrose need for something more
Something profoundly beautiful and amazing about our brains – incredibly efficient machines, in awe of it
Building AI and comparing to human mind will help us unlock what’s truly unique about our minds – consciousness, dreaming, creativity
Philosophy of mind – there haven’t been the tools, but today we increasingly have them
Lex – Universe built human mind which built computers to help us understand universe and human minds
Protein folding – AlphaFold 2 solved it
Proteins = essential to all life – workhorses of biology, amazing bio-nano machines, specified by genetic sequence, in the body they fold into 3D structure (like string of beads folded into a ball)
The 3D structure determines what it does – and drugs must understand this to interact with it
Structure maps to function, and is specified by amino acid sequence
Unique mapping for every protein – but it’s not obvious – and almost infinite possibilities
Can you by studying the sequence, predict the 3D structure?
Takes 1 PhD student an entire PhD to predict one protein But AlphaFold 2 can do it in seconds now – over Christmas can do it over entire human proteome space (!!)
Biologists can now lookup protein 3D structure in a google search
AlphaFold was most complex and meaningful system they’ve built so far
Started on games (AlphaGo, AlphaZero), to bootstrap general learning systems
His passion is scientific challenges – AlphaFold is first proof point 30 component algorithms needed to crack protein folding
About 150K protein structures had been found – that was their training set
Would put some of AF’s best predictions back into training set to accelerate training
AF2 was truly end to end – from amino acid sequence directly to 3D structure, without needing all the intermediary steps – system is better at learning the constraints on its own instead of guiding it
AlphaGo – learning system but trained only to learn Go
AlphaZero – removed need to learn from human games – just play with itself MuZero – didn’t even need to give rules, just let it learn on its own
Started DeepMind in 2010 – back then no one was talking about AI, people mostly thought it doesn’t work (even at MIT)
If all professors tell you you’re mad, at least you know you’re on a unique track Founding tenets / trends
-Algorithmic advances (reinforcement learning)
-Understanding about human brain (architectures, algos) improving
-Compute and GPUs improving
-Mathematical and theoretical definitions of intelligence
Early days – ideas were most important – deep reinforcement learning, transformers, scaling those up As we get closer to AGI, engineering and data become more important
**For large models – scale is clearly necessary but perhaps not sufficient
DeepMind – purposely built multi-disciplinary organization – neuroscience + machine learning + mathematics + gaming, and now philosophers and ethicists too
“A new type of Bell Labs” DeepMind itself is a learning machine building a learning machine
Top things to apply AI – biology and curing diseases (AlphaFold), but it’s just beginning Eventually simulate a virtual cell (maybe in 10 years) – “that’s my dream”
Drugs take 10 years – target to drug candidate – maybe it can be shortened to 1 year with this, AlphaFold as first proof point
Math is perfect description language for physics AI as perfect description language for biology (!)
Open-sourced AlphaFold (including data) – max benefit to humanity – so many downstream applications, better to accelerate research and discovery, used by 500K researchers (almost every biologist in the world!), amazing fundamental research, almost every pharma company is using it, “gateway drug to biology”
Also open-sourced MuJoCo – purchased it explicitly to open source it
One day an AI system could come up with something like General Relativity (!)
Big new breakthroughs will come at intersection of different subject areas (DeepMind = neuroscience + AI engineering)
We just don’t understand what it’d be like to hold the entire internet in your head (imagine reading all of Wikipedia, but much much greater) – no one knows what will result
Nuclear fusion – believe AI can help
In any new field, talk to domain experts for collaboration
What are all the bottleneck problems? Think from 1st principles
Which AI methods can help
Problem of plasma control is great example – plasma is unstable (mini-sun in a reactor), want to predict what plasma will do next, to best model and control it
They’ve largely solved it with AI, and now looking for other problems in fusion
Simulating properties of electrons – if you do it, you can describe how elements and materials work (fundamental to materials science)
Would like to simulate large materials – approximate Schrodinger’s equation
His ultimate aim for AI – to build a tool to help us understand the universe – to test the limits of physics
A true scientist – the more you find out, the more you realize you don’t know Time, consciousness, gravity, life – fundamental things of nature – we don’t really know what they are
We treat them as fact and box them off – but there’s a lot of uncertainty about what it is
Use of AI is to accelerate science to the maximum – imagine a tree of all knowledge – we’ve barely scratched surface, and AI will turbocharge all of it – understanding and finding patterns, and then building tools
If you’re good at chess, you still can’t come up with a move like Garry Kasparov, but he can explain the move to you – potentially AI systems could understand things we could never by ourselves, and then explain it and make it useful for us
We’re already symbiotic with our phones and computers, Neuralink, and could augment / integrate with these AI
His current feeling is we are alone (no aliens)
We could easily be a million years ahead or behind in our evolution, eg, if meteor that destroyed dinosaurs came earlier or later – and in a few hundred years, imagine where we’ll be, AI, space traveling – we’ll be spreading across the stars; will only take ~1M years for Von Neumann systems to populate across the galaxy with that tech
We should have heard a cacophony of voices – but we haven’t heard anything “We’ve searched enough – it should be everywhere”
If we’re alone, somewhat comforting re: Great Filter (maybe we’ve passed it)
Wouldn’t be surprised if we found single cell alien life – but multi-cellular seems incredibly hard
Another large leap is conscious intelligence
General intelligence is costly to begin with – 20% of body’s energy – a game of professional chess is same as F1 racer
Hard to justify evolutionarily – which is why it’s only been done once (on Earth)
AI systems – easy to craft specific solutions, but hard to do generally – at first general systems are way worse
Do AGI systems need consciousness? Consciousness and intelligence are double dissociable – can have one without the other in both ways
Eg, Dogs have consciousness and self-aware but not very intelligent, most animals are pretty specialized
Eg, some AI are amazingly smart at playing certain games or doing certain tasks, but don’t seem conscious
May be our responsibility to build systems that are not sentient
None of our systems today have one iota of consciousness or sentience – way too premature
Re: Google engineer who believed their language system was sentient – Demis believes it’s more a projection of our own minds, our desire to construct narrative and agency even within inanimate systems
Eliza AI chat bots in 1960s – already fooled some people
Neuroscience – certain pre-reqs may be required for consciousness, like self-awareness, coherent preferences over time
Turing test is important, but there’s second that differ in machines: we’re not running on same substrate (humans are carbon based squishy life forms)
Language models – we don’t understand them well enough yet to deploy them at-scale Should AI be required to announce that it is AI?
re: AI ethics, important to look at theology, philosophy, arts & humanities
Heading into an area of radical abundance and knowledge breakthroughs if we do it right – but also huge risks
Need careful controlled testing instead of just releasing into the wild, the harms could be fast and huge
Better to first build these AI systems as tools – carefully experiment and understand – before we really focus on sentience
How to prevent being corrupted by this AI power:
-Important to remain grounded and humble
-Being multi-disciplinary keeps you humble – because always better experts
-Have good colleagues who are also grounded
AI can learn for itself most of its knowledge, but will have residue of culture / values from who builds it
Globally we can’t seem to cooperate well eg, climate change
Need to remove scarcity to help promote world peace – radical abundance
AI should belong to the world and humanity, everyone should have a say
Advice for young
-what are your true passions? explore as many things as possible; find connections between things
-understand yourself – how do you deal with pressure, hone your uniqueness and skills
Perfect day in Demis’ life, habits
-10 years ago: whole day of research + programming, reading lots of papers, reading sci-fi at night or playing games
-today: very structured, complete night-owl, 11-7pm work (back to back meetings, meet as many people as possible), go home and spend time with family, 10pm-4am do individual work, long stretches of thinking and planning and some email
-quiet hours of morning – love that time (1-3am), inspiring music, think deep thoughts, read philosophy books, do all his creative thinking, get into flow, sometimes will go to 6am next day, and pay for it the next day (but it’s worth it)
Always been a generalist – too many interesting things to spend time on just one
Lex: Why are we here?
Demis: To gain knowledge and understand the universe, understand world around us, humanity, and all these things flow from it: compassion, self-knowledge
Feel like universe is almost structured to let itself be understood and learned – why are computers even possible?
If Demis could ask one question of true AGI: “What’s true nature of reality?”
Answer could be a more fundamental explanation of physics, and how to prove them out
A deeper, simpler explanation of things – leading to consciousness, dreaming, life, gravity