AlphaFold

... or 100 years of scientific development in 10 years.

Even if you look at just Go- um, AlphaFold, which is that program that, that Google DeepMind made, it automated the research around folding proteins, and I think it did what-- I think it was like 100,000 PhDs worth of protein-folding research-

Hmm

...

The extremes in these charts don't necessarily combine freely, but neither does high contagiousness guarantee low lethality.

What could a superintelligent AI create if it were trying on purpose? AI has already made major contributions to biology.

In 2024, AlphaFold2's developers shared the Nobel Prize in Chemistry for predicting proteins' three-dimensional structures from their amino acid sequences, a problem scientists had pursued for over 50 years.

A protein's shape helps determine what it does and the ability to predict that paves the way for effective alterations.

AI systems have also designed entirely new proteins that scientists manufactured and confirmed could perform intended functions.

But AI is also doing things that look almost miraculous.

DeepMind's AlphaFold won a Nobel Prize for solving a 50-year-old problem in biology and is now being used by over 3 million researchers worldwide.

And an AI-designed cancer drug is now entering human trials.

So what is the real threat of AI? Are we living through a genuine technological revolution that could transform humanity? Or is all this apocalypse talk just very good marketing for a bubble that has grown from $700 billion to $1.4 trillion in the space of a year and is on track to hit $2.4 trillion next? Well, our guest has spent more than two decades examining the intersection of technology and corporate power.

... mutations for that patient almost 30% of the time, which is higher than a previously existing ranking method.

Well, it sounds like an interesting case study, but I guess we'll have to wait and see, uh, whether this will be another, like, AlphaFold moment, which was another of DeepMind's tools that really transformed biology.

Yeah, it's funny you should say that.

I've got a direct quote here from this, uh, article from a molecular biologist who said, "This isn't an AlphaFold moment." [laughs]

[laughs]

Um, and they've said, uh, "The predictive performances of genomic models like AlphaGenome are still far from brilliant, and these models should not yet be used alone to make clinical decisions." So yeah, maybe not quite at AlphaFold level yet, but extremely useful.

Well, a sensible caveat to end that story on then.

[laughs]

We didn't immediately see, uh, a, a line from Move 37 to discovering a new material, for example.

And even if you think within a single organization, the line between Move 37 and, say, AlphaFold, wasn't a particularly direct line.

It's not like the AlphaGo agent, or indeed the AlphaGo training techniques really informed AlphaFold, um, i-in a very direct way.

But in principle, if, if you had had a generalizable, uh, discovery engine, then it could make a discovery about the weather and then figure out, oh, there's some part of that discovery, perhaps it's the architecture of the neural network that was used in those to make that model.

I wonder whether that applies to protein design.

Uh, and insofar as we have trained these models on data, the form of the models is, uh, reflects the data, um, and reflects reality in some ways.

It's teaching us about reality.

Um, and in fact, yesterday I was listening to your interview with John Jumper, and I thought he put this, um, uh, very succinctly and beautifully when he was talking about the advances of AlphaFold2 over AlphaFold1, where, as if I remember rightly, they used no more data than AlphaFold1, but in some sense they were just more in tune with reality in AlphaFold2.

Let's carefully break down how unprecedented this is, because the distinction is vital.

Previously, the biggest AI breakthrough in the realm of biology was Google DeepMind's AlphaFold.

Right.

But AlphaFold is fundamentally a predictive model.

It looks at an existing known sequence of amino acids and uses AI to predict how they will fold into a three-dimensional protein structure.

It is incredibly useful for drug discovery, but it is ultimately analyzing what already exists in nature.

I think this is a recipe that we're going to see over and over again.

So if you remember the history of AlphaFold, where it was originally a model, and then it was another better model, and then it was another better model, AlphaFold 3, a Nobel Prize winning model.

And then it was a database, the AlphaFold protein structure database, where you use the model to pre-compute the answers to basically all of structural biology, or at least the the proteomic portion of single molecule structural biology and you flatten an entire field bulk solved bulk solving a field seems to want to become a database of all the pre-computed answers to all the questions that can be asked in that field we saw this with alpha fold now we're seeing it happen with variant effect prediction taking every possible you know three times 3.1 billion base pairs equals approximately 9 billion possible single nucleotide variations.

This is the bulk solution for now all of variant effect prediction.

I think this is a formula.