Minus One · 3 min · May 21, 2025

Joscha Bach says AI advanced through scale, not a master algorithm

Joscha Bach · Executive director, CIMC

Joscha Bach says AI advanced by scaling compute and data, not through the master algorithm researchers expected.

In this episode

20:40 Science fiction gets mythology right more often than details5 min
13:27 Investors find more alpha by looking beyond the current AI wave3 min
1:07:19 AI may absorb emotional work pushed out of social life3 min
10:20Joscha Bach says AI advanced through scale, not a master algorithm3 min
40:14 Neal Stephenson does his best writing in the first two hours4 min
57:16 Cyan Banister looks for outlier founders to spot new themes4 min
4:20 Nuclear knowledge diffused because its discovery hit a narrow window6 min
28:08 Ken Liu distinguishes a thematic parallel from a prediction3 min

“We didn't think that the answer to AGI would be use more compute and more data.”

Joscha Bachat 12:08 · Minus One

Takeaways

He says researchers expected a master algorithm, but scaling prevailed.
He argues decades of earlier AI research were not on the critical path.
He sees technological change outpacing institutions’ ability to adapt.

About this segment

Joscha Bach says AI advanced by scaling compute and data, not through the master algorithm researchers expected. He argues much of the field’s earlier work was not on the path to today’s LLMs.

Executive director, CIMC

Developed cognitive architectures and leads CIMC's research into machine consciousness and models of the mind.

Aditya Agarwal, Host · General partner, South Park Commons
One of Facebook's first engineers, he helped build its first search engine; later CTO of Dropbox.

From “AGI, Sci-Fi, & Storytelling | Neal Stephenson, Ken Liu, Cyan Banister, & Joscha Bach”, May 21, 2025. Watch the full episode

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

10:23
Josiah, continuing with this nuclear analogy, you've been writing about cognitive architectures and there's all these terms of art, attention, transformer, retrieval, agentic behavior. Would you mind giving us a brief overview, a status update in the project of superhuman intelligence and what it looks like? Yeah, are we going to be disappointed? Do you agree with that? I don't think that we will be disappointed.
10:47
My perspective on progress and so on has, of course, changed. How many of you have read Snow Crash? Okay, Diamond Age? Anything? Okay, so we have an idea of a baseline, right? When we read Snow Crash for the first time, it read like a dystopia. At least I think it did for many of us. And now when we reread it, it's not quite so much a dystopia, I think. and in part because it's so there is so much vitality in it right this is uh it seems to me that we since then moved into a world that became slower as molasses that became more and more sinister and that we realized that every new technology doesn't get a few decades as in the past but only a few years until it gets regulated to death and uh there is a world that is still alive that is still moving forward it's moving into weird directions and so on but there is
11:37
stuff happening and so i think there's also this question is this technology that we are doing is it calcifying faster than we can innovate are things going to shit faster than new things are growing and that is in some ways of an open question but at the moment i have this impression that things are moving again because the future is again changing faster than the existing systems We can adapt and start to colonize them and destroy them.
12:04
And so there is currently a moving avant-garde that keeps things in motion. And the avant-garde is very different from what we thought it would be. We didn't think that the answer to AGI would be use more compute and more data. We thought there was some kind of smart master algorithm, I think. Or many of us did. And then we realized that basically everything we did in the last 70 years in AI was not in the critical path.
12:25
We could have stayed with Rosenblatt's work on the perceptron and then just continued and wait until Moore's loss catches up and we get the following thing, right? LLMs and so on. And I think that what I found fascinating when I read Neil's work was he didn't believe in AGI in his earlier book. So for some reason it doesn't feature. Maybe you can explain where that is. But in Diamond Age, you get an LLM.
12:50
And it's a really smart LLM, but it's not agentic by itself. It's really entirely supplemental. And there is no change of identity of the world due to these technologies, where you get a virtual world that is super smart, but there is no competition to us. It's still us. And the world that we're in right now, it doesn't seem to be about us ultimately, because the next us is going to be different. People are going to hybridize, I think.
13:13
They will have their decisions made by their personal AI that they share all their thoughts with. And so we will change into something else. I don't know if disappointing is the right word for that, Um, that's, that's certainly like significant, right? Um, all right.