00:18:37 · Ruby Justice Thelot, Herbie Bradley
[00:18:37] S1: I think Nathan Calvin had a tweet on this. He was like, form video... Like, time spent watching short form video has increased. Yeah. Yeah. Because, like, Americans, when they have to work less, they're just, like, spending more time doing talk. [00:18:53] S0: Yes. Yeah. And then the final reason, I think, is just I think Shyam Shankar had a recent tweet where he put it quite well that the atomic unit of productivity in a large enterprise is not really the individual. It's sort the workflow level. And so if you improve one individual on a team's productivity by 10x, then that doesn't necessarily mean that the overall output of that team improves a ton. Even if you have somebody at the head of the power law in your team, that doesn't necessarily improve things, unless you can do some amount of reconfiguration of the work that you're doing or some change management within the enterprise. And so I speak to a lot of people who are doing forward deployed engineering, doing AI transmission projects in enterprise, and they all basically give me the take that diffusion is kind of bottleneck on change management and getting people to accept different ways of doing organizational processes and work processes.
00:19:41 · Ruby Justice Thelot
[00:19:41] S1: And to me, like, this is an area where I guess I am more of a capabilities maxi because it just seems so easy and so intuitive to me. Like, you set the thing up literally by opening a new tab in your browser, and you use the thing literally by telling it what to do. And it's so simple. It's not like you have to learn how to use, you know, some complicated new piece of software like Photoshop or even an Excel. It's literally just you have a little virtual coworker that you can talk to and get to do things for you. And anybody with a credit card can have access to frontier intelligence that all but a few thousand people at the labs have access to. Like, it's just so crazy to me that there isn't much more and much faster diffusion. Like, it's not like you have to build out a whole new network of, like, fiber optics or something.
00:20:29 · Herbie Bradley
[00:20:29] S0: Yeah. Yeah. I mean, I think it is still in some ways overrated, like, that you don't need to do much education or training or whatever to get the average white collar worker to use AI effectively. It sort of reminds me of kind of teaching my grandma to use Google back in the day. It's like, here is this search box where you should, in favor, be able to just type in anything you want to know and get a link to the answer. And yet, it requires some familiarity and some reps and some training of some form in many cases, or it did in certain of the 2000s when enterprise were adopting the internet technologies for the first time in order to really get large productivity increases from that. And so I think this is partly what the job of the FTE is, is firstly helping people inside the company to figure out how to use the thing, and then secondly, figure out what are the organisational blockers to doing that and figuring out ways to route around those. And often that may involve kind of strange things like, oh, some middle manager has some internal politics reason to keep this data siloed, so the FTE needs to go around him to the skip level manager and persuade him to do it. There's just like a ton of internal politics which blocks diffusion in that way, at least.
00:21:38 · Ruby Justice Thelot, Herbie Bradley
[00:21:38] S1: Right. So I guess we should talk about the topic of the day, which is Anthropic S-1. They just submitted an amendment. They added a new line about RSI. They said their internal estimates indicate that they may imminently meet their responsible scaling policy's definition of dramatically accelerated progress attributable to AI. Obviously this is a big deal because anything that you put in an S-1 or any official filing with the SEC has to be completely true. You're not allowed to lie in these filings. So Anthropic clearly believes that they are on the cusp of RSI or that they may be on the cusp of RSI. So what do you think they're seeing internally? [00:22:19] S0: Yeah. They have their latest generation of models, which is not yet released, and I expect they're seeing much greater aggregate research and engineer productivity uplifts. Overall, do think we would benefit the public and analysts outside and the market, in fact, from much more high resolution reporting [00:22:37] S1: Right.
00:22:37 · Ruby Justice Thelot, Herbie Bradley
[00:22:37] S1: Right. [00:22:37] S0: From the labs of exactly what types of productivity uplifts they're seeing. Can we get more high resolution employee surveys? What are the latest unreleased models? I think in August, Anthropic had a system card thing where they reported, oh, we have this unreleased model, and we have seen this amount of productivity uplift from it. The responsible scaling policy in this case has, I believe, this threshold that they refer to in their S-1 of a two x overall productivity uplift, meaning they're making progress two x faster than they would be, I think, pre AI with this latest ML model, or I should say they expect to achieve this threshold with a near future model or imminently with the term I believe in the S-1. And so that 2x basically translates into, if I recall, around the kind of 10% researcher productivity improvement per unit of the Epoch capabilities index benchmark, which still, there's various different people trying to model RSI. It translates into an R of around 0.5 if you have some scale where like R equals one is a fully self sustaining acceleration loop. [00:23:48] S1: Right. So wouldn't that imply that we don't have a fully self sustaining acceleration? [00:23:53] S0: Yeah, that's correct. But they did make this threshold as concerning enough to report as an RSP thing. Right. [00:24:02] S1: What do you think it would take to get from where we are now to full self sustaining acceleration? Why do you think Anthropic thinks that they will go from 0.5 to R1 or greater?
00:24:02 · Ruby Justice Thelot, Herbie Bradley
[00:24:02] S1: What do you think it would take to get from where we are now to full self sustaining acceleration? Why do you think Anthropic thinks that they will go from 0.5 to R1 or greater? [00:24:16] S0: I expect just having internal theories about exactly how different aspects of the research pipeline can be automated. You know, one example is John Schulman, I believe, on the podcast, proposed that researchers could sort of distill their high level taste and knowledge about experiments into models by doing LHF or sort of becoming human raters for different ideas. I'm sure Anthropic have a whole roadmap internally where they have gamed out a bunch of this stuff and kind of see the rate of improvement and see that it's getting there. I mean, we know that they published this blog post a few months ago with the graph of different levels of RSI and sort of making progress towards more and more autonomous research. I personally still think it's quite likely that even if you do get a period of self sustaining acceleration or you can in theory close the whole loop, that it will be desirable for various reasons, including to avoid good hearting or to avoid safety issues to keep some researchers in the loop, because at some point, the speed of the overall process doesn't end up being bottlenecked by the researchers. It ends up being bottlenecked by compute Right. The time it takes to do experiments and data and this kind of thing.