How AI Is Shifting with Nathan Labenz

0:01Today on the AI Breakdown, we are talking to Nathan LeBenz of the Cognitive Revolution podcast. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, our newsletter, and our Discord.

0:24Hello, friends. Welcome back to the AI Breakdown. Today, we have on a multi-diverse threat in the AI space. He's a podcaster. He was a red teamer for GPT-4. He works with a company called Waymark. He advises other AI companies. He really is all over the space and can move pretty effortlessly from big picture macro type questions down to really interesting technical things. Hello friends. Quick note before we get to the rest of the episode, you have probably heard me talk about the AI education beta over the past few months. We've had a ton of you participate, which has been amazing. And now we're almost ready to announce something big and something new.

1:01If you want to be one of the first to hear about our new approach to learning AI that is hyper-practical, hands-on, immediately relevant, continuously upgrading, and anchored by community, go to besuper.ai and sign up to be notified when the project goes live. We're getting there in just a few weeks, and I want all of you along for the journey. Once again, that's besuper.ai. In this conversation, we kind of cover the full breadth of those issues with an eye to understanding what the biggest trends in AI are right now and really what everyone needs to know about the space. It's a great conversation.

1:35I encourage you to go check out his Cognitive Revolution podcast. Without any further ado, let's get to it. All right, Nathan, welcome to the AI Breakdown, sir. How are you doing? Doing great. Excited to be here. Yeah. So, you know, we're talking about this a little bit. I am gallivanting around Mexico right now as the listener is hearing this for my wife's 40th. And what I wanted to do is invite a couple of people on who are in this space thinking about some of these same issues, have a bunch of different types of conversations. And the one that I thought would be fun to have with you is almost sort of like a AI state of the union, sort of from a very broad level.

2:10Like you are, you know, you have a bunch of different intersections with this space. You think about it holistically in the context of your own podcast. You do stuff in it as well, you know, in a variety of different ways. So I thought you'd be a great partner for this conversation. So that's the idea. And I think where I want to start is just like, what is your perception of the state of the labs and their big tech partners right now? And I think notably, we're recording this on Tuesday, right after two of the three co-founders of Inflection have announced that they're going off to join Microsoft to be, in Mustafa Suleiman's case, the CEO of Microsoft AI, which is a new thing.

2:51So what's your sense of labs and their competition right now? Well, I guess for starters, I would say there are three clear leaders in the space today. The obvious number one position, I think, still has to go to OpenAI, though Anthropic with Cloud3 has arguably taken the lead in terms of having the best model in public for people to use. and I've always been a big believer that Google DeepMind has the strongest bench, you know, and just the most thorough research agenda. So even though they are a bit behind in terms of the polish of the products, as we've seen with some recent episodes, I absolutely think that they have it where it counts.

3:36I think those are really, and then you could say, well, who else, you know, might be next. Meta would be an obvious candidate if you were going to expand the circle. Mistral might be a good candidate. Inflection, I would have recently said was a candidate, but now I'm not really sure what to make of that. And, you know, beyond that, you could start to think about companies in China and maybe even, you know, go look for somebody in India. But I really do think the top three are a cut above the rest. I think where they are right now is, It's a little bit hard to tell to what degree they are racing with the technology versus trying to proceed with caution.

4:15Certainly, they're all saying that they're proceeding with caution. But looking back just two years, it was only early 2022 when the very first Instruct model came out. And now all three of those companies are basically at a GPT-4 class and promising to continue. So I think my best guess is that we're still in the steep part of the S-curve. It doesn't seem like it will be probably that long at this point until OpenAI makes their next move. And we're already at the point where the frontier models are closing in on expert performance on most routine tasks, even tasks that are very cognitively demanding, like medical diagnosis and things people go to school a long time for.

5:00So there's not too much room left before the AIs will start to really compete with humans, I think, in a very meaningful way. And overall, I think things are about to get weird. What did you make of, if you caught this, Sam Altman's sort of discourse around GPT-5 on Lex Friedman recently? There are a couple things that were sort of notable that he said. Like, one, he certainly played it off. He said that they have a lot of important things to release before it. Two, he sort of like was a little bit like almost dismissive of sort of where they are now. He certainly seemed unbothered by competition.

5:41I mean, what was your perception of that? I haven't seen the whole interview yet, but I have seen those clips that you're referring to. For context, I did have an interesting opportunity to evaluate Sam Altman's public comments for a six-month period between when I participated in the GPT-4 Red Team program and when it was released finally in March of 2023. So I had this window where I was among a very small group of people that knew what GPT-4 was and what was coming and had then the opportunity to watch his public comments with that knowledge and kind of assess like, is he being honest? You know, how should I interpret him?

6:22Basically, what I concluded in that window is that he is pretty honest and that his statements are a pretty good guide to what is to come. He's obviously leaving a lot to the imagination, but I found him to be pretty literally saying what was going to happen. Albeit, again, in a very high level abstracted way. So I would assume that that's probably still his style when he says things like GPT-5 is going to be a similar leap compared to 4 as 4 was to 3. you know exactly what that looks like is a little bit hard to say obviously but i would expect genuinely a very big leap to be coming and in terms of you know things that they are going to launch in the meantime yeah again hard to say i mean they're i think gpts have been a major disappointment really in my experience so i do think there's probably something coming there to make the agent side of kind of the current level of intelligence more useful.

7:24It seems like they have this general sense that they want to be ahead of the curve privately, but kind of dole out the power in, you know, in a strategic way so that people have a chance to sort of adopt and adjust. And I do think there is opportunity to make the agent paradigm work a lot better than it does today without necessarily getting to, you know, that super next level of model that, you know, whether that's Q star or some sort of planning or, you know, something that's capable of doing new science. I think they're going to want to see more action in the world before they maybe turn up that intelligence dial to, you know, the next order of magnitude of scaling.

8:05So I'm speculating, obviously, with those comments, but I do think he's generally sitting on something and speaking about something pretty concrete, even when he's making these sort of, you know, rough, very loose guide to the future sorts of statements. So I would not at all bet against GPT-5 being a huge leap from what we've seen today. GPT-5 being a huge leap would actually be one of the only things that explains how utterly unbothered they seem by all of the like bluster in the press and the competition and Claude three finally being like, you know, if they're sitting there with Sam feeling as dismissive as he was, you know, another set of comments from that interview is he basically said that GPT-4 sucked.

8:49And if that's if GPT-5 is so good that it makes it feel like GPT-4 sucks, it would potentially be sort of, you know, explanatory for how they're acting. I mean, certainly if to the extent that Sora's innovate, you know, how much better Sora is than everything else that we had seen in sort of the video generation space is reflective at all of their sort of behind the scenes aheadness. I think that, you know, they could really just be waiting for whenever they decide is appropriate to sort of reclaim the state of the art title. Yeah, I mean, they have had a year and a half since GPT-4 training was complete to figure out what to do next.

9:27They said for a while that they weren't training GPT-5 yet, but were working on the ideas that they would need for it. Certainly to scale up another order of magnitude beyond GPT-4 would take, you know, a lot of compute and a lot of data. And so no doubt there's been a lot of preparatory work that has gone into that. But yeah, it does seem like they are probably using tools internally that significantly exceed what we have on the outside. You know, even just the base model, from GPT-4, I think is a notable improvement on what is in ChadGPT today for multiple reasons. The RLHF process kind of hurts it a bit on performance in a number of ways.

10:12The training to give short answers and the sort of laziness thing. I'm quite confident that they are not using a lazy version internally at OpenAI. So I would strongly suspect that they do have a, even if it's not GPT-5, that they do have a notably better set of tools for themselves than they make available to the public. I want to come to the agent thing. This is sort of another part of conversation. You mentioned that GPTs have been sort of disappointing in their context as baby proto agents. And I think that that's true. So for me, I haven't been disappointed with GPTs only because it seemed very apparent to me right away that they were just advanced custom instructions.

10:53You know, like it was, it's, they're a great way to not have to enter the same sort of prompts that you use over and over and over again, right? So the way that I use GPTs is like things like creating images for presentations that are standardized and things like that, right? So it's just, it's really just a different way of prompting it. But I think that you are right that they were sort of presented as step zero, let's call it, towards an agent future. And in that they've, you know, people haven't been that stoked on them. How much do you think agents and sort of this focus on agents is like the next big bet in terms of reigniting consumer imaginations for all these labs?

11:30I think it could be a pretty good candidate. You know, I guess I kind of think of our modes of interaction with AI as being today, on the one hand, you've got your productivity assistant chatbot and the interaction there is real time. I'm doing something. I need help. You know, help me draft this. Give me some feedback on this. Help me write some code. I've got a bug in my code, whatever. But in that scenario, you are doing your thing and you sort of have the AI alongside available to help and you occasionally loop it in to help you. That's probably the most common mode of interaction today. And then on the other end is what I sometimes call delegation mode.

12:10And that is where you're really setting up a workflow with the goal of achieving high enough performance that you don't have to supervise every single task or every bit of output that the AI gives you back. And that also can work. You know, if you really dial it in, certainly at, you know, my company Waymark, like we're not, you know, it's real time experience and people often get very good stuff. And I've built a ton of these workflows. You can build them in Zapier. You can, you know, custom code them. You can turn it into a mature app. But, you know, you're dialing in sort of expected inputs.

12:42What are they going to be? What are the steps? You validate it pretty extensively while you're setting it up. But if you're successful, you can get to the point where you can trust it to at least a decent degree because you have kind of controlled the environment and dialed in the performance through prompt engineering, validation, evals, whatever. What's missing in between is the ability to delegate on the fly. It's like you don't really have the ability to say, I want you to go kind of do this for me and have any confidence that it's actually going to happen in a way that you would be pleased with.

13:15You can talk about it in real time or you can go set up a really structured, scaffolded system, but you don't really have anything in between. And I think that that is really what people want. You know, when you think about, and I'm an AI advisor to an executive assistant business called Athena, and really what people want is very often just the ability to take something that's on their plate and put it on somebody else's plate. You know, and it's sort of we have a voice memo app. And the idea is like you open the app. The the mic is immediately recording. You say your thing and then, you know, you it goes away and the app actually self closes.

13:51And this this whole paradigm is around like we want to make it so kind of quick and easy for you to get things off your plate so you can go back to doing what you really want to do. I think if AIs can start to accept that kind of delegation, it will be a major moment because it takes you out of that sort of, I have to be real time with this. And it's kind of, you know, and that could be super useful, but it doesn't require you to go so far as to set up all this scaffolding and, you know, dial in the performance and so on and so forth. So I think for a lot of people that would just be supernatural.

14:22I'm here, I'm doing this right now. Take this off my plate, book this, research that, give me a report back on this. integrate, you know, this API call into wherever. If you could actually just send those ad hoc tasks and get them done, I think it would be a really big deal. And the main reason that hasn't happened yet is just that they just don't work well enough. The core models that we have just haven't been trained on probably enough of that sort of use case to really get there. So going back to, you know, what is, what is Sam going to launch? What is the OpenAI team going to launch before for a GPT-5, if you think of GPT-5 as like a 10x or maybe even 100x compute scale up relative to GPT-4, they could probably do a lot on the behavioral margin with just training on these sorts of tasks.

15:09So they don't need to be orders of magnitude bigger to really dial in like, this is how you reliably make an API call, or this is how you reliably choose which button to click on a website. I think the models are smart enough to do that, but they just need to be dialed in relative to where they are today. So that's why I sort of imagined something like that being a potential intermediate step. Yeah, it's interesting. I think that there's a, one of the challenges for the agents to come to fruition is that on the one hand, you have to sort of nudge up their capacity and get specific about how you train them.

15:41But that also inherently involves guessing at what people are going to find the most useful for, which is really, really difficult. I personally think that a lot of the things that people are building towards for agents as sort of like early use cases, like, I don't know, faster DoorDash ordering are just not going to be that useful in practice, or at least not so much more useful than the ways that we do it now that is going to cause sort of sufficient behavior change. And so it's very difficult because you sort of like, what would be ideal is a generalist agent system. But to get there, you kind of have to have specific agents.

16:10And to get to specific agents, you got to guess at what those specific agent use cases are going to be. So it's a little bit chicken or egg. Certainly, I would say that from the evidence of just YouTube views and download numbers, agents are very, very top of the heap in terms of people's anticipation and excitement and sort of what they imagine AI turning into in the future. I find them fascinating, you know, just and I find them kind of entertaining to watch even. I like just experimenting to see if I can get multi-on to do stuff for me. And, you know, More often it doesn't, but it is getting close.

16:45I'm starting to see the, there's still a little fog between here and there, but starting to see how you could spin up really all sorts of workflows. I always ask people, and this is still more for delegation mode today where you are going to get structured about it, But like, is there a task in your business that either you can't keep up with or that, you know, you would love to scale 10 or 100x beyond what you can do today, but you just, you know, always kind of assumed that would not be possible. The most common things that people come back with are like lead generation. If I could make really high quality personalized outreach to the right targets, that would be extremely valuable.

17:29And similarly for recruiting. If I could identify the right profiles and send them a really good message, every startup founder, they would love to do more of those kinds of things. but they're just time limited. And it's also kind of hard to delegate. And it's, it is also kind of hard to set up. Like you got to first go get the bulk stuff from LinkedIn and then you put it into a spreadsheet and then you got to go through the Zapier or whatever. But the, the dream future is like go find 50 people on LinkedIn for this job description, draft each one, a personalized note, come back to me when that's ready and I'll review it.

18:06And, you know, eventually maybe even send it directly. But for now, I would definitely advise a human review before you send that stuff. But we're not too far from it being able, you know, from these systems being able to take on a task like that. And that number could be 50, it could be 100, it could be 1 ,000. You could delegate, you know, you often take somebody quite competent to do that. And the executive assistants that we work with, they can do it. But, you know, a lot of times they lack context and their judgment isn't necessarily awesome as to exactly which profiles are the best. And, you know, their writing isn't necessarily, you know, what the CEO would want to send, certainly in their own name.

18:41So I do think there is an opportunity for AIs to be on some of these tasks, just a better option, you know, than what people have at their disposal. And so again, if they can just kind of fire it off with a two line note and say, hey, go do this and come back to me when it's done, that will be a real kind of phase change for how a lot of people work. Yep, absolutely. I want to ask you a little bit more technical speculation around where Emphasis is going to be this year. So SWIX and the folks at Layton Space did this four AI wars defining the AI space right now. And one of them was generalized models versus specialized models, right?

19:22How much are the West going to be won by Gemini Ultra or GPT-5 or Claude-4 or whatever, versus lots of different functions having highly specialized models, right? Like, you know, is it going to be Dolly 3 inside GPT or MidJourney that wins? So I'm interested in your kind of take on that. And maybe as a subsection of that, just where you think we're going to see, you know, sort of emphasis from an experimentation and innovation standpoint this year, you know, are we going to just keep trying to plumb kind of ever larger data sets and, you know, make our large language models larger and larger and larger?

19:54Or are we going to see more and more emphasis on, you know, sort of smaller models that can do more with less, you know, or it's both. But I'm interested in kind of your take on some of those sort of, you know, technical prioritization questions that, you know, individual companies are facing right now. One of my mantras is everything everywhere all at once. So I do think it's definitely both. There's not a, you know, I'm not one for extreme positions on questions like this, because I definitely see a lot of value in both. Right now, for me, Cod 3 has recently really crossed an important threshold where it can write as me in a way that is compelling to me.

20:34And I had previously seen that a little bit with Gemini 1.5, which I was able to talk my way into the private preview on, but no other earlier model was really able to do that pretty much at all. And I had even tried I'd fine tuning some advanced models and, you know, just couldn't get anything to really put a draft together that I felt was better than a blank page. You know, you could I with Claude 2, I would do it and I would sometimes convince myself it was helpful. But really, I ended up rewriting everything. So with Claude 3, though, it is really a notable difference. And one of the things that is notable about that is I'm dumping in huge amounts of context.

21:12So I think the sort of rag versus long context window debate, long term, probably, again, it's both because there's a lot more context out there than even the long context that I'm dumping into Cloud 3 right now. But I recently compiled all the intro essays that I've done for my podcast. I do one for every episode, you know, a three to five minute little opening monologue. And now with a bunch of those dumped into Claude 3 and then the transcript of the new episode, I'm able to get a pretty good first draft. and the line of like, what did I write and what did Cloud3 write is actually starting to blur.

21:50The same thing with putting stuff on Twitter. I recently exported all my Twitter data for the express purpose of being able to put all of my tweets into Cloud3 so it could help me write in the style that I write in Twitter. And it really is very compelling. I'm still editing, but I feel like, you know, I'm entering into this sort of cyborg author mode and that seems to only be possible with the biggest models. So there are these kind of qualitatively different things that open up with further scaling. And I don't think we've seen the end of that. And I certainly don't see anybody making that miniature in the immediate term.

22:29That's probably going to be the territory for big models for a long time. Maybe not a long time. A long time in AI time, which might not be that long in other frames of time. At the same time, though, Claude Haiku is also awesome. So, for example, we have had this task forever around identifying which of a user's images are the right ones to feature in the video that they're making. And we've approached it in many different ways, many different models, blah, blah, blah, blah, blah. GPT4V basically solved that problem for us with the one caveat that it's kind of slow and a bit expensive. Although, honestly, it's still pretty cheap.

23:08but it can add up to like, you know, a dollar or even a couple dollars for a user that has a lot of images. So that's not nothing. Haiku can do the same task and takes that price down like an order of magnitude, 90 % plus savings. So that's exciting. And it's faster. I would say Haiku is definitely going to be a major unlock because it's really good and extremely cheap at just a quarter per million tokens. It's the kind of thing where, you know, one of the things that agents have have really struggled with over time is like, what's relevant to this situation? You know, if you're looking through my email, I've got an unbelievable amount of email.

23:46What is actually relevant? You know, you can do a keyword search and you pull stuff up, but then you kind of got to scan down the page and figure out what's relevant. And that's either too expensive, too slow, or just like not effective with most of the models. But with Haiku, I can really start to see how you could just plow your way through a lot of that stuff, you know, and would I pay a nickel for a actually really good search, you know, underlying search process that would find all the context to maybe then put into a cloud three to then write a draft as me for another 10 cents. Like, yeah, you know, I mean, that stuff is extremely, extremely valuable if it actually works.

24:26So I think orchestration is probably where a lot of stuff goes. And this kind of gets back to agent, you know, as well, this is essentially shaping up to be something like in an email agent, Right. Go find all this stuff, run all the way through it, figure out what's relevant, then come back, then draft. I mean, that's sort of the agent cycle and breaking down the key parts of that into what can only be accomplished with the frontier model, a.k.a. right as me and, you know, what can be accomplished with the cheapest model now that there is one that is super fast, super cheap and long context. I think that's where a lot of the sort of application development tinkering is going to happen that will ultimately make these things really useful for people.

25:08Yeah, super interesting. I tend to be on the same sort of it's going to be everything, you know, kind of tip. I think there's just so many reasons for there to be innovation and emphasis at the smaller scale. But to your point, that's not going to change the desire to have sort of state of the art continue to improve. So it just feels like we're inevitably in kind of both end territory. I want to talk a little bit about sort of broader societal level stuff. I'm interested in your perception of sort of like where we are in terms of public opinion, particularly around the safety discourse and maybe how it relates to sort of some other parts of the conversation like, you know, open source and things like that.