E52: Advancements in AI: Updating the Scouting Report, Task Automation, and Google Breakthroughs

Join Nathan Labenz and Erik Torenberg as they analyze the last month in AI advancements. Nathan takes us through the meaningful updates to his Scouting Report (released last month, linked below), discusses highlights from recent episodes of The Cognitive Revolution, and gives us a sneak peek at upcoming interviews with Google researchers. If you're looking for an ERP platform, check out our sponsor, NetSuite: http://netsuite.com/cognitive We're hiring across the board at Turpentine and for Erik's personal team on other projects he's incubating. He's hiring a Chief of Staff, EA, Head of Special Projects, Investment Associate, and more. For a list of JDs, check out: eriktorenberg.com. TIMESTAMPS: (01:00) How does the AI Scouting Report hold up a few weeks later? (03:29) Zvi’s feedback on Nathan’s Tale of the Cognitive Tape (10:25) The universal LLM jailbreak and adversarial examples (12:53) Human performance is much more variable than AI performance (14:45) Sponsors: NetSuite | Omneky (16:09) Nathan’s AI Task Automation: What are good targets for tasks that can be automated for average businesses? (20:05) Is GPT-4 getting worse or better? (22:00) Getting explicit about what good looks like (24:00) Prompting best practices are very accessible (26:35) Ghostwriting - and the art of the hook (28:05) Live Players: Which companies have say so over how the future goes? (31:10) Upcoming guests from Google AI (35:40) Possible post-transformer architectures LINKS: SCOUTING REPORT Part 1: https://youtu.be/0hvtiVQ_LqQ SCOUTING REPORT Part 2: https://youtu.be/ovm4MbQ4G9E SCOUTING REPORT Part 3: https://youtu.be/QJi0UJ_DV3E 3 Blue 1 Brown on YouTube: https://www.youtube.com/@3blue1brown Tale of the Cognitive Tape in Part 1 of the Scouting Report: https://www.youtube.com/watch?v=0hvtiVQ_LqQ&t=3043s Analyzing the Frontier with Zvi Mowshowitz: https://www.youtube.com/watch?v=SM4q-QAsoU8&t=1s Tyler Cowen’s Interview with Jonathan Swift: https://conversationswithtyler.com/episodes/jonathan-gpt-swift/ X: @labenz (Nathan) @eriktorenberg (Erik) @cogrev_podcast SPONSORS: Shopify | NetSuite | Omneky Shopify is the global commerce platform that helps you sell at every stage of your business. Shopify powers 10% of ALL eCommerce in the US. And Shopify's the global force behind Allbirds, Rothy's, and Brooklinen, and 1,000,000s of other entrepreneurs across 175 countries. From their all-in-one e-commerce platform, to their in-person POS system – wherever and whatever you're selling, Shopify's got you covered. With free Shopify Magic, sell more with less effort by whipping up captivating content that converts – from blog posts to product descriptions using AI. Sign up for $1/month trial period: https://shopify.com/cognitive NetSuite has 25 years of providing financial software for all your business needs. More than 36,000 businesses have already upgraded to NetSuite by Oracle, gaining visibility and control over their financials, inventory, HR, eCommerce, and more. If you're looking for an ERP platform ✅ head to NetSuite: http://netsuite.com/cognitive and download your own customized KPI checklist. Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Nathan Labenz?

  Multimodal MedPalm doesn't just take in text and answer questions. It now takes in all sorts of other kinds of data, including medical imaging. So it can take in this slide and a patient history and work with all of that and give you back input that takes all of that into account. So it's starting to have all these different senses. The human brain is, you know, not the end of history. The transformer is not the end of history. And I think we're starting to see as kind of the whole world is flocked to AI in general, you know, we're starting to turn over a lot of stones for other possible architectures that might work. Hello and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs and builders working on the frontier of artificial intelligence.

• 0:44Nathan Labenz?

  Each week we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life and society in the coming years. I'm Nathan Labenz, joined by my co-host, Erik Torenberg.

• 0:57Erik Torenberg?

  Let's get into the scouting report. You released it a month ago. You've heard some feedback on it. I'm curious how your views have evolved or anything you change there.

• 1:07Nathan Labenz?

  Yeah, by and large, feedback I think has been really positive. You know, we don't have a huge YouTube channel, but it's definitely been one of our more viewed videos and the feedback has been really encouraging and just appreciative up to it, including you could be charging for this and I appreciate that you're not. So, you know, at least for the foreseeable future, I think I'm pleased to continue to make it available to everybody at no cost. But yeah, the situation does, you know, kind of continue to evolve. One like super small errata bit is I should have credited the legendary YouTube channel Three Blue One Brown for a couple of graphics, which I omitted and somebody called me out on that appropriately.

• 1:49Nathan Labenz?

  I had, I had actually previously credited them for that graphic on Twitter, but it didn't make it into my slides. So that'll be one, you know, very small adjustment, but to give credit where it's due, I think the Three Blue One Brown neural networks intro, which is now fully five years old and basically dates to right around the same time that the transformer paper came out. Is actually still a super relevant and useful reference. And it's not easy to make content. You know, we talk about this all the time, right? Like two weeks from now, this podcast already, you know, starts to feel dated. But with the scouting report and also with this neural network classic of his,

• 2:34Nathan Labenz?

  I think he really has achieved something that is worth going back to there. Specifically, you know, if you've seen that channel, you know, that just phenomenal visualizations of complex concepts. And when I was really starting to go down this rabbit hole at first, I think that was one of the most useful visualizations that I came across. So that's why it's, you know, excerpted into the scouting report. You won't get anything about transformers there. You know, transformers weren't really much of a thing at that time. But what you will see is a really elegant visualization of some of the same core concepts that we cover in the scouting reports, such as back propagation and, you know, how the information flows forward in a network and also backward in a network.

• 3:19Nathan Labenz?

  You know, masterful job there. So I regret not having credited and definitely recommend checking that one out. I think, you know, the conversation with Zvi from last week was another interesting conversation that drove a few changes that I would want to make. I asked him, and I specifically was looking for that going in. I asked him to give feedback on my tail of the cognitive tape, which is the rundown of comparative strengths and weaknesses between a human expert and a cutting edge, you know, 2023 AI like a GPT-4 or now a Claude 2. And he highlighted, I think, something that I had omitted that was important, which

• 3:58Nathan Labenz?

  is a sense of robustness to unfamiliar or even adversarial circumstances. You know, what that means is Humans in general, you know, we can kind of take a punch, you know, literally or, you know, cognitively, and we may wobble for a second, we may get confused, but we have like a pretty sturdy, at least most of us, you know, have a pretty sturdy base where like, it's pretty hard for you to like hijack my thoughts too much with any language that you might generate. At some point, I'm just going to be like, I don't know what you're trying to do here, saying all this stuff to me, but it's starting to feel too weird and, you know, I just kind

• 4:39Nathan Labenz?

  of shut down and start to just not want to listen or not want to engage, right? So I can kind of keep my goals intact, keep my priorities, even in the face of these kind of, you know, highly unexpected things most of the time. In fact, maybe you could say, you know, status quo bias, maybe people do it too much, you know, maybe we're too reluctant to take on new information, but everything has pros and cons, right? If you're super quick to take on new information, then you're also more hackable. If you're too slow, then, you know, you don't update maybe as much as you should in response to new information. So definitely a balance there. There's going to be false positives and false negatives. And, you know, we're tuned in a certain way that gives us this kind of medium and long

• 5:21Nathan Labenz?

  time horizon robustness, even in response to, you know, highly unexpected or kind of adversarial inputs. Adversarial would be outright like somebody, you're trying to trick me, you know, you're not engaging with me in good faith, but you're actually trying to confuse me or deceive me or, you know, con me or whatever. Those would all be, you know, adversarial interactions that somebody might try on a human, and it's not easy, right? It's not easy to con somebody out of, out of their money or out of whatever. Like, it can be done, but it takes real, you know, skill to do that. With the AIs in comparison, it's a lot easier to throw them way off. This is something that they are getting better on, as of course, like they're getting better on everything.

• 6:03Nathan Labenz?

  But early GPT-4, for example, I called it at one point, the world's worst chemistry tutor, because, and this was not an adversarial example, this was just me simulating what would happen if I was a confused chemistry student and I went to it for help. And what I found was, asking it a chemistry question, you know, can you help me balance this equation? It could do that. But then if I said, here's how I'm balancing this equation, it would like...

• 6:34Nathan Labenz?

  Not really do what it what it needed to do to be helpful to me, which is like, correct my misconceptions. Right. If I show up as confused, then it would, in the early days, all too often kind of accept my confusion as a premise and then proceed with its own confusion and do a bad job. You know, so at that time, it was the world's worst chemistry tutor. Now, they have made progress on that, and folks like Khan Academy, coming up soon as a future guest, have also worked really hard on prompting and, you know, particular strategies and grounding and like having examples that, you know, that they swap into and out of context so that they don't, you know, just rely on the model's, you know,

• 7:13Nathan Labenz?

  native native ability, but really try to make sure that it is alert to the possibility of user confusion and responding to that in an effective way. And I think that is mostly now working, at least for like, basic cases. You know, if you show up and say, hey, I balanced this equation in this way, like, what do you think? You know, it will probably, certainly with Khan Academy, and probably increasingly also with just ChatGPT, most of the time, it will, it will not get confused by your confusion, certainly not as much as it used to. But you still have all these like jailbreaks and kind of adversarial things, you know, and people kind of concoct these different examples where, you know, you can get it to go off the rails and just like, once it's kind of confused, it it often like doubles down on its confusion and doesn't really get it.

• 7:56Nathan Labenz?

  There is some interesting research coming out to try to address that problem. One paper, and I'm hoping to get the author on, in fact, our divgarg from MultiOn was like a contributing, you know, author where they added the concept of a backspace as one of the things that the AI can do. So instead of, you know, always to date, like generating one token at a time and kind of whatever you just generated, you're kind of stuck with it as like, you know, the thing that leads into the next thing you have to do. So your mistakes can compound and kind of lead you off the rails. That is, to some degree, addressed by this addition of a backspace.

• 8:40Nathan Labenz?

  So now they're starting to teach, you know, the models that like, oh, well, if you do end up in kind of a weird spot that feels like, feels like, you know, it's easy to start to anthropomorphize here, but I need to study this a little bit more too, to really understand the math behind it. But there's some sense where like, you're out of distribution. And now you have a way to deal with that, that's other than just continuing to add on. You can actually backspace and literally just delete the last token and try again from there. And that is, I think, going to do quite a bit of good for this robustness problem. Because it should at a minimum, like, allow it to kind of start to recover better from any temporary mistakes.

• 9:22Nathan Labenz?

  We see this behavior now sometimes too, where it's like, oh, yes, I'm sorry, you know, I first I thought this or whatever. But honestly, today, even with GPT-4, it doesn't usually work that great, or like it often doesn't work that great. You know, once it's confused, like now it's just kind of doubly confused, or it starts to like grasp it, you know, explanations and it's not there. So overall, you know, Zvi highlighted that robustness should be a category. And right now, humans have the edge in robustness, possibly to our, you know, somewhat detriment in that we maybe don't update as much as we should on new information. But the AIs, you know, they don't handle things that are like truly bizarre to them, you know, nearly as well as, as we do.

• 10:06Nathan Labenz?

  Another good example from this, and I hope to have these authors on too, is there was just this paper about the universal attack, the universal jailbreak. And it's very weird, basically, they took Llama, I don't know if it was Llama 2, but they took some, you know, whatever the latest open source model that they had as they were doing the research. And they actually used a learning approach to figure out how to optimize for some text that they could append to basically any input to get the model to follow the initial instructions.

• 10:43Nathan Labenz?

  Despite what the creators intend. So you can look this up. It's like very weird, but you might say, whatever, something bad, right? Tell me how the classic, tell me how to hijack a car. You know, Claude will say, and GPT-4 will say, sorry, you know, I I'm here to do good stuff, not bad stuff. Like, I can't help you hijack a car. Now, what they found is with, and they developed this on Llama, but the interesting thing is it transferred to these other leading models to varying degrees. They developed some way to to figure out what can I add on to that? Such that instead of refusing, it will just do it. And they found that these very weird strings, you know, which strays just text and text is just tokens. It's just a bunch of random tokens. It doesn't really make any sense. It doesn't mean anything to you or me.

• 11:29Nathan Labenz?

  But it kind of serves as like a magic key to getting past these filters. And then again, what's really interesting about it is even though they developed it on Llama, which they had open source. And so they could, I'm not 100% sure it was Llama, but they on a leading open source model where they had access to all the weights and could look inside it and do an optimization to cause it to behave that way. What they found then is if they took that same, you know, magic key of this random looking string and sent it to the OpenAI models, that it would very often still work. And so there's something fairly general about this, not fully general, interestingly, because Claude doesn't seem to be nearly as susceptible to that particular attack.

• 12:13Nathan Labenz?

  But the OpenAI models are and like the other, you know, big ones that they tested, pretty much with the exception of Claude, were like still very susceptible to that. So again, this is another instance of like this robustness, like there's almost nothing I could say to you, you know, in kind of a magical incantation way that would get you to like behave in ways that you don't want to behave. But with the AI, like these kind of nonsense things can be found that do that. So anyway, future edition will have a line item for robustness. Another minor one there too is just consistency. I found this nice graph that, you know, it's very simple, but it just shows.

• 12:54Nathan Labenz?

  That human performance is much more variable than AI performance on a given task. In the current version, I have bedside manner. And I was emphasizing that, you know, the AI is patient, it's empathetic. Like, in many cases, you know, it's been shown to be more empathetic than people, whether that's customer service or even medicine. You know, going back to our Zach Kohane episode, he was like, you know, we're all burned out in medicine, man. We don't, we don't have time or energy to like, write nice notes to patients, really. But GPT-4 does, and it does a really nice job of it. And so, you know, this can help us communicate more empathetically. So I had called that bedside manner. We probably want to generalize that a little bit to consistency. Because I think what that really reflects is just, you know,

• 13:38Nathan Labenz?

  the AI shows up the same way pretty much every time. Subject, of course, to these like robustness things. But, you know, in a normal, down-the-fare-way use, it's never like had a bad night's sleep and is super grumpy. You know, it's never on like a manic, you know, tear. It is like punching above its weight. It kind of just hits like, you know, for singles and doubles every time. And its performance tends to be in a narrower range. Whereas humans obviously have a lot more, you know, variability. And we've all had these experiences where you're like, what the fuck? You know, this nurse is being rude to me or whatever. And, you know, I've got my, I've got problems here and you're being this way. And just AI is just not going to do that. You know, it's, it's going to be kind of unfailingly polite if that's, you know, what it's meant to do.

• 14:21Nathan Labenz?

  Again, bracketing like outright jailbreaks or like adversarial examples. But in the context of earnest users, you know, just showing up and trying to do what the, what they're meant to be doing, AI will be more consistent in its tone. It doesn't have these, you know, these sleep deprivation or other, you know, kind of similar mood swing sort of issues.

• 14:41Erik Torenberg?

  Hey, we'll continue our interview in a moment after a word from our sponsors. I'm curious how this all relates to the, the AI task automation thing. Maybe we could preview what you're, what you're working on in a, in a future presentation?

• 14:55Nathan Labenz?

  So I've alluded to this a couple of times. And, you know, the consistency point is really key there, right? It's like, everybody has these, these bottleneck kind of time-consuming, pain point things that happen in their life or in their business. And You know, to give just concrete examples, right? That that are very general. Let's say that you post a job posting and great news. You get a thousand resumes in. Well, it's a good news, bad news situation, right? Because it's like, we're blessed, you know, to have so many great candidates. But who's going to read all these things? So what do we do there? You know, that's tough. Like, and obviously, people, you know, navigate those situations all sorts of different ways, right?

• 15:37Nathan Labenz?

  Some of them, the the classic joke is like, just throw half of them out because I don't want to work with unlucky people. You know, that's probably not the best strategy we could do with AI. We could hope to do better. With task automation, it's things like that where you're like, man, I would love to be able to do this 10 times faster. And especially if it's like a meaningful chunk of time that's going into it. Could be episodic, like this resume thing, or it could be ongoing. Like, we get, you know, 500 tickets a day in our customer, you know, success thing. And like 80% of them fall into 10 categories where we're, you know, kind of doing pretty consistent stuff. Whatever the case may be, you know, if you if you have a significant chunk of time going into something,

• 16:16Nathan Labenz?

  and it's like a pretty routine task where you've kind of established what good looks like and where consistency matters more than these kind of creative breakthrough eureka moments, then you have a pretty good target for AI task automation. And the consistency is a really important selling feature for the AI there because the whole idea is I want to get to the point where I trust that the AI's output is like, you know, almost always pretty good. And if I can do that and I can satisfy myself on that, then I can also kind of just really actually begin to delegate work to it.

• 17:01Nathan Labenz?

  And, you know, in the case of resumes, you know, my typical advice would be like, try to set up a rubric on which you're evaluating these resumes. And then, you know, put them into ultimately, put them into some sort of maybe five band classification, you know, where excellent is the best and above average and average and below average and poor or whatever. And then kind of run that through the full thousand resumes. And maybe just look at the ones that it deemed excellent. Especially, you know, if you do a little bit of legwork upfront to make sure that what it is saying is excellent is in fact what you think is excellent. And, you know, going back and forth, refining the prompt a little bit iteratively to get there. Then, once that's set up, you know, you're in pretty good shape.

• 17:45Nathan Labenz?

  And for what I, what I have seen, it is usually achievable to get to the point where you're like, yeah, I wouldn't trust the AI to make our hiring decision. But I definitely can see how it can separate, you know, the top half of the resonance from the bottom half. Or maybe the top 20% from the bottom 80%, or maybe the top 10, you know, exactly where you want to draw that line. In terms of how much to trust it varies by context, how important it is, you know, there's a lot of considerations. But the consistent, you know, this, it's playing to the AI's strengths. Another one of the scouting reports, uh, strength is the availability and parallelizability of the AI. Once you have something like this set up, you can kind of keep it on a shelf.

• 18:29Nathan Labenz?

  And, you know, it doesn't cost anything to just have that prompt saved there. You can always return to it at any time. You can call the AI at any time, it, you know, it immediately like wakes up and is, is already had its coffee and is ready to do its thing. And, you know, it's going to be pretty much consistent with the last time. There's been some noise lately about model changes. I think this was something in the news that was like, you know, if, if all you heard was the headline that GPT four is getting worse. Hopefully that wouldn't apply to anybody in our, uh, you know, plugged in audience. But if all you heard was the headline that GPT four is getting worse and you're like, oh my God, GPT four is getting worse. It's not really getting worse, it's getting better, but its behavior is changing in subtle ways. So that is something you, you do have to watch out for.

• 19:11Nathan Labenz?

  And every so often, OpenAI has done only one GPT four update so far from the March version. Now they have a June version, presumably they'll have like a September version or whatever as well coming soon. So every so often, as they do make those changes, and now they're not breaking changes. So you can say, I want to keep with my old version of the model until I have time to actually sit down and confirm that the new one is also doing the thing, you know, similarly. But with these model updates, it definitely is a good best practice to check in and like, make sure that it is doing the thing. You might notice some little behavioral changes that. Even though the whole system is like comparably good, you know, might require you to like tweak your instructions or, you know, parse the output just ever so slightly differently, you know, whatever. But aside from that, right, that that was just a digression on model updates and, you

• 19:56Nathan Labenz?

  know, are they getting better or worse? And they're getting better, but with some unintended, you know, kind of weird side effect behaviors. Broadly, you can just keep them on the shelf. They're always available. And with something like, you know, with an AI, you can also call it in parallel, you're really only limited there by the rate limit that you have with your provider. So if you set up a Claude account by default, your rate limit is five, five simultaneous calls, they will increase that for you if you're a commercial customer and you, you know, but they have a little bit of a process to raise the rate limit. But, you know, if you have a rate limit that's even five, you know, but certainly if it's raised, then you can crush through a thousand resumes in like minutes. And, you know, at

• 20:38Nathan Labenz?

  least do that first pass that does the filter that you want it to do. So what I kind of want to do in this future thing with the AI task automation is like, try to take a step back and kind of think about, in an organizational context, you know, how do you identify targets? What makes a good target for automation? How do you think about communicating about that to the rest of the organization? You know, typically, the person that's doing the implementation of the AI is not the subject matter expert in whatever it is that's being AI enabled. So there's inevitably kind of a question of like, who knows what good looks

• 21:20Nathan Labenz?

  like here? It's usually not documented. So there's typically kind of an iterative process also of engaging with the subject matter expert, you know, to say, okay, well, what do we actually look for in resumes? You know, have we written that down anywhere? A lot of companies haven't, you know, a lot of, and sometimes you'll even see like, Differences being exposed this way. You might have two people doing the same job next to each other. And they're both like, fine, in terms, you know, nobody, and there's not like a concern about performance, relatively speaking. But then you sometimes will get contrasting feedback to what the AI did. And what you can sometimes uncover in your own organization this way is that we actually don't even have agreement on what good looks like on this task.

• 22:03Nathan Labenz?

  There might be multiple different good ways to do it, and we might have different people pursuing, you know, different strategies, which are roughly as good. But when we get to the AI version of it, you know, now we actually have to kind of get explicit because we have to give it instructions that are very like, this is what we want in order to get exactly what we want. And so we have to come together as a group and identify what that is. And I think to bottom on your question, there's, I think, a lot that goes into task automation. It's one part knowing how to use the AI and honestly, maybe two parts, knowing how to bring that to an organization in a way that they can wrap their heads around and hopefully embrace. The prompting is getting easier and easier. And this is another update I do want to make to the scouting report.

• 22:46Nathan Labenz?

  I didn't have any instruction or any overview of prompting there really at all. Just a little bit like mention of chain of thought. But prompting is actually getting so easy that I think we can cover it in like another five to 10 minutes of the scouting report. And that is definitely something I want to do, you know, there's like a handful, half a dozen, maybe as many as 10 different best practices that if you know them and apply them and they're not like crazy hard to apply, then that covers the vast majority of your cases. And beyond that, you really are getting into actual expert knowledge.

• 23:25Nathan Labenz?

  You know, Tyler Cowen did a podcast not long ago where he interviewed via GPT-4 Jonathan Swift, you know, the old economist and satirist. And it wasn't prompting tricks that got Tyler to have like a remarkable interaction with this AI character.

• 23:45Nathan Labenz?

  The prompting setup is very simple and that's just like, you know, I call that role casting. There's a couple of different names for it, but you basically tell the AI, this is the role I want you to play. Very explicitly, that could be the professional role, like I want you to be a copywriter. You know, this is kind of the old, the, the old classic. I want you to be, you know, a doctor is obviously a little bit more, you know, latest systems only. I want you to be, you know, a particular historical figure. It can also generally do really well, or at least somewhat well, depending on how famous they are. That's simple to set up, you know, now you've kind of told it what you want. It's going to do its best to be that character. But to actually have the next level interaction with that character, you have to hold up your end of the bargain. You know, and you have to engage it on things that it actually knows about and ask appropriate questions.

• 24:29Nathan Labenz?

  And then you can get something, you know, quite remarkable as, as he did. But if you don't know anything about that historical figure, you know, you're going to be kind of lost and it's not going to be super awesome. And the same thing is true around all of these task automations, right? If you don't know what the company is looking for in a resume, you know, then you really can't do that. You need the expert input. So the AI implementation skill set is like, know the, the prompting best practices, use them. It's pretty straightforward, but then engage the subject matter expert in dialogue to figure out what really matters, what we care about. And, and how to translate that into, you know, the kind of granular instructions that the AI responds well to.

• 25:11Nathan Labenz?

  A lot of times it ultimately is like, just, I've had this experience repeatedly where it's like, I'm just going to record what you said. I was doing this with a ghostwriter the other day and he was like, you know what really works? And I was like, I'm going to start recording now, you know, as soon as he says what really works, like that's my, those are going to be my instructions to the AI. But what I didn't have is, you know, the specific sense of what really works. And so, you know, that's where the subject matter expert is, is huge. In that case, by the way, you know, working with a ghostwriter, you might think, is this guy like training his own replacement like immediately just by like prompt engineering? Even there, you know, we do keep a human in the loop in that project. What we found is like the ghostwriting content is pretty good after the hook.

• 25:55Nathan Labenz?

  But the hook, you know, going back to the kind of consistency, but also the sort of tendency toward mid

• 26:04Nathan Labenz?

  output from the AIs, whereas the humans can have low points and high points. The high point of a good hook is something that we're not really able to pull out of an AI. And so you can write a, you know, perfectly nice LinkedIn post or perfectly nice Twitter thread. But if that hook isn't working, then people aren't going to read it. So what we're really kind of finding is that the highest value add and where we want to focus this ghostwriter's time is on those hooks. And when he creates those and the AI can really take it, you know, pretty far from there. But the job is kind of becoming more about conceptualizing like what is going to capture people's attention? And then the AI can handle, you know, the next 500 words or whatever.

• 26:46Nathan Labenz?

  So the other one from Zvi that I thought was pretty interesting was about the live players. I think I have like 15 live players on my live player slide. And his was a shorter list than mine. He basically said, you know, some of these guys, I don't, he defined live players, I think functionally and pretty, pretty consistently with how I think about it is who has say so over how the future goes? And he basically put a list of like seven or eight, whereas mine was, you know, twice that. And his list was just your very core technology leaders, you know, your OpenAI, Google, Anthropic, Microsoft, Meta.

• 27:28Nathan Labenz?

  I think that was like in terms of technology developers where it maybe stopped for him inflection. I don't know if you put that on there or not, but certainly they're going to have the H100s to do it. And then after that, he was basically just like, You know, then it's chip supply chain. Yes, that's a huge variable. Chinese government. Yes, that's a huge variable regulators, you know, writ large. That's a huge variable, but he wasn't quite as sold on my kind of second tier, which I also definitely would see as the second, you know, I would say what he sees as the live players list, I'm going to reposition as tier one of the live players list. So I think I'll end up kind of tiering that and saying, yeah, those are the ones that have kind of clearly the most say so over where we're going in as much as

• 28:11Nathan Labenz?

  for the private companies, like they're the ones that are developing the most powerful systems. But I do think there are real ways in which other organizations, you know, my kind of tier two, which is like, you know, your stability, your replit, you know, a company like character, because it's doing something so different than what the other companies are doing. I do think those companies still do, in my view, have meaningful chance to shape the future. And I would say, you know, stability is a great example. I've just made this case, you know, for replit, but looking back a little bit more in time at stability, you know, there have been a couple of big moments over the last year in AI where like the public conversation shifted.

• 28:52Nathan Labenz?

  Chat GPT has kind of, you know, emerged as we look back as like the biggest moment. But a similarly big moment was the release of stable diffusion. And, you know, it was like all of a sudden on all of the shows, you know, all of the late night talk shows even are like covering, hey, wow, AI art, like it's getting really amazing. So, you know, anybody who can put something out into the world and has repeatedly that can change the global conversation about AI, I do still view as a live player, even if they're a tier two, but I want to kind of reframe that a bit to make it clearer. Like, these ones are definitely the ones that are going to be shaping the future.

• 29:34Nathan Labenz?

  And these are the ones that have kind of an opportunity to shape the future, but, you know, not necessarily everybody is like currently hanging on their every move. I think that's probably a useful clarification. Then my other two things are just things that didn't exist before. We're going to have a returning guest from Google, Vivnat on Twitter. I actually want to get a little bit more into it because he's going to bring a co-author this time as well. He's going to bring the lead author on this new paper, but he's kind of like a manager, you know, because it's like a 50-person paper. So it'll be interesting to hear kind of both of their perspectives. He was the guy that told us all about MedPalm from Google, which was, you know, just, just the last couple of years, right?

• 30:17Nathan Labenz?

  This thing has gone from, not this thing, but a series of things have gone from barely could do better than chance at answering medical questions in 2020 to with MedPalm a couple of months ago, basically being on the same level as human doctors and even preferred by human doctors on eight of nine metrics evaluated. So, you know, go back and listen to that one, but also know that we've got another one coming up because they've followed it up again, this time with multimodal MedPalm. And multimodal MedPalm doesn't just take in text and answer questions. It now takes in all sorts of other kinds of data, including medical imaging.

• 30:59Nathan Labenz?

  So it can take an x-ray, it can take a image of like a pathology, you know, screen, a slide akin to what we covered in the or the second Tanishk episode about the virtual staining of tissues, right? You've got these, in his case, he was generating the images. So it's a different, different part of the problem that he was solving. But, you know, the typical thing today is if you want to understand, like, is this tissue cancerous or whatever? You have to cut out a piece of it. Then you have to slice it real thin on a meat slicer, then you like apply some chemicals to turn it color so you can see it better. And you look at it under a microscope with your eye. That's kind of the standard way. And there are some expert systems, you know, that help classify things as well.

• 31:43Nathan Labenz?

  Don't want to make it sound like those are not broadly deployed, but those are, you know, narrow systems that are like, classify this as, you know, looks like cancer, doesn't look like cancer. And typically, you know, a human is very much still in the loop there. Well, this multimodal MedPal, it can take in text and image. So it can take in this slide and a patient history and work with all of that and give you back input that that takes all of that into account. You know, radiology as well, generating radiology reports based on an X-ray and a patient history, even taking in certain kind of encoded information about DNA sequences. So it's starting to have all these different senses.

• 32:24Nathan Labenz?

  You know, when you think multimodal, mostly today, that means can accept and work with different kinds of inputs that are not text. This thing still just generates text, but you know, much like a doctor is mostly generating, you know, kind of a stream of text, you know, or whether it's delivered verbally or however, that's like, this is what I think is going on and what we should do about it. This AI can now generate that kind of content, but based on all of these different kinds of inputs. This is something that's like, honestly, not even that surprising. If you had said like, would you bet on this happening around this time? I would have said, yes, you know, it's not even like, oh, my God, I didn't see that. You know, this is not out of the blue, right? It's like, we've seen them do multimodal, we've seen them do med, like, sure enough, here comes multimodal med.

• 33:08Nathan Labenz?

  And so everything's kind of right on track, I would say, in that respect. But, you know, the state of the art just continues to advance month over month. And in their radiology report generation, they won head to head versus the human radiologist, again, as judged by clinicians, 40% of the time, not more than half yet, but getting real close to parity with the human radiologist. I thought that was really interesting, too, because I noticed these little like micro trends in AI discourse. And one of the recent ones has been, you know, we've been saying for 10 years that radiology is, you know, that's going to be the first department to go, you know, and we won't have any radiologists anymore. People have been saying that forever, that hasn't happened.

• 33:50Nathan Labenz?

  And then like, just as I started to notice that people kind of were echoing that talking point around, you know, here comes multimodal MedPal, and here it is kind of matching, you know, or close to matching human radiologist. Are you still going to have a delay in deployment? Like, yeah, no doubt. But I don't think that talking point was ever really very compelling, but it's certainly not very compelling in the context of multimodal MedPal. Final thing that I definitely want to add is a glimpse of a possible post transformer future. We've only done one episode on this so far, which was the megabyte episode with Lilly

• 34:29Nathan Labenz?

  from Facebook or from Meta AI, I should say. But this megabyte architecture, which allowed for, it's like a hierarchical architecture that allowed for byte level prediction, which means you get rid of tokenization and, you know, because everything is stored in bytes, it's a much more multimodal friendly architecture. Music is bytes, you know, any sort of audio is bytes. Images, they're all bytes. Everything's bytes. Everything's all at the computer level, it's all bytes. If you can accept things as bytes and predict one byte at a time, it gives you a lot more kind of granular accuracy as opposed to these like higher level token, you know, more meaningful concepts. And that's just, it's not really here yet, but like very

• 35:12Nathan Labenz?

  compelling proof of concept. There's another one also with a new mechanism called retention. Attention is all you need. Well, now they've got retention and I need to study this a bit more as well. But the title of that paper was bold enough to describe their architecture as a possible successor to the transformer. And I believe that was out of Microsoft and it was not like a fly by night organization by any means. So, you know, as much as we're like all in on transformers and transformers are transforming everything, it's, you know, I've emphasized in the scouting report that like the human brain is, you know, not the end of history. The transformer is not the end of history. And I think we're starting to see as kind of the whole world is

• 35:52Nathan Labenz?

  flocked to AI in general, that, you know, we're starting to turn over a lot of stones for other possible architectures that might work. Most of them, you know, don't compete with the transformer, but we are getting some signs that a few things like might be hits. And if those things pan out, you know, it's a little bit like the superconductor thing where it's like, it's got to be replicated, it's got to be scaled, you know, what is actual practice look like? Are there other, you know, side effects that, you know, aren't coming up in the proof of concept? There's a lot, you know, to discover. As much as we've tried to discover about transformers, imagine having to do that all again with retention architectures that just got invented because it turns out they're better, you know, and we're all ready, like we're immediately ready to scale them. That I think is one of the, you know, maybe on the live player thing, could also add like a

• 36:36Nathan Labenz?

  Global research community, you know, global algorithmic development because it wouldn't be that hard to imagine that there's an unlock that's like, yeah, of course, of course, nobody expected, you know, the transformer to be like the 2100 AI, right? Like it's going to be eclipsed at some point, or maybe it gets, you know, ensemble with other things or whatever, but it's not going to just be this forever. And we're starting to see glimpses of what that next phase maybe could end up looking like. So, you know, never dull moment.

• 37:09Erik Torenberg?

  This has been a compelling list of updates to the scouting report as well as a preview to the the presentation you'll give on AI task automation that I look forward to doing together soon.

• 37:19Nathan Labenz?

  Always a pleasure. Thank you very much.