Oct 26, 2024 ยท 2024 #38. Read the transcript grouped by speaker, inspect word-level timecodes, and optionally turn subtitles on for direct video playback
Edit labels for this show, save them in this browser, or download a JSON override for the production folder.
Transcript Playback
Agents and Robots
Human Transcript
Timed transcript
Blocks are grouped by speaker for readability. Expand a block to inspect word-level timing.
Speaker 2
Hello, everybody. It is Saturday, October the 26th, 2024, the last Saturday in October. In fact, the last that was the week tech summary that Keith and I are going to do until, of course, November. We are joined, Keith and I, with a couple of guests from a company called Notebook LM. I'm going to introduce them.
Yes, specifically how AI is becoming more than just a tool.
Words and timings
Yes,specificallyhowAIisbecomingmorethanjustatool.
Speaker 2
Right. It's becoming almost like an independent actor. Well, we're all sick of those two voices, but we're not sick of Notebook LM. It's really an incredible product. And Keith ran the newsletter this week through Notebook LM to come up with that clip. I think, Keith, you've got the whole video. At what point are those two people going to replace you and I?
You know, if I had a couple more hours, I could take their script and put you and me on as avatars with our voices, and we could not bother with this real-time live conversation that we're having now.
This charade, this farce, Keith, this real-time human-being conversation, is this, in a few years or months or perhaps even days, going to appear increasingly superfluous and pointless?
No, because I think our ideas are quite unique and more interesting probably than the bots' ideas. I did listen to the whole 32 minutes, by the way, which I've embedded in this week's newsletter so people can check for themselves. But it's a bit tedious, it's a bit repetitive, and it doesn't really push the edges of thinking. So I would say you and me are more likely to be entertaining than it, at least for now.
Yeah, at least for now, but... These products and companies are moving so fast. And of course, as always, it seems on That Was The Week, we're focusing on AI. This week, again, terrible art, Keith. Is this James Bond and a robot? Is that the name?
Oh, God. I mean, that's even worse than the voices on Notebook LM. The title of this week's newsletter is Agents and Robots. It's definitely... It is definitely happening. What exactly is happening, Keith?
Well, the developments in AI, which we've talked a lot about, the foundation models, this week, most of the announcements coming from the main players were about what they call agentic or agent-based AI. where they can create custom agents, if you will, specialized to a set of tasks, and actually carry out the tasks, like a sequence of events to achieve a goal, if you will. Satya Nadella did a whole day in London demonstrating Microsoft platform latest capabilities in that area, and that video's in the newsletter. Anthropic made the announcement that Its clawed AI can now act as a user of a computer, doing everything a human would do with making decisions.
Although that didn't impress that many people, the anthropic application, so far. I mean, people see potential, but it's still very early, isn't it, Keith?
It's all very early, but the fact that so much happened in a single week from so many different players implies this is a significant part of the next stage, if you will. And we should take notice because this is where the conversations we've been having in the last few weeks about how SaaS and the cloud is threatened by AI starts to come to life and be real. And so it is a a fairly major thing of course it leads to both good things and bad things most of the most of the um curated essays this week focus on the good things but there's one in particular that uh that focuses on uh the downside it basically talks about ai agents being trained to take human vulnerability in terms of our ability to be emotionally tweaked and persuaded of things in the way that advertising tries to do or politicians try to do. And that millions of agents are going to be set loose, focused on that as the end goal. And we probably won't know their agents. We'll think they're human. And so this whole, if you will, brainwashing attempt that you see in the election this week with Harris saying Trump's a fascist, which clearly is not. And with Trump saying immigrants are coming out of asylum and prisons, which clearly they mostly are not. You know, that attempt to reduce everything down to memes and get us to believe in things, the fear is that agents will do that on steroids.
Yeah, although I wonder, given that's already happening with characters like Trump, whether it will actually make a great deal of difference. You talked about it being too early. I always think in Silicon Valley, everyone always talks about it being too early until it's too late and then no one can invest and the whole market's been wrapped up and whether there is a middle period. You talk about something called agentic AI in your editorial. You borrow the term, I think, from Andrew Maynard. What exactly is that, agentic social AI?
Well, probably most people are already familiar with bots. A bot is like an automated social media poster. Agentic AI is the next generation of that where you could ask an agent, to become an ex-subscriber, it could know a lot about you personally. And that would be true across millions of people. And you could tell it that its goal is to get you to vote for, let's just say, Trump for the sake of argument. And it would look at your history and try to create arguments in favor of Trump that appeal to you due to who you are. And it could do that at great scale across millions of people. which a bot can't a bot's just delivering the same message to everyone and it's pretty dumb so so agentic ai i i used a bad example a good example would be um it could you could set it off looking for criminals and it would look at everyone and
profile them and well that's a good one i mean that's just even more paranoia who knows who a criminal is i remember in um In Berkeley, when my kids were growing up, there were parents who would run algorithms to find out whether there were child molesters on their street. So I'm not even sure that one's a good.
But you could also, you could say, I want to do the shopping for this recipe using Instacart and have it delivered tomorrow morning and it could figure it all out and you'd get a delivery and you didn't have to do anything.
That doesn't sound particularly revolutionary. Anyway, what is revolutionary, I think, is Notebook LM. You ran this week's newsletter through... the LM algorithm and platform. And you came up with some dominant themes. We've already talked about the rise of agentic AI. A couple of other themes. What is a robotics renaissance? I didn't know that robotics ever went away, Keith.
KEITH STONE- Robotics hasn't gone away. It's been around a long time and clearly made some progress in fixed installations like production lines, for example. Making cars is a good example. The renaissance was the word picked by Notebook LM. I think, like you, I probably would push back on that word. I think it's more a transformation than a renaissance. It's when human-like, you know, two legs, two arms, a head and a brain-style robots start being able to carry out agentic-like tasks, many of them.
And we're already beginning... I wonder whether both Trump and Harris in their own way are really examples of these kind of digital Manchurian candidates. They both seem automated in their own different ways.
Well, I think that's always been true. That's what the advisors and the media people around each candidate are for. They train the candidate. against a set of learned messages to have a certain emotional impact on a certain demographic. So there's always a massive amount of control over these candidates. I think what we're going to find is that the ability to extend from the candidate to direct engagement with each of us as individuals is the next stage. And then physical interactions are what robots are about. That's a slightly different set of things, and I don't think politics is really the zone for that. I think military might be. Yeah, well,
certainly, and we'll do many shows we have done and we will do on the impact of robotics on war and also, of course, on loneliness. Sherry Turkle has written extensively on this. Another of the...
You know, it's been knocked down significantly. both because of Tesla-related issues, like its sales dropped short last quarter by a bit, but also by Musk hatred. That clearly does have some kind of... And you're not a Musk hater, are you? I think Musk, you can love him and hate him at the same time, depending on which part of him you're thinking about.
Yeah, the editorial this week focuses in part on that. Maybe... The markets woke up to the fact that you just bought a Tesla, Keith. Is it possible that the price went up $150 billion in a day because of you?
We'll have to run that through Notebook LM. And then the fourth dominant theme, which is really part, I guess, of Tesla's trajectory, is what we call... Elon Musk's trajectory, Musk's influence and conflicts, which are always newsworthy, but particularly newsworthy this week. What happened?
Well, firstly, the New York Times published a piece talking about if Trump wins and Musk is actually given the job predestined for him of rebuilding America, if you will, efficiency is the key word that's used.
I think it's both of them. And they're really talking about making America more efficient. And the New York Times points out how many conflicts Musk would have in doing that, given all of his business interconnections with the U.S.
Well, I think for two reasons, yes. The first is, I don't think Trump's going to win. We could debate that. But even if he does... You don't think he's going to win? And then the second thing is, I don't think he'll appoint Musk.
I think it's all... But that's another question. So that's where you and I strongly disagree. So we're 10 days before the election. What are we going to bet, Keith? I bet you that Trump will win.
I wouldn't bet a lot because I think it's super close.
Words and timings
Iwouldn'tbetalotbecauseIthinkit'ssuperclose.
Speaker 2
How about your Tesla? It's probably not worth much. No, I'm definitely not going to bet my Tesla. How about our wives or our children? Children's a better deal.
yeah i think i'll stay clear you can have my youngest that actually isn't a bad exchange i'm convinced but that's maybe we can talk about that next week just before the election um so what else so i agree with you on the second point i mean if trump is elected which i think he will be i think that I'm sure there'll be a very early honeymoon and then something will happen and one of them will be annoyed and the other one will just get out of the game, pick up their marbles and leave.
Yeah, I think that's very lucky. I don't know if you noticed, but this morning SpaceX landed the astronauts that were stranded on the International Space Station that Boeing had left up there. And Boeing is now talking about selling off its space business.
It's astonishing. Yeah, I mean, I've made it clear many times on this show. I'm not sure I'm a hardcore Musk hater, but I certainly don't like the guy. I'm not as ambivalent as you are. On the other hand, I take your point, Keith, in your good editorial this week, that he is a remarkable innovator on so many levels. So there is this contradiction between Musk as this ultimate annoying jerk on the one hand and on the other hand this it's just astonishing i'm guessing in an age of social media probably edison would have appeared in the same way or or carnegie or certainly henry ford there is something about these remarkable innovators that all the also make them really dislikable
well i don't dislike him i uh you know at If you think about, if you separate out what motivates him, it really, there's nothing very offensive other than his views about immigrants, I think. And that's the one area where I really object.
How did you enjoy the play, Mrs. Lincoln? I mean, that's enough to make him profoundly dislikable. Yeah. And also, I mean, his shameless embrace of Trump. I mean, Trump is many things, but he's not a... he's not someone who's sympathetic to Musk's view of the world. So it's clearly just an opportunistic move from Musk. I mean, he could have just as easily tied his money and his brand to Harris. It might've been much more useful.
Yeah, I mean, they can't. I mean, he's not only the world's most, the richest man, which obviously is significant, but also increasingly the most powerful man. In some ways, he might be more powerful than the president.
Yeah, I don't know if you saw, but there's rumors this week that he's been talking to Putin about Yeah. Putin got him to agree not to turn on Starlink in Taiwan at the request of the Chinese premier. I think it's all BS, but... Well,
I wouldn't... I mean, it does point... You've talked about this, about the dwindling of the nation state and the rise of these new agents, these new super innovators, multi-trillionaire types like Musk. Other key... Facts and insights with our help from our friends at Notebook LM.
The future of LLMs, AI agents in the workplace. Keith, what else is interesting? EB dominance, Stripe embraces crypto. They've laid it all out for us. We don't need to do it ourselves.
Exactly. So these are not your pet topics, are they? But Marc Andreessen's piece that the future of LLMs focuses on is saying that You know, the unit of cost in LLMs is called a token. And he posits that the price of a token is trending towards zero. And so that ultimately LLMs will be free. And that owning an LLM model, therefore, won't be a differentiator.
But you can download the 70 billion parameter Facebook model called Lama 3 point something. And you can run it on your Mac with an M2 chip or above. And you can basically get your own version of open AI that's perfectly capable of doing all kinds of things, all on your own machine at home. But you do have to pay for it, don't you? No, it's open source. You just run it on your machine. You've got to be a bit technical to set it up, but there's YouTube videos taking you through the steps. And then you've got your own, literally within your house, you've got your own version of open AI. It's kind of interesting that it made me think there's probably a business in producing a home AI where it wouldn't be a robot, just be a little box and you put it on your home network and that becomes your AI and it's fully distributed and free, basically.
I don't think we've had a, and maybe we never will, but there hasn't been a Netscape moment where suddenly all this stuff becomes accessible to ordinary users. Even you've acknowledged that actually putting Notebook LLM using this show and your and my voices and faces is quite a lot of work. But companies are all working on fixing that.
Except, and this is just this week, but Apple released iOS 18.2 to developers like me. And if I do this, ask ChatGPT what the best season is for oysters in Inverness, California.
No, no, because you're talking directly to OpenAI and the quality of the response is fantastic. So you've really got in your pocket now, access to all the world's intelligence in the form of questions and answers, including very complicated, hard questions.
You've got oysters on the mind, Keith, in Inverness, which are delicious, making me hungry. A couple of other interesting key facts and insights from this week, at least according to Notebook LM. EV dominance. I always get two messages on EV dominance. On the one hand, you always get these articles suggesting that consumers don't want EVs. So the prices are coming down. On the other hand, inevitably, in the long run, as Keynes would say, we're all dead. So everyone who says, well, in the end, EVs are going to win. Who knows? What's the news on EV dominance?
So this is a really, really good essay by Noah Smith, who has a newsletter called No Opinion. which is the detailed look at the rise of EVs and the likely timeline to their ultimate success.
And I know in Norway there are more EVs now than gas cars.
Words and timings
AndIknowinNorwaytherearemoreEVsnowthangascars.
Speaker 1
So America's a little bit behind, surprisingly, which is, I think, more to do with Americans clinging on to their Ford Mustangs and the like. But it is completely clear that that evs will not just win but take over and even government policy is is uh focused on that happening by a certain day that day keeps slipping of course yeah i
mean and notebook lm have been kind enough to give us some notable quotes the quote from smith on evs so evs are just a superior technology to combustion cars they're soon to be cheaper they're more convenient they're much easier to maintain They're more powerful and far more silent. They will mostly replace combustion vehicles. It's just a matter of when. Well, it's not just a matter, because that could be 100 years, 10 years, or 1,000 years.
traditional cars so that's another interesting it isn't in the newsletter but waymo raised another five to six billion dollars this week to and that's another uh well
i guess google or google yeah alphabet company so uh and another let's since we've got the notable quotes here here we've got um Anthropic on its computer-used API. Available today on the API, developers can direct Claude to use computers the way people do, by looking at a screen, moving a cursor, clicking buttons, and typing text.
We already talked a little bit about this, but is it true, Keith? I haven't used it, to be honest, Andrew, because I'm not sure I have a use case for that yet. So I haven't played with it, but I think it's true. I've seen some YouTube videos of other people doing it. So it looks true. And you can imagine, you know, you and me are both old enough to remember, I don't know if you're techie enough to remember this, but batch files. Batch files were something that DOS used before Windows. And you could program a sequence of tasks in a text file. And then you would just call the text file and it would carry out the sequence of tasks. So to some extent, there's something dumber and not very new about this. Where it gets really interesting is when you just give it, you know, book me a United flight to New York on December 3rd, leaving around noon and make sure it doesn't cost more than $500. And it just does the whole thing for you. That's the interesting bit. And I think we're getting to that.
Just to go back to the Maynard piece on agentic AI, the notable quote that Notebook LM added from him was, all it needs is AI models that are capable of influencing people to achieve specific goals by utilizing patterns and associations between behavioral cause and effect that are extricable from zettabytes of human-centric data. Isn't that basically any different from advertising, from the marketing industry? from the thinking of marketing pioneers in the 20th century?
I'd say advertising does that at an abstract level, and it puts you into a demographic bucket where maybe there's 100 buckets and you're in five of them. But after that, it's pretty much the targeting is poor at best. You've probably seen ads on Google to buy things that you've just bought, for example.
Now, what this is, as we discussed earlier, a whole level beyond that. This is a real-time thinking agent, if you will, that's able to look at you and your history of posts across... or within contexts, and specifically look for behavioral cause and effect on you personally that would only work on you? Is this chilling or good?
it only gives us what we want, but it knows us, as Eric Schmidt once famously promised incorrectly about Google, it knows us better than we know ourselves.
Yeah, well, this really depends on whether it's working for you or working for somebody that wants to influence you. And then the second thing is how independent-minded are you? I mean, I don't know about you, but... I'm not independent-minded. I only do what you do or say. Yeah, of course. You're the most follow-the-crowd person I know.
Yeah, I'm crowdsourced. And the final, one of your pet themes, Keith, is... and I agree with you actually mostly on this, is how Europe is lagging behind the US when it comes to innovation. The final quote is from Guillaume Ronstang, sounds like a fictional character, on Europe's tech scene. To them, Europe is akin to an emerging market, kind of like your little cousin doing antics on Christmas Day. Mildly interesting, but you can't wait for them to graduate to the adult table. Is Europe ever going to graduate to the adult table, Keith?
I think it's been there and left, hasn't it? He's slightly wrong in characterizing it as a young cousin. It's more like an aging senior that has got dementia. Not to be cruel, because we're all going to end up there at some point. But this story is actually about an AI company called 11X that raised money, but is moving from Europe to the United States, which I put in because it's a trend. My talk in Brussels recently was all about Europe's challenges and why entrepreneurs choose to leave and come to the States. And this is just a fantastic current example of that. So it reinforces the things we've talked about in previous weeks. And it's just more evidence to that is sad, actually, in a way.
I mentioned Maynard Keynes earlier, John Maynard Keynes, the great 20th century economist who famously said, in the long run, we're all dead. Our video of the week is one of mine. You've been very generous to put it in. I interviewed Robert Skidowski, who's the most authoritative scholar, writer on Keynes. He's the author of three-volume biography of Keynes, as well as a one-volume bestseller. He has a new book out. It's called Mindless, the Human Condition in the Age of Artificial Intelligence. The conversation I thought was most interesting when we talked about Keynes. Is there such a thing, Keith, do you think, as the human condition in the age of AI? I'm always a bit skeptical of that.
I'm with you on being skeptical. I actually think Contrary to popular opinion, especially elite opinion, human beings are incredibly smart and intelligent and very, very difficult to persuade to change their ideas. Just like you and I, right? Just like you and I. So the idea that... I haven't read the book, so tell me if I'm reading too much into the title. But Mindless, I think, is like what... what Hillary Clinton talked about, the deplorables, is this elite idea that people are just easy to program, if you will. I don't agree with that at all. I think people are determinedly self-interested based on their experience of life and really, really hard to brainwash.
I think if that's true, then you're wrong about Well, maybe you're right about Clinton. I mean, if that's true, maybe you're right. Who is more likely to win the election? Who's the bigger brainwasher, Harris or Trump?
Well, it may seem like a weird answer, but I think Trump tells it like he sees it, which is, you know, there's too many foreigners coming across the border and coming into American towns and cities. He exaggerates, of course, but that probably aligns somewhat with people's real-world experience. He talks about the establishment not caring about ordinary people, which, given that ordinary people are voting for him mostly, it's not rich people. There's probably some truth in that. And so I think he's like an accidental clown who's captured something that's already in the air
Yeah, it's also right on foreigners because I'm not an illegal immigrant, but I'm an immigrant. And I have to admit, I eat a lot of my neighbor's pets. I have a barbecue, the cats, the dogs. The turtles, the goldfish. We enjoy eating our neighbor's pets, so I'm not alone on that.
But funny as that is, Andrew, none of us believe his BS, right? But we do actually think he's keying into something real, which is why he ends up being more popular than he should be.
Yeah, well, the other thing is, is that, yeah, as you say, he's sort of weirdly more honest than she is in the sense that he exaggerates the truth and she just basically avoids even addressing it. Anyway, we'll do it. Let's do a show next week, Keith, on the election, because by next weekend, that's all anyone will be able to think about. Let's move on to your startup of the week, which I think in many ways, is in keeping with the themes of agents and robots. It's a machine learning for healthcare documents. T E N T E N N R tenor. I'm guessing is this a, why'd you choose this one?
Well, cause I think when addressing this move towards agents, it's important to look at the human, uh, improvement that can happen because of them, as opposed to this idea that we're all children and we're going to get brainwashed. So I put this in because it's a great example of a good use case. We all know that health documents is one of the
I mean, yeah, the healthcare system, and maybe I'm about to make a kind of neoliberal argument here, but because it's a kind of monopoly, they're so inefficient. One of the things that's most astonishing to me as, for better or worse, like everybody in America, a user of American healthcare, is the only time one ever either gets a telephone call or has to make a telephone call is to your doctor because they just don't care. They won't do email. They don't do text. There's no competition.
Yeah, and the industry, the healthcare industry is so, I don't know whether it's prone or vulnerable or open to innovation, but at some point, whatever Trump or Harris does or doesn't say about healthcare, all this new technology must inevitably change it, make it more efficient, won't it?
Yeah, well... maybe that's an idea for another election. Um, so, uh, tenor is your startup of the week. And finally your post to the week and you took it off sub stack, not, you know, what from somebody called Andre Vura cast. What does a Vura cast say this week?
This is very funny. Uh, know i i work with data in my day job and he's talking about statistical modeling inferences and designing experiments which none of which require machine learning they basically require skills in statistics and and heuristics and uh he's basically saying if you uh if if if you really want to understand data and do machine learning what data scientists do with LLMs is not even in the same domain as what you would do with data. And he's distinguishing between the two. I thought it was great because that distinction is lost on most people.
A machine learning engineer is a term that encapsulates quite a lot of different jobs. But there's data preparation. There's data querying. There's data platform management and organization. There's all kinds of stuff to do with data. well before you do machine learning on the data and the whole soup to nuts process of data up to machine learning is about 10 different jobs and they've all got titles and names none of which are data scientists that's his point and this ai
Yes, it will, but it's not there yet. In fact, it's weird, but the one part of the AI ecosystem that is still highly manual is data. Learning from data is what models do, and machine learning acts on data. But most of the work to get to the point where you can even start doing those things is still highly manual.
Well, maybe it will be fixed by next week, by November of 2024. When we will be back, we'll do a special. I know you're not keen on talking about politics, Keith, because you do the Gilmore Gang, which is a more political broadcast. But I think next week, we can't avoid talking about the election, can we?