Latest / The Bleeding Edge Podcast / Astra 6, Consumer Agents, Responsible AI, Github Hydra
Transcript
- Ralph: I'm your host, Ralph Benker, and with me in the studio, in his bedroom, somewhere in London, Em: Yeah. Ralph: is Emil Augier. Emil, welcome to the show. Em: Great to be back, Ralph got the momentum going again. Let's get into it. Ralph: Nice. Well, let's get let's get into it. So we've actually been very excitedly talking about Astra, which is OpenAI's new model, GPT six Astra. And the big news this week is that Jensen Huang has put a bit of a spin out there and he has hailed it as the start of AGI. and I believe on X. I don't really spend much time on X. So Emil, you'll have to confirm this one. He congratulated OpenAI of heralding in this new era. Now of course a little bit of tongue in cheek really in terms of trusting what Jensen is saying that's not taking away from the capability of the model but it was trained on a hundred thousand Nvidia chips and apparently it cost about a billion dollars to train the model Em: Wow. Ralph: they actually have and that was on the blackwell chips and they have the next version Astra 7 is already in the making and that is reportedly going to be trained on 400,000 Vera Rubin chips and so we're talking about just the next level of capability and if you look at how the market has been responding to Astra six then I think you'll agree with me. Things are getting very interesting indeed. So we'll talk a little bit about that. but what do you make of Astra six, Emil? Em: I'm loving it. Ralph I can only speak for personal experience on homeland projects, but I'm just finding the token allotments incredibly generous and the results so far seem really smooth on Astra. So more to update you on next week as I dig into it deeper. Ralph: Well, we'll talk about some of the use cases that the creators and makers out there have been improving. And it's a pretty interesting blend, really. we've also got some really interesting news that is continuing from the guys at Meta. so they've basically obviously released released this Muse model and they're now really packaging it up. as your personal AI agent. So there's definitely a move here from Zuckerberg to really kind of take the consumer space. Obviously Anthropic have been pushing very hard on enterprise, Chat GPT, OpenAI, they're just pushing on every door. but Zuckerberg has obviously had huge success with the likes of Instagram, WhatsApp. and Facebook and so it completely makes sense that he's going down the route to actually open up a consumer personal AI agent. And the difference this time is this thing's got your credit card. and bowl accounts also incredibly generous. There are two tiers. so there's obviously a very low cost accessible tier which is around twenty dollars and then there's a hundred dollar tier so exactly what you can do with that we'll be putting through the paces as well on the show. but the big question is will consumers trust it? do consumers trust AI at all? in and I guess in some other news, one of the reasons why people are kind of questioning where all of this is going is we have a 27-year-old ex-openAI and ex-anthropic pre-training researcher. And he just left his job at Anthropic a couple of weeks before his. Shares vested and he's called out the risk that AI represents. And so one of the things we have to understand before we get these nicely packaged, very friendly models, there are hundreds of researchers out there that are working on the bleeding edge of AI trying to tame these beasts. And by all accounts, they are terrified by the capability that they are having to to train. And so hence we've seen you know some very interesting folks leaving their AI labs and and writing some very interesting essays, basically warning everybody of the dangers. in the same week, we've also had Sequoia, which is a a very large VC and they have told their portfolio of 80 founders or a portfolio of 80 founders. That they need to stop renting intelligence and start owning it. And I think this is a really interesting moment because when you have capability like Fable 5.1, Astra 6, and everything that is coming out of the Chinese labs, it's very difficult to see how you can start building your own AI. but of course, as we've discussed on the show, there are tools like Inkling from Mira Marathi's thinking machines that actually allow you to start building your own capability. and I actually think that that is definitely one of the models that we're going to see increase as we look at founders trying to build a moat, and indeed any company trying to build a moat. I did also have one other kind of news point that was just something that came out in the in the last week. Alex Karp. He's the the CEO of what's the company, the big one, the one that we're all terrified of. Em: And through quick? Ralph: no. No, we we like that one. the one that's watching us, that's got control over all the satellites and Yeah. Em: Sky Net Ralph: I'm gonna have to Google him now. why can't I remember Em: G Ralph: it? Em: go on, find it. Ralph: Alex Carp from Palantir has basically Em: Of course. Ralph: called out if you're not an AI company, you know, then you are doomed. So you better become an AI company. So, and this is really being echoed by many big investors and and folks that are on a variety of podcasts and YouTube channels. and essentially, you know, the die is cost for us. If you don't have an AI core, then I think the writing's on the wall. just because you know so much of what we're gonna be executing will be supported by AI. and if you don't believe that that is coming, then you probably aren't going to be working in the next five years, or even sooner, who knows. But I just you know quite interesting to to see some of the sentiment. around AI. we still have a lot of the doom and gloom folks out there. So there's a few reports recently just about how AI is gonna kill us. And I think there have been 25 different ways that various groups have now completed research on how AI is going to mean the end of mankind in the next few years. I'm not gonna subscribe to all of that at the moment because I'm just seeing too much excitement. and really positive productivity coming out of a lot of the solutions that I'm working on certainly. And I realise that it's pretty scary at the cold face when you are working on the researching of these tools. but my hope is that we will get somewhere in terms of the alignment of AI with human goals and and so I'm just gonna hail this as a wonderful moment with Astros six really being called out by Jensen Huang as having delivered AGI. Now, actually a lot of people out there, you know, if you kind of the the folks in the know, they will say well we've had AGI since around 2020. and I guess it depends on your definition of AGI, but a little bit of a populist comment there from Jensen, but we'll we'll take it because it's a very special model. So so those are the top stories this week. we'll jump into a couple of other interesting ones as we deep dive to you know you know just get into a little bit more into the detail of those. anything catch your eye, Emile? Em: share your same optimism in AI. You know, we we're enjoying the journey and I don't want to dwell on it too much, but just to add colour and some tr transparency over what's happening with that researcher that's handed in his resignation, by his own account on postings, he has worked for three years as a RL researcher at both Anthropic and OpenAI. And just to be clear, he's not saying he's afraid of current models. Where the concern is actually lying is that internally what they're recognizing, and he's not alone in feeling this, is that the speed of R L to the race to superintelligence is happening faster than they anticipated that will take them to to superintelligence. And I think that's the core aspect to why the story's blowing up. But as always, the Doom Mongerers and specifically the likes of online reporting news agents are taking the story and putting very scary titles around it. So I just wanted to add some shape behind, why we're seeing what we're seeing. And I did have something to share on Astra. another posting that I'll read out directly from Greg Eisenberg. I've seen a hundred plus people post GBT six Astro demos where it builds 3D games, and those are cool, but that's honestly not the part I can't stop thinking about. Millions of people can turn an idea into a printable object, a CAD file, a blender model, or a circuit board without having to become a full-time 3D designer first. how many thousands of one plus million businesses come out of that. Well, you pick a niche that constantly needs small custom parts. A few examples include gyms with broken equipment pieces, restaurants missing some weird proprietary clip, cyclists who want custom mounts, small manufacturers stuck doing the same handwork every day with a jig they wish existed, dentists and clinics needing custom jigs, guides or tool holders. You put up a simple page where someone uploads a photo, adds measurements, and describes the part. Astra can do the first CAD pass for you. You sanity check the dimensions and the material, and the print farm ships it for you. Of course, you need distribution first, but you're in business. You just became a manufacturer, and it's kind of like your factory is a chat window. for those that are interested. You just need to go to ideabrowser.com to find out more as he adds. And the TDL on this is in 2024, lovable bolts, replets turned anyone into a vibe coder. The effects of that are still playing out, but we're seeing basically anyone create websites and applications. In 2026, OpenAI turned anyone into a vibe manufacturer. This shift just started. It's a really big deal and it's under discussed. Ralph: Yeah, and that's a really fundamental shift. Now to some degree I'm kind of going, this is a little bit old news. why am I saying that? Well, Claude Claude Code has been very effective in terms of its ability to drive Blender for quite some time. And Blender is obviously a 3D design tool. Astra Bowl accounts, you know, is able to do the same thing. but it just seems like everybody has somehow caught on to it a little bit. you know, well, they've just kind of caught on to it, whereas with Claude it was just something it did. but I I I don't want to kind of take away from that, but it's definitely something that the developer community has converged on. So some of the things that they've been testing with with Astra is code based cleanup. this is one where I will be putting Astra through the bases. I've got a repo with quite a bit of code in it that definitely needs a bit of a cleanup. So that's gonna be a really good test. It's got quite some lines in there. obviously codex has been Chat GPT's or at least OpenAI's coding solution. And I've used it quite extensively, so I'm really looking forward to seeing what Astra can do with that. 3D reconstruction from reference images, exactly as you were saying, you know, just take a take a photo of anything and 3D print it. So you know, that's just amazing, I think. the other thing is one-shot iOS apps. So not only is anybody a vibe coder, and anybody can build a website, but you can now also build an iOS app. You could do the same thing with Claude by the way. Claude has also got it's actually embedded in the Apple Developer subscription. So and it's as far as I know, there was an option to basically use Claude code for free if you are signed up to the Apple Developer program to build apps. so so that's just another one where I think Chat GPT or at least OpenAR are catching up. reverse engineering hardware protocols. video prep for final cut. and and I think all of that really came about with you know the the community actually putting this thing through its paces in the first 48 hours. And so it kind of gives us a little bit of a different way also of evaluating these models. So on the one hand you you have the the the news cycle So whatever the story is that the Frontier Lab actually releases, then you actually have the benchmarks. and let me tell you, the benchmarks mean almost nothing to most people. but you know, we get lots of fancy graphs. but the thing that really kind of helps us understand these models is when creators and developers and makers out there actually put it through them their paces and by all accounts, the first forty eight hours are quite defining. so that's been another really interesting aspect of this model, just to see how the community's actually benchmarked it. and so we don't just have to believe, you know, what the lab says. We can actually just sit tight for forty eight hours and the community will get to work and tell us what it's really capable of. and so in this case, you know, I think we've kind of got a junior technical labor force across four unrelated domains at once. That's how it's been described. and and I think that's that that's pretty impressive. So what what are your blender skills like, Emile? Em: I don't have any, but you're absolutely right about that this story isn't as current as it sounds. why this is actually getting a lot of attention now is the increased capability and offerings on computer use to people and at a much reduced cost than before. So when you add that capability to what we could do before, it's just become a whole lot more accessible to people. And I think that's why it's trending. Ralph: Yeah, and it probably does give us a good sense of the difference between Anthropic and OpenAI. You know, OpenAI is so consumer focused, I have lost count of their user base, but they were kicking around nine hundred million users. Anthropic obviously less than that, but more focused on high paying enterprise customers. And so the news cycle just runs differently because of different user bases. also in the same week we had Claude Fable five point one, Quen three point eight, Gemini three point eight, flash, Muse Spark one point three. So five model updates across four labs, essentially inside one news cycle. and the real Em: don't forget that Gemini Flash has also had three cycles in three upgrades in two weeks. Ralph: Right. And so which kind of really brings me to the point of that little news snippet. If we thought things were speeding up and going fast before, what is being predicted by the end of the year is essentially daily releases from these Frontier Labs. and that is a really scary prospect. I have no idea how we will cover the news then. especially when you look at the capabilities. and just to illustrate the capabilities. So we were also chatting about this Navier Stokes maths problem that I can't possibly pretend to fully understand. but I think it's been one of these, let's just say, difficult problems, a really hard problem to solve. and essentially it tries to predict whether in some kind of ideal fluid whether if you stir it the right way you can actually create a singularity, essentially a black hole. And Astra 6 was able to prove that it is possible. and this is really interesting for two reasons. Number one, one of the the the podcasts that I was actually listening to where they were getting into this a bit, the chap had emailed a counterpart in OpenAI asking him if he had done any work on this Navio Stokes solution on the 28th of August. and the guy said he hadn't at that point. just a few days later, on the first of September, he basically came back and said, Yeah, okay, you know, we have we we we've started work on this. and and a few days after that. They said they'd solved it. and so the point here is that this is not this is a a generalist model that in a handful of days was basically trained on this Navio Stokes problem, and within a few days it basically was able to solve it. So they ran 10,000 agents. they basically ran those 10,000 agents for 88 hours. They spent something like 130 billion tokens, and it cost around six million. And so you might say, but Ralph, that's a lot of money, and nobody has that kind of compute. But the the real story here is in about six months' time, we are talking about a hundred X in terms of the speed capability. And reduction in cost of similar processing. In a year's time, they're talking about potentially a million times more capability. And so, you know, it can it be that in a year and a half's time, in order to solve a problem of that magnitude, you're looking at six dollars of cost. You know, that is the real story here. And this is what actually the leading researchers are saying, this is what we're moving to. This is why I'm so excited about this moment in time. Em: I just want to add, to what I understand, what's happening around that story, it is part of the Millennium Prize, which to whomever solves it, there there's several of these problems. there might be twelve or so. only ones historically has been solved, and as you called out, despite it being an awarded prize for a million They spent multiple millions solving it, but as that gets cheaper, it's gonna unlock so much potential, especially in the medical field. And another thing I just wanted to add around that was nothing, Rolf. Ralph: great. We can have a quiet moment to ourselves. yeah, no, it's a a a very a very pivotal kind of time and but by all accounts OpenAI have actually solved multiple of these challenges. you know, this Navio Stokes story is just the one that happened to break and capture everybody's imagination. Em: I have remembered what I was gonna say regarding the solving the solution. The singularity is actually about the aspect of tracking and understanding liquid motion and where why this theory is really important. It's used in medicine for blood pressure and it's used for weather forecasting and also some aspect to aerodynamics for planes. and where the model gets complicated as as viscosity increases, liquid stops becoming one single form and starts going into a particle shape, and that separation is a singularity. And I think the theorem can't get past that point. Ralph: Well, you clearly know me, you're you've been looking deep into the the tea leaves. Em: Yeah. Ralph: but you know, the point is if the last guy to do something special here was like Demis Hasabias from Deep Mind and he he figured out how to basically synthesize something around unfolding proteins. So it was a bit of a big deal, right? And he and he won the won the prize for that. and now really the question becomes you know, who's doing the solving? Who wins the prizes? Because at the end of the day, all you need to do is stick it in the prompt. It's this is the same as asking Chat GPT you know, where you should go for dinner tonight. the chips are all falling, so let's see. Let's see where it goes. And I think this is why Alex Carp is really calling out, you know, you need to become an AI company. and I really believe that myself. I think every single business has the opportunity to build an AI digital twin. And indeed the you know, you you should you should be doing that. If you're not doing it, then you you need to start. but anyway, it's gonna take a while for all of that to play out and no doubt we'll have to do a little bit of a deeper dive on exactly how you can do that, Emil. I think that's the big project for the remainder of this year. Em: before we go, a tip to our listeners. this week's practical tip is giving AI agent a definition of done before it starts with agents taking on longer jobs. Looks finished is surprisingly a weak test. For example, try something like build my reading page. It is done when I can mark the story, read, close and reopen the page and see that same marking. Check that on a narrow screen that it works for me and tell me what you tested and what you could not verify. You can use the same idea for everyday research, such as compare three options for me, link the evidence, and flag anything you could not confirm. This turns a vague request into observable result. Anthropic engineering work highlights the value of explicit feature checks and testing the actual user journey. Our practical spin is simple. Describe what success looks like from your side of the screen. it's a small thing, but it helps. Ralph: I love it. Yeah, I Em: I love it. Yeah. Ralph: just always use the outcome I'm looking for, but I think I can refine it a little bit in terms of what done looks like. But thanks for that top tip, Emil, and thanks for all your great stories. We'll see you on the edge. Em: And yours. Adios.