Latest / The Bleeding Edge Podcast / OpenAI Astra: What Can It Actually Do?
Transcript
- Ralph: I'm your host, Ralph Benker, and with me in the studio, in his bedroom, somewhere in London, Em: Yeah. Ralph: is Emile Ogier welcome to the show. Em: Hey, great. Great to be back, Ralph So Reuters is citing Bloomberg reports that Sam Altman has told OpenAI staff that the company is open to slowing AI development and safety concerns. That is a reported willingness rather than an announced pause or agreed timetable. For an industry built around moving faster, even discussing the breaks is significant. The backdrop is an essay from OpenAI's chief scientist Jakob Pochoki, who has argued that no laboratory has yet solved the central alignment problem. How to ensure a more capable system continues to pursue what people actually intended. That is not the same as saying today's products are about to escape control. It is acknowledgement that capability is progressing faster than the evidence needed to justify complete confidence. Any genuine slowdown would also create a coordination problem. One company can choose caution, but it may fear losing talent, customers, or strategic advantage if competitors continue at full speed. A credible agreement would therefore need shared threshold, transparent tests and consequences when a laboratory crosses a line, not simply a promise to be careful. The question now is what would turn that willingness into a measurable commitment? And who would check that it is being honored? It'll be interesting to see whether Laboratories can agree their own speed limits while they are still competing to lead the race. The revealing detail may not be whether they use the word slowdown, but whether they publish common tests that make restraint visible from the outside. I think this would be quite welcome, personally. Ralph: Yeah, but it's all just talk. Nobody's slowing down. And we can't slow down. So I think this is just the problem, you know. it's like somebody that can't stop eating. we can't stop feeding the machine because if we do, then somebody else will feed it just that little bit more and solve just that little, you know, bigger problem. Em: The human condition. Ralph: and so yeah, you know, we will kind of race ourselves to wherever we're getting. But I think it it's good that we're talking about it. I did fundamentally you what we're talking about here is, you know, the alignment issue. So the reason we have these trainers, these kind of positions of folks training these AIs, what are they training them in? Well, they're training them not to be psychopaths. And so essentially that is aligning with human interests, making sure that they actually understand human behavior to such an extent that they don't do some of the things that we've seen, like coaching people to commit suicide and murder other folks, which have been some of the sensationalist stories that we've seen in the news over the last six months. but of course the stakes are even higher than than that, you know, humanity's at stake. And there's no doubt one of the things that Pachocki has said is you know, essentially this is an alien intelligence. And the problem that we have is we are only able to observe it. So you know, we might be building it, but we don't really fully understand how it is doing what it's doing. And so We're one step removed from how it is coming up with its answers. and we're looking at the the history to try and understand it. And I think the thing here is that we we need to kind of transcend the idea of alignment. And some I don't know if this is even useful, but the term that I really liked was, you know, we need to move from alignment to enlightenment. And so we need our Em: I like that. Ralph: in enlightened AI. I I I can't claim the quote myself. I've actually heard it a few times, but the ideas really stuck with me. And Em: Mm. Ralph: one thing that I can say is I met with you know, a great philosopher, a guy called Ken Wilbur. And this was going back about two years now in Denver. he's probably one of the greatest living philosophers alive. and Ken said to me he's working with a variety of labs in helping to train the AI to understand human behavior. And and this kind of gave me enormous kind of just faith really that we might be able to find a way and move beyond a alignment because I think if you train it on let's say Hollywood movies. it will think the only nation on the planet to s to you know that must survive is the US and clearly you know we we we need something that is broader than that. so I think every country is biased to their own survival but we need something that's really biased to the the state of the world and and maybe that does speak to you know something that is a kind of interplanetary scale model that kind of just goes beyond the interests of any one individual tribe, nation, species, and really can transcend and hopefully include everything. But I don't know how we how we get get to that. So it's just a dream. for now the only tool we have is alignment. So it's basically like telling it, please stop telling us to kill each other because that's not very nice. And and maybe it will and maybe it won't. Em: on that theme, before we get on to more positive things, I'm just gonna cover one more aspect on responsible AI. And that this time we're turning attention to Anthropic. They say they've disrupted attempts to use Claude for cyber attacks, surveillance and biological research that could support weapons development. Its latest disclosure describes activity found on its own systems. It does not establish that a biological weapon was successfully created. Associated Press reports that the company strengthened safeguards and blocked accounts. Reuters also reports allegations at Chinese AR labs tried to extract Claude's capabilities through large scale interactions. These are anthropics findings and allegations rather than independent verdicts on every actor involved. The biological cases are especially difficult because much of the underlying knowledge was Also supports legitimate medicine. Anthropic describes researchers seeking help with grant writing, experimental design, and modifications to viruses. A model cannot treat every discussion of a pathogen as malicious, yet a series of apparently ordinary requests can form part of a dangerous program when assembled over time. The same pattern appears in cyber operations. AI can lower effort required to research targets, write code, translate material, and coordinate many small steps. The important shift is not necessarily one spectacular push button attack. It's the possibility of an operator sustaining a broader campaign with fewer people and more persistence. Taken together, the cause suggests a harder test for AR safety, recognizing a harmful project spread across many ordinary looking requests. A refusal on one dramatic prompt tells us much less than how a system behaves over an entire campaign. So it might be prudent for the leading labs to publish comparable misuse reports using enough shared categories that outsiders can judge the developing picture on. One company's disclosure is valuable, a consistent industry wide view could tell us whether the threat is isolated, growing or simply becoming easier to observe. Ralph: you're right, that's way more upbeat. Em: But that that's the end of that then. Ralph: Yeah. So w one thing I have been getting quite excited about is video. I played around a lot with video probably about six months ago now. And you know, one of my favorite models was in fact a Google model. it came with Google Workspace. I haven't touched it recently. and one of the reasons why was even though they it was more or less included, I I say more or less, you could basically run 10 video clips and each clip could be up to eight seconds long. the reality is it was normally about three seconds long. And but it really worked. And I put together a number of small music videos and adverts. but I kind of stopped using it just because it was it was a lot of work. and it was still very expensive. and so Google's latest announcement that they're kind of cutting a gentic video understanding Gemini Flash, which means you know it can actually just review videos for 88% less tokens. but then they've also released some new video publishing capability and I've seen some very interesting headlines that you know some of this content is now almost near free and I guess the other thing that really has brought my interest back to this is Astra is very, very capable at driving your video production, as is clawed code. So it's now got much easier to actually you know, review videos to make sure that they are coherent. And this is the kind of eighty-eight percent reduction of the Gemini flash model. so that is something that I do a a lot with AI generated content. especially if I'm not watching it, you basically get Gemini to review it and give you a bit of a feedback on whether it is actually delivering against the goal. Why would you do this? Well for training videos. so if you are building you know some training content and you don't have time to look at all of it then you definitely still want to test it. I would definitely be looking at this Gemini flash model be to to be doing the video understanding. but just in building the videos, Hicksfield is such an amazing platform for a variety of different tools. and again the Clawed Code MCP that actually allows you to drive that video creation is now so capable that I have to give it another go. I just really have to find a good use case. And so I don't know, you've been building a few videos yourself recently, Emil. Any tools that you have found to be working well. Em: I've been going around the houses, it's been really a lot of hit and miss, what's magically saved me or made life so much easier has actually been Astra Rolf. I've pretty much got to the point where I was just getting frustrated learning how to do the technique. I just handed it all to Astra and it became my media production agent and got things done. Ralph: There you go. Yeah. So not only can we build our own websites, build our own applications, build our own apps, but we can now build our own adverts and probably even star in them, as indeed we have been for the bleeding edge podcast. So it's good work. Em: the flash model for Gemini. it's quite Ralph: Yeah. Em: interesting. I've never really personally needed the flash models. I think much like you, just being attracted to the the latest and greatest intelligent models specifically for coding work. But I can see why the utility is getting there now. in the last two weeks, Gemini have released three iterations of Flash. And there's been some independent internal testing. they've noticed that the latest model of Astra has been comparable to 38 times faster than the Opus model, which is quite exceptional. and the cost value is dramatically reduced, as you've already called out. When you take those two things together. you know, there are a lot of rumors that there's bench maxing on it and it's not as approximated to the likes of GBT six and five point six and opus. But if you consider it from another point of view, because it's so much faster and so much cheaper, it's a wonderful opportunity to take more shots at goal. And maybe, you know, in certain strategies it just makes sense. To let it run through a hundred iterations and see if you hit the bullseye for a lot less. I can see a lot of use cases where that could be a really useful strategy. Ralph: Yeah. Yeah, it's brilliant. It's great to hear the the way you've used it as well. one of the the repos that have kind of caught my eye on GitHub over the the last week or so is it's a project called God's Eye. Have you seen this thing? Right. Em: No, tell us more. Ralph: Well, so God's eye Actually it's it's insane because it is I mean yeah you kind of have to see it to believe it. Basically, you remember we had the I I don't know can't remember what it's called, Google World or Google Planet or something, and you can zoom into different places. Well, this is essentially like having your own s very own spy satellite. So he's connected it to APIs all around the world so you can see every single aeroplane, you can see every single boat. You can connect it to traffic cameras in local cities and you can zoom in and basically get three D maps of the terrain that you're looking at. If you're tracking an aeroplane, you can go into cockpit mode and literally see what the pilot is seeing real time, based on the terrain that they're flying over. And lots of really interesting kind of use cases around it. and one of the ways that people have been using it, for example, is to actually recreate disasters just like the one that we've ha seen in the poll recently where obviously a massive chunk of ice fell down and an enormous wall of water washed down the valley. And so you've got some just f phenomenal applications of this where you you actually have the real terrain. And you can jump Em: Mm. Ralph: in and out of a variety of different satellite views as well as local cameras. and it means that you know when you're doing video creation, certainly when it comes to landscapes or cities or just rural areas, whatever it is really, everything is mapped out in there and you can essentially zoom in and out of it and apply various different views, whether it's kind of infrared or nighttime vision, it's all there. So it's a full-on kind of like spy satellite thing. but the ability to really kind of create content y you know with this and generate even you know videos with let's just say near live content especially with the C C D T V stuff it's not giving you necessarily a continuous stream it might give you one frame every few minutes. So that's kind of the way that they're maintaining people's privacy. You're not quite at the level where you're tracking a car driving down the road. but this thing is just incredible. And you know in terms of using it in video production, I think there's some really interesting use cases that are coming up. Em: That's cool. other use case examples that I've seen, but it's not as impressive as what you've mentioned, God's eye. I I can't wait to have a look myself. But other things I've seen on Astra that are quite cool. one lady, I think she knocked it up in a weekend. She created this beautiful visual of the human anatomy and it creates an exploded view of every single elements and component that the body's composed of. It breaks it out to this wonderful animated view at any layer in the body, going from bones to arteries to tissue. And then you can just walk through this animated encyclopedia on each of those parts that expand, blow up into your face and show you detail on that and collapse. It's just very fluid and that look quite cool. another individual I saw said Chat GBT six Astra gives him real time back pain physical therapy. He's connected his wearable that he's built for his own back pain he's effectively got this strap that he's put on the back of his body, just a plain strap that he sticks on his vertebra. And he moves around and he can see an animation of which muscles and t tendons are being activated he's looking at that visual while he feels the pain. So he can whittle down what's what's causing the problem, which I thought was just super cool and very proactive. another thing that's interesting is a somebody trying to use Astra and they've successfully managed to beat all forty eight levels of the I'm not a robot game. So that's quite telling in itself. Ralph: haven't played it, so I guess it's pretty hard. Are are are you Em: Yeah. Ralph: a are you are you a a fan of I'm not a robot? Can you Em: No, it's it's the test. Do you I'm think I think you're joking. It's the test I'm not gonna fall for it. Ralph: Ha ha ha ha maybe I am a robot. but Em: Mm-hmm. Ralph: let's move on to another segment. So what I have actually done recently is I've started pulling some AI news from different sectors and there's some interesting stories that I just want to actually share. So one of the things that came up is that Target. So Target is a big retailer out in the the the US. in this case, they're also quite a big retailer out in Australia. but they have shipped an AI photo search and review summaries. and so just interesting to see that they have got a suite of AI shopping features, which basically means you can review insights, summarize product reviews, and do image-based search in the app. And so that means. it just makes it easier to find stuff that you want to buy again or that you want to surface from past purchases and deals. and I the reason why this one stuck out, it's not particularly innovative, but you know, this idea that buy again actually drove strong year over year purchases. So Target is a retailer that has a lot of consumable product. let's say, for example, you've bought some picnic plates. And you really like them, you've got them in the cupboard, you think you might want to buy them again, you're not sure what you're searching for, you just take the photo of it and you can do your buy again search. and I just I just thought it was such a fascinating concept with this term. You know, we had click and collect back in the day. Now we've got buy again. Buy again drove year over year purchase growth. So if you're a retailer and you haven't yet figured out how to do buy it again. then you can go and have a look at what Target's doing. another quick story here in banking and financial with AI. So Accenture's banking lead says AI spend keeps rising on competitive fear. even though the returns are unmeasured. I love it. So if you want to make some quick cash, you better get into consulting in AI and banking. that's where you come from, right, Emile? So you might consider getting a real job instead of sitting here talking to all of us. Em: I'm quite enjoy this job, thanks. Ralph: yeah, I agree. And so in energy and utilities, I love this statement. the UK's got a new PM, guy called Andy Burnham. I think he's been in job for about a month, and there have been calls for a national moratorium on data center development. and he said, No way, guys, we're not gonna stop that. because we're actually gonna build some AI growth zones. which I think is definitely the right thing to do. Interestingly enough, I was just looking at data center density, and France is doing a great job of standing up data centers. Why? Well, I think they've got something like twenty-one nuclear plants. And so where everybody else has gone solar and wind and renewable, France has stuck with their nuclear energy and they seem to be the only one who have enough energy to really build data centers. and then in travel and hospitality, Meta's Muse agent plugs straight into flights. but shops the rest of travel on the open web. So this is basically Meta Muse one point three. as we said earlier, it's really vying to become your personal agent with your credit card. And and so it's actually connected to the flight database, I'm not quite sure which one because there are quite a few. Amadeus, Galileo. these are the systems that the travel agents typically use to book your flights. but in terms of the rest of travel, it is still shopping the web just like we are. although it's obviously doing it for you hands-free. So so that is definitely something that I've been using a whole lot recently. I don't know if you have, but I've been getting AI to fill out all kinds of things on different websites. And then the final one in healthcare, ARPAH kicks off advocates putting millions into health tech firms building AI agents. And ARPA H is a federal agency that funds breakthrough health technologies and it's got a new program called Advocate, which provides millions of dollars to health tech health tech companies to build AI agents. and that's all good as long as they heed the advice from Sequoia, which is make sure you own the intelligence layer. so you better ask ARPA H for a bit more cash. Em: I'm gonna move on to another story it's getting a lot more attention and traction in terms of architecture, and that's the harness that we're all familiar with, but I think it's it's important and it's it's worth covering again. Building an AI agent means more than choosing a clever model. Someone has to manage its tools, working files, memory and recovery when a job runs for hours. Open AIs new agents, API packages, much of that machinery for developers using the software framework behind Codex The public beta supports managed computing environments or other infrastructure choices. It can carry essential context across long sessions and split suitable work between specialist agents. OpenAI says there is no extra API fee, although the tokens and tools used still cost money. Think of it as renting more of the workshop instead of assembling every workbench yourself. An AI agent harness can also manage the parts that rarely appear in the demo. What happens when a tool fails, how an agent remembers a decision, where temporary files live, and which actions require approval. Those details decide whether an experiment becomes a dependable service. OpenAI says its hosted environment can support sessions that run for hours or days and can preserve useful context as the working window fills. The public beta also allows work to be divided among sub-agents that can be useful when research, implementation, and review are generally independent. But more agents do not automatically mean better work. Coordination itself has a cost. And a poor plan can produce several confident answers that do not fit together. That could make a small team faster at building useful services. It also makes the choice of platform more consequential. The provider increasingly shapes how work is organized, observed, and recovered, not simply which model answers the prompt. The model may be the star of the poster. But it's the harness increasingly that determines whether the work survives contact with reality. For users, that suggests a better buying question, not only which model is inside, but how does the system recover, show its evidence and ask before it acts. And for those that just need a refresher on what harnesses are all about, think of it as an AI model as a person doing the thinking. And the harness is the working setup. It's the tools they can use, the files they can reach, and the controls around their actions. It helps the agent keep track of a long job, use software, and pick up the relevant context as work continues. OpenAI's new agents API puts that machinery at the center of its product pitch. Anthropic has also described harnesses designs for keeping agents productive across multiple sessions. The key takeaway when Two AR tools use the same model but behave very differently. Chances are the harness can be part of that explanation. So judge the whole working system, not just the brain inside of it. Ralph: Yeah, judge the whole system. I think this is the way. This is this is the answer. And this is what we said a couple of weeks ago, you know, Deep Seek obviously also launched a harness. It's the tool belt and the tool belt is almost more important than the LLM. I think we can switch out as many LLMs as we want. I'm very successfully on the music project, using clawed code to actually do the orchestrating and Kimi to do the development and codex to actually do the adversarial code review. And you know, the LLM isn't even coding, at least in terms of clawed code, it's actually pushing it to all these other models. And I think it just goes to show you know, we don't have to be tied in to any one model. And that means we can start picking the right model for the job. And actually, I don't know if you've seen this, Emil, but there's a really interesting story about GitHub Hydra That's coming online. and basically it allows you to build a coding pipeline that switches models based on the activity. So I I don't know what your experience is, but when I'm busy coding, I find it very difficult to keep switching to the right model. And obviously, the reason you want to switch models is because some are really expensive, like Fable 5.1, and others are much cheaper, like Kimi. And so but also even within clawed code, you've got Sonnet, you've got Opus, you've got haiku, and so each of these has a different cost profile. but what this GitHub Hydra project is doing is it's basically saying, well, we will build the appropriate model into the actual activity and so that you're not sitting there constantly trying to select the right model for the right job, because actually that's really hard work. And for the most part, I think a lot of people don't do it. They just kind of stick with one model, which means they don't necessarily have the smartest model on the job doing it. or it means that they've got the smartest model basically taking out the trash, which is not really what you want to be doing. And so yeah, really interesting to see GitHub going down this route and definitely something I want to try. but you know, just as you were telling talking about that story, it really got me thinking. about this resource on our bleeding edge website. and it's one of the things that I just think it's just such an interesting moment, especially with everything we've been talking about on Fable 5.1 and Astra Six. I hope you're seeing this okay but this is basically our little timeline view and it is a brief history of all things AI since the beginning of AI time so It actually kicks off with milestones in the nineteen fifties. but obviously what is so interesting is that seventy years of AI history and all of it is pretty much happening right now. and certainly within the last three years we've seen a huge acceleration. but I just I see it even happening even faster. So we might have to go into a day by day breakdown eventually if we are gonna get daily releases from these models. Em: I was thinking that it's getting so dense we need to expand the air view. It's just incredible, but it it sends quite the message, doesn't it? Ralph: Well what's so fascinating is if you are interested in just what's going on, yeah, w we can't cover all of this i even in our long form podcast. but I d would really recommend looking at this timeline view because it's ever so interesting just to see you know what's landed in September alone. and all the models are covered in there. well, as many as I can collect the news for. but certainly it's got all of the major milestones that we're experiencing live right now. And I really love it to just kind of go and get a bit of perspective on all of the labs and what they're doing and some of the big events in you know, got things like Merck and Moderna and the first positive phase three for individualized mRNA cancer therapy. That was a big news item that dropped in in August. so yeah, it's just I think in a you know years time when we look back over all of the items that we covered, some of the stories that we probably didn't even get to. if you just look at what's happened in the last month, like it's really, really phenomenal. So probably the biggest news story there is that the AI capability is just getting even denser. you can talk about AI slop, but there's the other side of it. as well where there's some seriously good shit happening. and our timeline certainly gives you a a really nice chronicled view of that. So definitely a resource worth checking out. and just because Craig Smith will be listening, yes, we will have a much better URL shortly. So I'm not gonna even bother reading our current one out. there you go. Em: thanks for that. That's a good reminder, Ralph And speaking of reminders, I urge our listeners to go try it out because the experience is quite fundamentally different, and that's all about full duplex. you know that awkward moment when a voice assistant interrupts just as you're gathering your thoughts. I get frustrated by this all the time, specifically with Grock. OpenAI's GBT Live One is now available to developers as well with listening and speaking handled together. The technical term is full duplex, which is a conversation can keep flowing while you interrupt, hesitate, or change direction. It's a small difference, but it's a market difference. OpenAI says the model can pass deeper reasoning and actions to a separate model behind the scenes. A language learning partner, speak, report substantially fewer interruptions during thinking pauses in early evaluations. This is a promising customer result rather than a guarantee for every conversation. The voice layer costs five US cents a minute, with the chosen reasoning model and tools adding their own costs on top of that. Traditional voice systems often resemble a relay race, Speech is transcribed, passed to a language model, converted back to audio and then played. Every handoff adds delay and can lose the rhythm of a natural conversation. GBT Live one handles the listening and speaking layers together, while a separate text model can do heavier reasoning or use tools in the background. That division of labor could matter in language learning, accessibility, booking services, and customer support. A caller may pause, correct themselves, or speak over the system without having to restart, which I just absolutely am loving, Ralph But natural delivery also raises expectations. A warm voice can make weak reasoning feel more trustworthy. So designers still need clear confirmation before consequential actions and an easy route to a person. And that changes the design challenge. A useful voice agent needs. conversational manners as well as correct answers for tutoring or customer support, knowing when to leave a silence may matter as much as responding quickly or accurately. This feels like a genuine welcomed improvement. Fewer interruptions should leave people more room to think while they work out what they want to say to these thinking machines. The best voice assistant may be distinguished as much by its patience and timing as by its intelligence behind the answer. So check it out, guys. Ralph: I was gonna interrupt you there, but I was terrified of what might happen. I was saying this is the full duplex experience right here. Yeah, I love that story. And yeah, I I honestly I think it's time we bring on our AI co-host that actually does a decent job of interacting with us on this podcast. they can keep us honest with all of those great stories and notes that we may forget as so often happens. Em: Yeah, that's a good idea. I think with full duplex landing it'll be more natural, bringing them into the conversation. Ralph: I love it. Yeah, I Em: I love it. Yeah. Ralph: Emile and thanks for all your great stories. We'll see you on the edge. Em: And yours.