Latest / The Bleeding Edge Podcast / Weekly AI Brief Sep 11 2026
Transcript
- Ralph: Take twelve. Emile just asked me a very interesting question. He said, Ralph, are you recording? And I wasn't because we were very distracted trying to make mics work. And here we go again for the twelfth time. Hi, and welcome to the Bleeding Edge Podcast. my name is Ralph Benker and with me in the studio today, Emile Ogier. also known as Eagle Eyes 'cause he spotted that the recording Em: Ha ha ha. Ralph: wasn't actually running. Welcome to the show, Emile. Em: awesome to be back for the third time, lucky. Ralph: Yeah. Well, I think it's fantastic because I actually think I know what the news is now. I hope I can remember all the stories that I was just riffing off the top of my head. but obviously the top stories all about Nvidia, they are really making waves. And the big wave we reported last week was all about their acquisition of Hugging Face, which is still progressing. and they're spending th thirteen billion on buying Platform that is hosting around 3 million open weight bottles. It's their largest acquisition, outright acquisition, that is, which is kind of interesting. They have been on a real buying spree. And if you kind of just put that in context, SpaceX bought Cursor for 60 billion. So hey, Hugging Face is pocket change for these trillion dollar valuation companies. But must be quite fun if you're working in the acquisitions department. Nvidia also separately signed a reported six billion licensing deal with Poolside to build more openweight models. they're all after the the open weight models. I think this is a really Em: Yeah. Ralph: interesting story because they're essentially building the market for what will be future buyers of their hardware. so open weight models allow you to basically you know run LLMs on your own computers and that means that if you're gonna do that you're gonna need very powerful computer which means you probably are gonna be buying Nvidia hardware and that means you're gonna be spending lots of money with them. So no wonder they are building out their own Nematron range of open weight models and doing deals with poolside to help them build more open weight models. that is what you talk call like investing in your future. the next news story SpaceX. So continue with a bit of the NVIDIA theme. They are buying NVIDIA's Vera CPU specifically for a genetic AI workloads. So I think the Vera CPU they launched some some months back probably about six months back now just kind of recalling a new story that came up in our feed and I don't actually know how much one of these Vera CPUs are. Emile, why don't you have a quick look and see if you can find a price for us by the time I get to the end of the news? I really want to know if I can afford one of these things. so SpaceX is obviously going to be dropping quite a lot of money into NVIDIA to buy those CPUs, and no doubt we will be seeing NVIDIA talking about investing in SpaceX because we know how much these AI companies love a bit of circular economy. And Hopefully we can also find how much they're actually investing in that. on the next news item, Salesforce that we're doing some analysis. and they have basically come up with a report that says the average company is now running 13 AI agents, which is actually up from five in early 2025. it made me think about a story from Microsoft about a year ago where they were talking about the agent boss, and they basically forecast something like every s you know manager will be managing something like 150 agents in an agentic swarm and really enjoying their role as an agent boss. And so we've got a long way to go in this AI journey. if we're still only running 13 agents in an average company. I think I'm running 13 agents just my podcast. but In other news items, you probably run thirteen agents just to look after that fancy mustache of you yours yours, Emile. so this kind of flipping to China news, GLM five point three flash stealth launched on open router. I don't know why they do this, it's kind of really weird, but they've got the name Ox Alpha. Why wouldn't you just call yourself GLM five point three? Maybe thought they thought it was gonna be crap. it's a little bit like writing a book. With a pseudonym. But it's topped the the token charts because you get so much value out of these Chinese models. but it's been outed now as being identified as the ZAI model. So there you go. If you want something cheap and high performing, GLM 5.3 will give you a lot of value for your tokens. Quen 3.8 Flash next is also out. minimax fast H3 and 10 cent. HY4 ship in the same wave. So I don't know what's going on in China, but they are absolutely smashing it at the moment. The other big news item that actually didn't make it on my feed this week is Deep Seek launching their harness. and everything is a plug-in. so that is a really, really smart play because, as we know, you use more of the harness than you use of the LLM actually. and it's one of the reasons why Anthropic has been so successful, is their harness is fantastic. So the fact that Deep Seek is opening up the next iteration of their harness is really interesting. So that's definitely one to watch. and I think we'll just jump into another couple of stories here that are very interesting. Super Groc Heavy. I know this is a favorite of yours, Emile. We were talking about it last week. Quietly removes near unlimited usage. From its pricing page. So, yep, it's getting a little bit pricier. You you didn't get in there fast enough, so you're gonna be meted now. so you know you'll you won't be able to make those funny cartoon strips for near zero dollars. Cloudflare ships a browser and a wallet for agents. I love this. Cloudflare runs about 25% of the internet, and they've got so many fantastic AI tools. but shipping a browser and a wallet for agents. let's see if anybody uses that. Obviously, we've had lots of browsers from different companies. Atlas was the last one from OpenAI, which they have basically shelved. so you know, just because you built a browser doesn't mean anybody's gonna use it. but we'll have to give that one a test and see if it's gonna survive the test of time. Nearly at the end, I think this is probably the last of the news items. Reddit's citation share in AI search results has reportedly collapsed. Now, this is very interesting because what we found very early on is that if you kind of think about answer engine optimization, AEO is they called it. or you know geo there was kind of lots of terms for this but basically you know your lm search it was relying on Reddit a lot and one of the things you needed to do was make sure that if your brand wanted to be top of the LLM feed then you needed to make sure that you had a Reddit page and there was you know a good share of conversations about your company and your brand on there. But apparently they're not indexing that as much anymore. So we'll have to take a look at what they're looking at instead. So are they scraping in websites? I really like the fact that they were pulling lots of information from Reddit because obviously lots of people talk about real experiences about brands on that. But a bit of a change there. So a bit of a varied news set this week. I hope it was interesting for you. What caught your eye, Emile? Em: the Reddit thing, I know what you mean, Rolf. I've also, love Reddit through the years, but unfortunately it's got penetrated quite heavily with bots now and that level of quality feedback information from real users has dropped and I think that's contributed a lot for devaluing the index on Reddit. I don't know a great deal more than that. I certainly will look into it. It's an interesting one. And already have an answer on your question. It's not as much as you might think, Rolf. If you turn your piggy bank upside down in your living room, I think you should just have enough change for it. So it looks like it's anything for the Nvidia Vero, it's anything between twenty thousand dollars and forty thousand dollars for enterprise buyers. And apparently if you're looking strictly at the hardware components, it's only seven hundred and four dollars is the estimated cost for the actual silicon CPU core, but it's nineteen thousand three hundred is the cost for the massive one point five terabytes of specialized memory that's integrated directly onto the board. So that that's quite a lot of money and as always we can see why Nvidia remains the most profitable company in the world right now because their margins are just so healthy. I think it's something in the excess of fifty Ralph: Well, no wonder they broke their Em: fifty to sixty percent, I think. Ralph: Yeah, well that's insane. So that's that's why they you know broke their earning model by forty five percent. so they were you know really ahead of earnings, because geez, I mean that's a crazy price. I wish I could sell something for ninety thousand. You s did you say ninety thousand? You did say ninety thousand. Em: twenty to forty thousand dollars for enterprise purchases. Ralph: Okay, for for each one of these things with Em: Each one. Ralph: the whole memory system and everything. yeah, I don't know. I think I might I might buy a car instead. Em: Stick with Claude Max. Ralph: Right. So that's the news items. We've got a couple of stories that we're going to deep dive into, but why don't you kick us off with your first story, Emile? Em: sure, no problem. actually off the back of last week's podcast, I know I had covered Grokbot and OpenAI computer use and we did do a bit of a d deep dive into that, but this is an important topic that's gaining a lot of traction and this time we're gonna focus more specifically on looking at the two comparatively, understanding where they diverge, why they're diverging and what it means to us in the market. So after looking at Grokbot And OpenAI's new computer use and Chat GBT work capabilities, there is an interesting way to close out this story. Because these companies seem to be heading toward the same destination, but they're taking noticeably different roads to get there. Both are betting that the next important phase of AI isn't simply a chatbot that gives you better answers. It's AI that actually can do the work for you. But look at how they're packaging this idea, Ralph XAI is making the metaphor incredibly explicit. With Gropbot, you can create AI teammates. They get persistent cloud computer. They can sign into tools you use, work across applications, remember how you like things done, and continue working for you when you're not there. And importantly, you can create several bots working in parallel. That means you could have one handling research, another working after sales, another working on operations, and potentially have the work. handed between one another. XAI even gives skills for teaching those bots repeatable processes and routines for running them automatically on schedule or where supported in response to events. So the XAI proposition feels something like don't just get an AI assistant, build yourself an AI team. OpenAI's approach, however, feels subtly different. Rather than asking you to assemble a workforce OpenAI appears to be turning ChatGBT in itself into a control center for getting work done. ChatGBT work can take larger goals, research information, use connected applications and files, create finished work and continue across longer running tasks. On desktop, computer use can actually operate your machine. Clicking, typing, and moving files across applications and the browser. And then schedule tasks add persistence, where it can happen once, repeatedly, or when something meaningful changes. So OpenAI's proposition looks closer to give ChatGBT the job and gradually let Chat GBT reach whatever tools are necessary to finish it for you. And that difference is more significant than it might first appear. XAI is organizing AI around the workers, but OpenAI is organizing AI around the work. These two philosophies could produce very different experiences. With GropBot, you might start thinking like a manager. Who owns this job? Which bot should I give it to? What skills should that bot learn? Which routines should they run automatically? You effectively create a small organizational structure around yourself. Open AI seems to be reducing that organizational complexity. You tell ChatGBT what outcome you want, and ChatGBT increasingly decides whether that requires research, files, connected applications, a browser, or an interaction with your actual computer. That could make OpenAI's approach feel simpler for mainstream users. You don't necessarily need to think about in terms of agents, you simply think. Here's what need to get done. But XAI's approach could become extraordinarily powerful for people who do want to think that way. Entrepreneurs, developers, small companies could potentially create as specialized digital teams without employing separate humans for every narrow workflow. And there's another fascinating difference. Gropbots currently share persistent cloud computer, including files, browser sessions, and application logins. OpenAI, meanwhile, is spreading agency across several environments, connected applications, cloud browser capabilities, and importantly, your own desktop computer. we could be watching two different models of the future emerge. One says move the work into the cloud, and the other increasingly says bring the intelligence into the environment where you're ready to work. And that could have consequences well beyond these two companies. For decades, software companies have competed to become the application you spend your time inside of. AI agents potentially turn that relationship upside down. if your AI can operate half a dozen other applications for you, how much do you personally need to interact with those interfaces anymore? Suddenly, the most important software layer may not be the application. It may be the agent sitting above the applications. And if that happens, the company controlling that agent could occupy an incredibly powerful position. The agent knows what you're trying to accomplish. It knows which services you use. It potentially remembers how you work. And eventually it may decide which pieces of software are worth using to accomplish your goal. That raises some uncomfortable questions too. such as who controls the permissions, who sees the data, who's responsible when the agent makes the wrong decision. And perhaps more importantly, how much autonomy will people actually be prepared to give these systems? there is a huge psychological difference between allowing an AI to draft an email and allowing it to send one. Another step again to purchasing something and another step yet again to changing a production software. And another step again to allowing 10 autonomous agents to make these decisions while you're asleep. Both companies are already building approval systems around consequential actions, which tells you something important. The technology problem isn't simply making agents capable. The commercial problem is making them trustworthy enough that people are willing to use their capability. And that may ultimately decide this market. Not which model scores. two points higher on a benchmark, not which chatbot writes the smartest answer, but which company convinces millions of people to say, yes, you can actually do this for me. XAI appears to be betting that we'll manage collections of specialized AI workers. OpenAI appears to be betting that Chat GBT can become a broadly capable interface between you, your information and your entire digital working environments, neither approach necessarily has to lose. We may eventually use both. A general AR orchestrator handling our lives and work with the specialist autonomous workers underneath it doing particular jobs. And that leaves us with perhaps the most interesting question of all. For the last few years we've been asking which company has the smartest intelligence. The next few years may focus on asking something very different. Which AI do we trust enough to give the job to and then walk away? Because once that happens, We are talking about a new layer of labor. And the race to own that layer may turn out to be far more important than the race to build the best intelligence. Ralph: Yeah, I think this is the story of our time. And and let's be clear, you know, you don't have to go out and try and build an agent to experience this. Opening up Codex or ChatGPT or Claude, you know, you are in the agent that is the agent already. And I think a lot of people kind of have this view that they have to go away and build an agent and stand up something like a Grokbot. and I think I'll just you know speak directly to this point from some of my own experience this week, Emile. So as you know. it's year-end tax time a little bit late for me and I've had a massive amount of expenses and all kinds of admin that I needed to solve. And it was just frankly too big for me to tackle in the traditional way. And so of course, what did I do? I reached for Claude and I Claude, I've got a problem here. I gotta solve this. And literally from there it has grown into the kind of agenc experience where I have literally outsourced to Claude the entire workload around managing the filing of various year end accounts for various companies. And I have gone as far as not just letting it draft the email but letting it send the email. And not just that, I gave it the job of finding a new accountant. and it went out and evaluated a bunch of accountants. so mostly these were people that had reached out to me over the years and had emailed in their credentials and it evaluated one, it checked to see what information there was about their services. It even evaluated how prompt they were in filing their own accounts. And in the end, I actually went for the more expensive accountant that filed their accounts before they were actually due. Whereas the second option accountant always filed their accounts late. So do you want an accountant that files their own accounts late or do you want one that is actually organized and delivering it ahead of time? That was the decision making factor along with a couple of other points. but the point being, you know, I gave it a goal. I told it this is the outcome I want. And I gave it access to my Gmail account that I use for for for these particular companies. And it found this particular accountant, it reached out to them, it broke at the deal with them, and of course, you know, I made the payments and everything to actually set it all up, but it filled out the paperwork, sent it all off to them. And of course, you know, the accountants are are constantly saying, Wow, Ralph, thank you for your very comprehensive reply. because Em: Thank you. Ralph: I think the one feature that you get with Opus V is it's quite verbose. it likes to make sure that you know that it knows that what it's doing, Em: Okay. Ralph: and and so I think they've been somewhat overwhelmed. My outgoing accountant has also been on the receiving end of this. and suddenly, you know, it seems like I've really got my act in gear. And of course, Claude is finding all kinds of holes in previous submissions, kind of going, Why haven't you claimed this? And why haven't you claimed that? And why is this information all wrong? and it's all kinds of things. birthdays, you know, for my kids that are obviously on my various tax submissions that somehow have been entered with dates switched, wrong addresses, literally hundreds of things. and some of them actually with some really, really large tax implications. And so the point of this story is this. when you actually start to give the agent, and in this case the agent is clawed, clawed code, when you give it the kind of context that really allows it to operate. with a bit of breadth and you then actually give it authority to go away and engage in this case three different accountants that are actively working on various submissions. and I've even gave given it access to my bank manager by the way who has now also got about 12 different emails from me and when they're wondering what is going on with Ralph because I I hardly ever email them. and I'm emailing them for some pretty obscure requirements to go away and find documents that they filed years ago. and and you know what? I am really trusting the model. So to your point, you know, you will give it as much let's just say room to maneuver as you trust it to maneuver in. it's definitely told a few things that I wouldn't have wanted to share necessarily. thinking that it was the right thing to do. it is definitely thinking on its own what it should be writing to these people. I have taken to reviewing every single email before it goes out, just because there's a few oddities in there. no harm done at the end of the day. It is driving the process pretty hard. And we are talking about hundreds and hundreds of decisions made over the last five days to actually get everything up to up to date and submitted. And you know, in terms of an agent operating independently with good context, with a clear outcome, this is probably the best example that I have seen personally. And I've built a lot of agents over the last year and a half. I've never seen anything work quite as broadly and unattended as this particular use case. And so it's actually been one of my favorite learnings of AI. I'm constantly surprised at. its capability and it's I'm just I c I cannot reconcile in many ways, you know, what people are saying that AI is is a lie. it's not delivering on the promise. because I can honestly say that it has improved my productivity over the last Em: Yeah. Ralph: week by a hundred percent. I I you know stuff I haven't been able to do in a year, it is doing in minutes. and maybe that's just my own you know, biased towards not doing admin, but it has completely unlocked my interest in solving these problems because it's suddenly become so much fun. So I think that really speaks to the use case that you're describing. and I agree with you. I think dedicated agents, but also kind of broad outsourcing of work to a platform, in this case, you know, something like ChatGPT or Claude Code. I think I think the latter is easier. I think the kind of grokbot example I will use as well for very specific use cases that are repeated, but for work that is a little bit more less let's say structured, that requires a range of inputs, that requires a range of communications, I'm finding the agentic case very compelling. Em: it's great that you said it. You had inspired me to look into it myself, having spoken to you and another lawyer friend that we mentioned on the previous podcast. slightly different strategy to you, but same result. It's just been wonderful. so you're not imagining it, Ralph I've got a fairly complicated tax system I need to sort out and both my wife and I were dreading it. But thanks to my wonderful wife, Kelly, she got on it and set up Claude and connected it to Clearbooks. and it was just wonderful seeing it working. It's using a regulation called PSD two, which authorizes banks to share your information with third party suppliers that act as aggregators on your behalf. And there is no charge behind that. the bank that's sharing that information can't charge. So it does give users a lot of power. You can effectively do lots of wonderful things with PSD2. But one of the key things that people are really enjoying, especially myself with this accountancy connectivity, is I've got various banking feeds, all those feeds go into ClearBooks. I don't even have to press a button. Those transactions get updated daily. I wake up the next morning and all my bank statements are aggregated, filed, and you know, all that painful admin work of having to work through all those row items line by line are gone. nothing to think about, everything was reconciled. just saved me days of work. So yeah, I'm really excited about it. I'm looking forward to the next personal project to Put the bots on to get things done and to free up more time. Ralph: I mean, this is literally the perfect w use case. And I think a big opportunity actually for the accounting industry, you know. I think they are able to set these systems up for you in a trusted environment. I think this is absolutely what most people can actually leverage. And I was going through a couple of resources here in Spain. And I actually found a service that is offering for about two hundred and forty euros. They are basically saying, you know, we will submit your accounts. And not only that, we will look over your history for the last four years and we will find anything where you've been overcharged or you've basically you know you know just missed the opportunity to reduce your tax burden and all of that for about two hundred and forty bucks. You know, I'm When when I'm asking my accountant to revise something, they w they're charging me a hundred euros per revision. And I had eight revisions that I was talking to about. That was eight hundred Euros. And I thought to myself, well, that's absolutely crazy. Obviously I haven't chosen to do those because you know, I'll be using my my air AI to to do those on on my behalf. But this is an incredible use case and I do think it is a big shift. For the accountancy industry. And I really hope they kind of tackle it because you still want an accountant at the end of all of this. It just makes them so much more effective. I don't know what the tax man is going to experience because I can imagine just an absolute deluge of AI findings and you know, a lot of clawbacks, which you know, ultimately I think governments have been relying on people's. lack of understanding and knowledge of what they could claim, how much they could claim. and so this is kind of fascinating. I predict that there is just going to be a lot of people that are way more sophisticated at at filing the right tax. And so obviously that's welcome. you know I think it's a bit of a a break for all of us. certainly for me, I have found lots and lots of holes in what has been submitted on my behalf in the past. And a lot of that's probably down to the information that I provided. and I'll just go one step further. So here in Spain you have to pay you know a variety of different things and they give you these little PDFs that appear in your bank account which are essentially the receipts. they're so small I can't actually even read them. I have basically downloaded them all, sent them in to to Claude which has read them I've taken photos of literally every single receipt, some stuff that I didn't even fully understand because I just didn't understand the language of what was on it. And it has found such incredible information about the valuation of you know one of my properties, for example. and it found that I was basically applying the wrong reduction in the but basically in the amortization of the the the value of that particular property, which meant I could have claimed a lot more back. And you know, I I've been looking at this document. I get it every single year. It's got a number on that I I never understood. Claude figured it out. And what's amazing is if you are in multiple tax jurisdictions like I am, and you do need to transact in multiple languages, that is now not a barrier anymore. you know, I was talking to somebody and they said, you know, it's complicated because you have to deal with stuff in Spanish. And I and I was gonna say actually, yeah, I don't, you know, the AI, my claude will write in Spanish to anybody. but I just said, Yeah, it's it's a bit complicated. But of course it's not. And so if you are an expat, this really unlocks a lot of opportunities to be a bit more sophisticated. So yeah, I'm pretty excited by this. And I actually think it's incredible. Like, you know, we kind of come to this every week you hear a story about somebody that is quite advanced in their AI journey and they kind of go something like, I wish I knew this earlier. I wish I got to this earlier. And I feel like that is literally every week for myself, even though I have been going really, really hard at it in terms of using these AI tools and leveraging them. and so yeah, if you haven't started on the journey, it's never it's never too late. But even if you are hard at it, I think there are still you know use cases that we can unlock and it just delivers in enormo enormous personal benefit. So enterprise, what's the problem? that's my question. How are you not unlocking value and and benefit? and I think yeah. This is just something, you know, we have to kind of consider. But maybe trust, security, you know, enterprise has to be locked down for lots of good reasons. but in our private lives, maybe it is a little bit easier to kind of open things up. Em: I was thinking about it from my point of view, it's been several years since I actually paid for an accounting c service, 'cause of my good fortune, my wife's a CA, so yeah, that that really helps things. But when I was paying back in London in the day several years back, prices easily start from 120 pounds. I'm sure it's gone up a lot more since inflation. so that alone, if we we're looking at old prices, not assuming for inflation, you're looking at a bare minimum 1500 quid a year to do the bare basics for a a sole traded limited company. And that's not even asking for the special requests such as the the things you needed to do, Ralph If you think it think how much that can unlock for you. in AI, how many of subscriptions you could pay for for the year to automate that all, have your time back and money back too. That's a pretty attractive proposition. But you're right with that that swell of requests that the government's most probably going to be hit by. I think consumers are a lot more AI enabled and red pilled than the government is. They they always a bit slow to to catch up with the latest in tech. So I can imagine their human workforce getting inundated by all these very intelligent requests. But without further ado, I'm gonna move on to two news stories. Just drilling in a little bit more about two things you've already covered, Ralph NVIDIA has crossed one of those thresholds that sounds almost fictional. The company is now generating more than a billion dollars of revenue every single day. It's the latest quarterly sales more than double from last year, with the overwhelming majority coming from the data centers. That tells us this is no longer mainly the gaming chip company that many people still picture. Nvidia has become the toll booth on the road to artificial intelligence and almost every major laboratory and cloud company needs to pass through it. The billion dollar figure described revenue rather than profit, but the scale remains extraordinary. Outlook points even higher, suggesting that appetite for AI infrastructure is still accelerating despite constant questions about whether the industry is overspending. And that is the real story. Nvidia is not simply selling more chips. It is giving us a live reading of how much conviction the world's largest companies have placed behind AI. One supplier can now generate the annual revenue of a sizable corporation in a matter of days. Before many of its customers have fully proved what their enormous AI investments will return. So why this really matters, Nvidia's results measure the enormous demand for AI infrastructure, but not yet the return customers will earn from their spending. And I do wonder, does this prove that the AI boom is becoming productive or mainly prove that everybody is still buying shovels? from our own experience, Ralph the positive returns we're getting out of all the utility in AI, I think the answer is quite clear for us and will be for others as they get on board. The second story that manages to be technically fascinating and alarming at the same time. Obliteration.ai has released a modified version of Z.ai's GLM five point three, which is a powerful Chinese open weight model built for coding and agentech work, and it has deliberately weakened the model's instinct to refuse sensitive requests. This goes much deeper than persuading a chatbot to role-play an uncensored character. The company says it finds the mathematical patterns inside the model associated with saying no, then edits those patterns out of the weights. The reasoning and coding ability remain, but the barrier that normally slots a model from completing certain tasks is dramatically reduced. Obliteration AI is positioning the result for authorized cybersecurity testing. red teaming and vulnerable research. Its own testing suggests the model retains strong coding performance while becoming much more willing to complete offensive security work. Those results come from a company not perfectly controlled but the underlying capability is what makes the story important. The pitch is your AI, your rules with customers expected to supply their own policies and controls That may unlock legitimate work that mainstream models refuse. It also shifts far more responsibility onto whoever deploys it. Once guardrails can be edited out of an open model, safety is no longer something we can assume comes packaged with the intelligence. It has to exist throughout the system surrounding it. So Open weights make safety alignment modifiable, the governance access controls and sandboxing must therefore sit outside the model as well as inside of it. So Ralph, is this user control safety a sign of maturity in the market or are we handing out the brakes separately to the car? Ralph: Yeah, look, we really saw this with the ChatGPT hack into Hugging Face. you know, what some five, six weeks ago, when ChatGPT, a model of ChatGPT, hacked into Hugging Face in order to access the answers to a benchmark test that it was going to take. And Hugging Face actually started to use Claude to basically see, you know, what this hack was about, where it was coming from, what was attacking them. And immediately Claude's model shut it down and said, Well, you know, I I can't access those features because that's not allowed. And so Hugging Face deployed, I believe they deployed Deep Seek. to then essentially diagnose the hack because the frontier models wouldn't do it they were basically said this is not allowed this is an illegal use of the tool and so exactly as you're describing you know if you want to actually have an offensive strategy so you know where you you can actually defend by y using models that are essentially unlocked. it's the only way you have to do it. and so obviously that means you could also tell it to go and hack other you know organizations. but you have to take the gloves off. You have to take the gloves off to defend yourself. To to defend means you have to offend. And if that makes sense. And so that's just the way of the world now. And of course you know the political correctness of the frontier models is actually going to cost them a roll. we cannot use them for defense. They will not you know cross those lines. And I think that's a real problem. and it is absolutely what the Chinese models are going to exploit. And it is no surprise that people are even trying to, you know, essentially jailbreak them further so that they will be even more aggressive in their ability to defend you. so you know it's kind of like the saying, I don't know who invented it, but it you know, the attack is the best defense. and so preparing yourself with an attack capability is the only way that you're going to be able to defend against you know, hackers in the future. now, you know, the future is now, it's yesterday, it's today, it's gonna be tomorrow. and so you will definitely invest in a frontier model in order to do the work, but you will absolutely have to invest in one of these Chinese models that essentially have the guardrails relaxed and it's the only way that you're gonna be able to defend yourself. Em: Yeah, I totally agree. And if you look at nature and sports and game theory, offense beats defense. Attack is the winning strategy. Ralph: Yeah, I guess with that, that's going to announce the end of the show. So without further ado, that's Em: Can't hear you. Ralph: been another great episode. Emile can't hear me. All of you can. he's waving at me, which is hilariously funny. But thanks very much. We'll see you on the edge.