Latest / The Bleeding Edge Podcast / Weekly AI Brief Sep 20 2026
Transcript
- Ralph: welcome to the Bleeding Edge Podcast. My name is Ralph Benker, and with me in the studio, as ever, Emile welcome to the Edge. Emile: Great to be back, Ralph: we're back and we're committed as ever to bring you the AI news. And there are lots of wonderful stories, so let's dive straight into it. Without further ado, it. It brings me to a story that is maybe a a little bit more of a just a question in many ways, you know, what's actually going on? I did a bit of deep research on what I would term as probably a little bit of the a backlash really in terms of AI content. So Probably what you've seen as well, Emile is that there is quite a lot of negative sentiment around AI content. and a lot of people are saying they a lot of people are saying they are saying no to AI in terms of AI content. and we've had some really landmark, you know, interesting landmark rulings over the last couple of weeks, with Australia, for example, saying That they will not allow AI generated music to enter the charts if it's significantly generated with AI. my question to you know the industry is what constitutes AI? Because if you think about electronic music, I think the line is incredibly blurred. and and of course, as you all know, I've been generating an enormous amount of AI music myself. I tend to write most of the lyrics. certainly for the the key songs and and albums that I release. but in other cases, you know, I use a combination of AI generated and human generated and nature generated sound. I tend to master all of the music outside of the AI tools, but the tools that are doing the mastering were built by AI. And so I think you know moving away from AI in this sense is going to be very hard. I think the industry is almost certainly you know built and layered and becoming more so. I think a lot of platforms are reporting like 75% of the new content that's being uploaded is essentially AI generated. And and I think you know there is a deluge of AI slop that probably people don't like, but there is also increasing sophistication in content generation. And so are we really going to reject good AI music? Are we really going to reject images that are useful and communicate, you know, key points, so whether they're diagrams or just actual images, are we really going to reject AI generated article content If it's actually giving us the information that we need, are we really going to reject AI-generated adverts in movies? and of course, even today, you know, we have seen some significant developments in the movie technology. So one of the the the tools that spoke about a few months back that has come about, you can essentially act in front of a green screen, and that is a real human actor in f acting in front of a green screen. And the AI will then generate dynamically the scene behind the actor. And one of the things you can do is you can go and record elements of the landscape and the scene that you want in the background and then actually feed that into the AI model. And it will then just generate the human moving through that scene. And so again, you know, this is more of a hybrid approach. So exactly where is the line of that? Are people going to reject that? I think AI already runs so much of what is going on in the world that maybe folks don't even know all the time when they are consuming AI content. and what it seems is that although people when interviewed, a lot of surveys and studies, you know, people will say no, actually when it comes to it, they will watch it. And so we've got something like sixty percent of consumers. Find an AI label off-putting. But put it in a blind test, more of them pick the machines writing. 95% of enterprise AI pilots show no measurable profit, but the hyperscalers will spend around 700 billion this year anyway. So these are kind of two stories that I think are somehow interacting with each other. On the one hand, people have got a bit of an aversion to the AI label, on the other hand, we've got the enterprise customers trying to prove you know why this you know why the investment in AI is even worthwhile. Hyperscalers are spending enormous amounts of money. is there actually a a true business case for this? And and really kind of just the intersection of those two is somewhere where the truth lies. and so I think you know if I just summarize what what what I think is happening on the one hand enormous amount of content being produced by AI. On the other hand, people saying they don't want to consume AI content, but they are consuming it anyway. the volume has gone up, the consumption has gone down, but the kind of net consumption is essentially is still the same. If if you you know just because we've got so much more content, we have a little bit of a a a drop in the consumption, but Overall it's still the same as what it was. and then on the other side, you know, we're kind of calling out that AI is not delivering and it's not delivering the kind of profit that enterprise is looking for. But on the other hand, every single engineering department is adopting AI, every single marketing department is adopting AI. We're seeing AI effectively writing books, generating you know, photography and visual content, generating video content, generating music content, and all of those things are working incredibly well, to the extent that you know people aren't even able to identify the differences anymore in some cases, if it's good quality. and so I think it's a very confused state. So one where I personally am seeing a huge amount of benefit and I'm actually seeing a huge amount of automation, I'm seeing a huge amount of Optimization, quality improvement, content generation. But on the other hand, you know, there's this backlash and then enterprise still struggling to make it profitable, but enormous profitability in the circular economy. That seems to be you know what the hyperscalers have created. and so I just thought I would kind of share that. I don't know where all of that's going to end up, I don't know where it's going. but it just feels like there's a lot of tension in the system. And a lot of people calling out that AI is a lie. I'm just seeing enormous amounts of YouTube videos calling this out, basically saying that the business model of data centers cannot work. It's like the bumblebee that can't fly. but there it is, you know, we've got 700 billion of spend in the next year that is still going into this industry. So anyway, I just thought I'd share that because I just think it's such an interesting tension of our time. Emile: Absolutely. And I I do have things to say on a draft. It it's particularly interesting. I can't remember who, in addition to that, some heavyweight in financial markets also recently drafted and there's been a backlash on that but there's been a big divide in it because the substance of the article is them. they prominent VC player that that hadn't lost in the markets consecutively for something like two decades. unprecedented record. And they admitted to using AI. So there there's two stories going on is their conviction and the substance of the article was them, but they used AI to produce the surface layer Ralph: AI to produce the the the Emile: if somebody's not putting any work into it, not proofing the final article, that I understand. But if it's just syntax, you know, if the substance there I I think maybe people are getting too critical on that point. I don't think it's going anywhere. And It's really how you approach AI. it's a phenomenal tool. where AI really makes a difference is when you use it in the middle. And by that I mean you have the original idea, you know what you want to say, you've got the industry knowledge. And then on the other end of the spectrum of the middle is you finalizing things, making sure that things still connect and make sense and the final polishing. But there's absolutely nothing wrong with using AI in the middle for idea formation, getting better structure, asking it to soundboard your ideas, finding flaws in your own thinking, or finding ways to enrich it. Ralph: Yeah, and I think you know, maybe there is some context here. there are many things that the AI actually just can't do very well. So we know holding a coherent argument across long form content without drifting, you know, it is a bit of a reliability issue there. I'm seeing it even in my coding agents. I have to really put guardrails in because it'll start fixing stuff. I didn't ask it to fix the next minute. I know it's three layers deep. Into something that I really didn't actually want it to touch. And you know, the main feature goes unbuilt. And so this is really a problem of kind of guiding it towards the right direction. And I think maybe, you know, where all of this comes in is the skill of our time is the guardrail, is the guarding, and and so ensuring that it doesn't actually drift when it's doing work. but in terms of image generation, short form video, actually, you know, short short form commercial text, I think it does a reasonably good job. Is it at the same level of creativity and quality as the best human writers? No. is it better than the average Joe? Yes, probably. and that's really helpful. And I think you know we just need to be able to use it in the right way. I don't think it does everything incredibly well. but what it does do, well, I I think you know is enormous enormously useful. And I'm actually seeing that there are companies that are able to leverage this. It does r seem like, you know, for the most part, people are not able to get short term value out of AI. I think enterprise still has an enormous issue with data and quality of information. and that is really, you know, as we've said many times, an absolute foundation of any agentic AI system is that it has good quality information. The minute you give it crap data, it's gonna give you a crap result. and I think most enterprise are still suffering from the quality of their data and availability of their data. So so there you go. but it's not going away and obviously we're seeing so much regulation kind of coming in and obviously a little bit of let's just say subjective opinion about what people are saying they do and do not consume from an AI perspective, but the reality actually seems to be opposite from what a lot of people are saying. So I think we're gonna struggle to make sense of all of this for quite some time to come, given the level of investment. I think you're right. I don't think this is going away anytime soon. given the level of automation and ease of use, you know, we've just been talking about Crop Pot. I really think that most people and most businesses are going to be able to have at least a handful of business like use cases that actually provide business value. and maybe that will start to kind of shift sentiment overall. but it's an interesting moment. Emile: just wanted to add to that final thoughts if you think about it, we exiting wave one of AI where we know the responses have been very verbose, there's obvious markers. But we're gonna be moving into wave two now where the responses are gonna get more concise and more Ralph: Where we know the response is being very perverse, there's obvious markers. But we're gonna be moving into one three you know the are gonna get more. Emile: succinct. so it's a temporary blip and I think it's gonna become less obvious as the the models iterate. Ralph: Well, that's very exciting and I look forward to wave two. I love a bit of AI surfing. So when that second set arrives, I am ready. I'm out there paddling in the wide open AI sea So Emile what's your story? Emile: It's a continuation, chapter two about automation and computer use. This time we turn our heads to OpenAI. It's another development that I think deserves more attention than it's getting. OpenAI is increasingly giving ChatGBT something remarkably simple, but potentially transformative. It's a mouse and a keyboard. With the newer computer use capabilities, ChatGBT can move beyond telling you how to do something on your computer and it can actually start doing it. On the desktop, chat GBT can work across applications, tools and websites, clicking buttons, typing, navigating pages and moving files on your behalf. And these actions can form part of a one-off job or even a recurring schedule task. Think about the difference. You may ask AI, can you analyze these figures and tell me what should go into my presentation? Computer use moves us towards find the figures, open the spreadsheet, analyze them, update the presentation, and leave it ready for me to review all open on my computer. That is a very different relationship with a computer. And OpenAI is wrapping this into something broader called Chat GBT Work. An agent designed to take a goal, break it into smaller tasks. Work across your apps and files and potentially keep going for hours rather than waiting for another prompt after every single step. There's now a built-in browser too. ChatGBT can move between multiple tabs, work with websites, download files, and deal with pages where you need to sign in. The browser has its own separate states, while OpenAI also offers a Chrome route for tasks that need your existing Chrome profile and logged in sessions. That may sound like a relatively small feature, I don't think so, because traditionally, if you wanted two pieces of software to work together, somebody usually had to build an API integration, and that was a lot of work. Computer using AI introduces another possibility. If a human can operate the software by looking at the screen, clicking and typing, perhaps AI can too. Suddenly, thousands of applications that were designed exclusively for humans. potentially become accessible to AI agents without every software company having to redesign its product around them. And then OpenAI adds scheduled tasks. So instead of saying check this dashboard, you can increasingly say things like check this dashboard every morning, compare it with yesterday's results, update my report and tell me only if something important has changed. That brings us So a really interesting question. What happens when AI stops being something you visit and becomes something that operates your existing digital world? There are obvious risks. If AI can click buttons, it can potentially click the wrong button. And if it can use a logged in website, identify permissions and security becomes enormously important. But this is precisely why computer use matters. The next major AI benchmark. might not be. Can the model answer this question? It might be, can I give the system a real job on Monday morning and trust it that it will hand something back to me safely and securely? The chatbot era taught computers how to talk to us. The agent era may be about teaching them how to work for us. So yeah, just a common theme playing through. Ralph: Yeah, I love that. And I've been using you know com computer use extensively. And what I can say is that it's a real game changer in terms of productivity from my point of view. There's a lot of repetitive stuff, especially when it comes to the music side of things, just actually interacting with a vi variety of distribution platforms. it means that I can actually, you know. get the AI to upload the descriptions, the the lyrics, and set you know a variety of different attributes. but I've also seen the same thing in terms of you know some of the tax accounting and it's been so much better at navigating these websites than I have. So finding stuff, especially platforms that I'm not on every day. One of the biggest frustrations you know, I was just d setting up some Google password admin yesterday and I literally couldn't find this button. and Claude was effectively able to n navigate that in a much shorter time than I could. So they're just way more capable at exploring the full capability of a website. I think there are issues with it. So there are certain things that you know I have asked the AI not to do, so it tends not to do the the sign in and the submission at key points. So I still have this kind of human in the loop which actually means I am the blocker in a lot of occasions. it does give me an element of control and I think that's the next part to to go. But obviously what it really starts to indicate Emile is that I think the user interface is rapidly being disintermediated. what is the point of a user interface when you've actually got your AI agent actually interacting with that website. This is where CLIs and MCPs become way more capable. And and of course you know it means that we don't have to build a front end. We can really just build some way for an agent to actually interact with the data set that we have sitting behind it. And the user interface might be a relic of our time. but obviously we have to kind of keep things somewhat human accessible and human readable as well. but it's just a so it's it's a much slower way of transacting. and and I think you know, it might just be that the user interface is getting in the way. and so it's something that probably does need to be disintermediated. but you know, where does that leave us as humans being able to do the work? and so I think it will exist just so that we can actually see what the AI has been interacting with. and should our system go down for whatever reason, then it you know we could still press the buttons. but I think it's a a big shift. And probably what it means for application builders and website builders is that you're now really building for two use cases: the the human interaction and the machine interaction. And that's something that I think we've seen some very early stages of that creeping in. I know on Archive you know they published a lot of papers and I think we mentioned again on the show a few months back, one of the the first papers that I I saw a few months back that was actually built for machine reading. So they didn't expect the human to read the paper, they expected your chat agent to basically ingest it. and so what was different structurally about that? Well what they did was they said, do not summarize this, do not exclude these points. you know, these are the structural elements of the argument of this particular paper. So these are the things that you really have to extract from this and do not change them, do not reinterpret them, do not recalculate them, you know, state them verbatim. And and that's clearly, you know, a bit of a a change in the way that. That particular research paper was written. and no doubt you know we we will see the same for websites and web applications. So that's really a new frontier, and I think it's very exciting for people that are building websites and that you know are building web-based applications because you actually get to now build for essentially a a new form factor for a new customer, for a new way of ingesting, and we get to really rethink. the customer experience. So hopefully you know that kind of ushers in a whole new capability and a whole new skill. important for for us to know and learn. Emile: Ralph I remember us speaking about this at length around early part of the year. pretty sure it was January, where we were wondering where this abstraction layer was going. I think it still is moving in that direction, but I think where my viewpoint has changed, I was effectively saying the GUI is going to be decimated, it's we won't need it, it's getting in the way. and I I do think that is the case from an execution point of view. but from an observational point of view and the human in the loop point of view, I don't think it's going anywhere. So I just I do think it has utility, but in the past we used the GUI to execute and observe and I think that is going to dissect and it's gonna become a observable interface for humans just so we can understand what our operators are doing for us. Ralph: Yeah, so it'll be a lot more you know what happened, what still needs to happen, what the results and the outcomes that were. So maybe a bit more of an audit trail of Emile: Correct. Ralph: you know what your agent has actually gone ahead and executed for you. So yeah, I love it and I'm very excited to to see what comes next. And obviously with GrocBot you know, this is kind of what's happening. It is essentially driving computer use. it's interesting to see that OpenAI are also embracing this and achieving some success. it's been one of the things that Claude has done exceptionally well over the last 12 months. Claude in Chrome has been you know a wonderful feature, flawed and not always capable, and a lot of websites actually blocking you know what they would term as bot activity. but I think this is gonna change. people are gonna have to really open up for registered bots potentially to be allowed to to to interact and transact on our behalf. so there's a lot to think about structurally. I don't think we've got our head around this. We are still trying to control bot access to so many sites. Cloudflare are actively blocking bots and And obviously, you know, that this is something that I think is in response to maybe some let's just say lazy and illegitimate or fraudulent use of bots. but I think you know what's next in 2026 is that we're gonna grow all of that up and actually give it a formal role. So I'm very excited for you know architects of web applications the internet as a whole because I think we've got such a big problem to sort out there. How do we actually build the internet for agentic AI? And and that probably needs a rethink of a lot of the architecture. So I'm smiling yet here because it's the kind of thing that I really like doing. And in fact, you know I'm I'm already building systems that are really trying to take that into account just you know how is all of this going to interact with an agent on the other side. I think that kind of brings us to the end of our episode today. so thank you for listening and we will make sure to get the next episode out promptly. see you on the air, Emile Thank you so much. Thanks See Emile: Thanks everyone. Ralph: you next week. Ralph Behnke: The bleeding edge of technology and artificial intelligence where innovation splits the known world and a new wave of intelligence emerges, pushing boundaries, transforming the future.