Full Episode Transcript

Good afternoon. It's Friday, and while there's moderate rain in London, we're here to sift through the latest developments in artificial intelligence. Alan Turing once remarked "The search for new techniques must be regarded as carried out by the human community as a whole rather than by individuals." This week that sentiment resonates strongly as we examine the ongoing challenge of separating signal from noise in a landscape cluttered with claims of breakthroughs and innovations. It's a familiar scenario — the promise of the next big thing often overshadows the substantive progress being made. Today, we'll look at what's genuinely worth paying attention to and what's simply noise. This is Turing's Torch: Artificial Intelligence Weekly — the bits that matter, minus the hype. Right then. Let's have a look at what's been bubbling under this week. It seems we're at that familiar point where the shiny promises of new technology are colliding head-on with the rather more mundane realities of cost control and of course responsibility. Let's start with the whole question of artificial intelligence accountability. It's no longer just a philosophical debate for late-night seminars. It's moved into the boardroom, driven in no small part by the looming spectre of regulation. We're talking about a push to make companies answerable for the decisions their automated systems make rather than simply shrugging and saying "the algorithm did it."

Now what does that look like in practice? Imagine a bank denying you a loan based on some artificial intelligence's assessment. Under these proposed rules, you'd have the right to understand why, to see the criteria used, and to argue your case. Or picture an automated hiring system rejecting your job application. You could demand transparency about the reasons behind the decision. The implications are considerable. Suddenly businesses need to invest not just in systems that can make decisions yet also in systems that can document and justify them. They'll need to establish processes for handling appeals and resolving disputes. It's a whole new layer of overhead. Of course the cynical among us — and let's face it that's most of us — might suspect this is all just window dressing. Will companies truly embrace transparency, or will they create elaborate, meaningless explanations to tick the boxes? Will individuals genuinely have the power to challenge automated decisions, or will they be met with bureaucratic resistance? My guess is the answer will, as usual, lie somewhere in the middle. There's a growing recognition that these systems, while powerful, are not infallible, and that human oversight is essential. After all, algorithms reflect the biases and priorities of their creators. Without proper scrutiny, they can perpetuate and amplify existing inequalities. It's a laudable aim, yet the road to good intentions is, as ever, paved with potential for unintended consequences. And that brings us neatly to the issue of cost. We're hearing whispers that some organisations are discovering that their ventures into as a result-called "agentic artificial intelligence" — that's the idea of setting loose an artificial intelligence system that can act somewhat autonomously making its own decisions to achieve a specific goal — are proving rather more expensive than initially forecast. The cost of developing these systems is one thing. Yet the ongoing costs of running them — the computing power the monitoring the security and all the other things that turn a prototype into a working service — seem to be the sticking point. It's easy to get seduced by a promising demonstration yet then the bills start arriving and suddenly the sums look rather less appealing. This matters because we're talking about real money and real business decisions. If companies aren't accurately predicting the true cost of these systems, they risk overspending, misallocating resources, or even abandoning projects altogether. It's another example of the broader challenge of managing expectations around new technologies. The promise of autonomy and independent action is alluring yet it's easily offset by the rather mundane realities of budgets and operational overheads. It also raises a question about accountability, doesn't it? If these artificial intelligence agents are making decisions that affect the bottom line who is ultimately responsible when those decisions turn out to be expensive? Perhaps it's time to start thinking about a kind of financial audit for artificial intelligence systems to ensure they're not running up bills faster than they're generating value. Otherwise, we might find that these autonomous agents are simply agents of chaos, financially speaking. The devil, as ever, is in the details. We're seeing this play out in other ways too. Nvidia, for example, has just released a new open-source tool called AITune, intended to make it easier to deploy deep learning models. The challenge here is that training a model and actually running it efficiently in the real world are two very different things. AITune aims to automate that process, figuring out the fastest way to run a given model. The potential benefit is a significant saving of time and effort, particularly in ensuring the model continues to perform well after adjustments. Yet it remains to be seen how effective it will be in practice and how deep its optimisations can go. It's another example of the industry grappling with the practical realities of deploying these systems at scale. The focus is shifting from simply building models to actually making them work efficiently and cost-effectively. We've seen a lot of similar promises over the years, of course. One hopes that AITune is less like snake oil and more like a useful spanner in the toolbox. It will be interesting to see if this helps to democratise artificial intelligence development. or merely concentrates power in the hands of those with the resources to implement it. All of this, of course, is happening against a backdrop of increasing regulatory scrutiny. Washington is considering regulating artificial intelligence more closely, driven by worries about the technology's rapid advancement and potential risks. The discussions involve government officials, financial institutions, and those leading the artificial intelligence industry. The aim is to curb what's described as "dangerous" advancements, though the specifics remain somewhat unclear. This isn't about preventing innovation altogether, yet rather about addressing the potential for misuse and unintended consequences. Think of it as trying to put guardrails on a very fast-moving vehicle. The implications are wide-ranging. These artificial intelligence systems are moving beyond novelty and could significantly impact the economy, security, and other aspects of society. The challenge for policymakers is to create regulations that address risks without stifling progress. It's a delicate balance, and one that's becoming increasingly urgent as the technology becomes more sophisticated. This reflects a broader pattern we're seeing: a growing recognition that artificial intelligence isn't. just a technological marvel yet a powerful tool that needs to be managed responsibly. It's the same conversation we had with the internet itself, two decades ago, and we all remember how well that went. Of course, whether these deliberations will lead to effective regulations or just another layer of bureaucracy remains to be seen. Sometimes the best intentions pave the road to, well, you know. The tension between technological advancement and societal well-being is also playing out in less obvious ways. Take the Philippines, for example. They're now using social media monitoring to prepare for natural disasters. Essentially, government agencies are tracking online conversations to get a better sense of what's happening on the ground during emergencies. By analysing posts comments and messages on platforms like X Facebook and even local chat apps they can understand what people are experiencing in real time. It's a digital early warning system. By identifying spikes in mentions of flooding power outages or requests for help authorities can get a faster and more detailed picture of the situation than traditional reporting methods might offer. This matters because the Philippines is particularly vulnerable to typhoons, earthquakes, and other calamities. Quick and accurate information is vital for effective disaster response. If authorities can anticipate needs and deploy resources more efficiently based on social media insights the potential to save lives and reduce damage is significant. It also raises a broader point about the role of technology in crisis management. That said, relying on social media isn't without its risks. Misinformation can spread rapidly, and verifying the accuracy of user-generated content is a challenge. Moreover, not everyone has equal access to the internet, which could skew the data and leave some communities underserved. As a result, while the ambition is laudable, it's a reminder that even the most well-intentioned applications of technology can have unintended consequences. And speaking of unintended consequences, let's not forget the ongoing efforts to improve the efficiency of these systems. One of the persistent challenges in applied machine learning is getting those clever powerful models out of the lab and into the real world. It turns out that a technique called knowledge distillation is gaining traction. as a way to streamline the deployment of complex artificial intelligence systems. The core problem is that some of the most accurate artificial. intelligence models are actually ensembles: collections of many individual models working together. While these ensembles can achieve impressive results, they are often too computationally intensive and slow for practical use. Knowledge distillation offers a workaround. Think of it as training a smaller simpler model — the student — to mimic the behaviour of the larger more complex ensemble — the teacher. The student model learns to replicate the teacher's predictions, yet with far fewer computational resources. This matters because it could unlock the potential of advanced artificial intelligence in areas where speed and efficiency are paramount. Consider applications like real-time fraud detection, autonomous vehicles, or even just speeding up responses from artificial intelligence assistants. The ability to distill the knowledge from complex ensembles into smaller more manageable models could significantly broaden the applicability of artificial intelligence in these domains. It also touches on questions of energy consumption and sustainability. If we can achieve similar results with smaller models, we can reduce the carbon footprint of artificial intelligence systems. Yet here's the rub: I wonder if we're simply finding more efficient ways. to perpetuate the same inherent biases that are already baked into these systems. After all, if the teacher model is flawed, the student will likely inherit those flaws, albeit in a more streamlined package. The efficiency savings are welcome, yet not if it simply amplifies existing problems. Finally, let's touch on the ongoing quest for control. There's been some movement in the world of local artificial intelligence, specifically around controlling how these systems interact with your own machines. Someone's put together a system called OpenClaw which aims to let you run artificial intelligence agents entirely on your own hardware with a strong emphasis on security. What does that actually mean? Think of it as a sandbox for artificial intelligence. Instead of sending your data to some cloud server to be processed, this system lets you keep everything on your computer. OpenClaw tries to ensure that the artificial intelligence agent can only access the resources you specifically allow it to and that it can only execute commands in a controlled environment. It's about creating a walled garden where the artificial intelligence can play without messing with your prize-winning roses or worse your financial records. A key piece here is limiting network access as a result that the artificial intelligence can only talk to itself and only after it proves it is who it says it is. The real-world impact here is about regaining some control. For many, the idea of handing over sensitive data to third-party artificial intelligence services is a non-starter, for perfectly rational reasons. If you're dealing with confidential information or you simply value your privacy being able to run artificial intelligence locally and know exactly what it's doing is a significant advantage. It also raises the stakes for the artificial intelligence industry itself. If these kinds of local secure systems become more popular it could shift the balance of power away from the big cloud providers and back towards individual users. Of course, building and maintaining a secure system like this is not trivial. It requires a certain level of technical expertise, and there's always the risk of misconfiguration or undiscovered vulnerabilities. After all a locked door only keeps honest people out and sufficiently motivated hackers will find a way around even the most sophisticated security measures. The question is always whether the added security is worth the added complexity. And that, as ever, remains to be seen. It's a recurring theme, really, this tension between promise and peril, between innovation and unintended consequences. We're constantly being sold the dream of a better future powered by technology yet the reality is always messier more complicated and more demanding of our attention. And that's before we even get to things like Alibaba's new VimRAG system designed to help artificial intelligence navigate and understand visual data like images and videos. The core issue it addresses is this: while current artificial intelligence models are quite adept at processing text they often struggle when faced with large amounts of visual information. Think of trying to find a specific scene in a movie or identify a particular object in a series of photographs. Current artificial intelligence techniques can become overwhelmed by the sheer volume and complexity of visual data. VimRAG attempts to solve this by organising visual information into a "memory graph." This graph structure helps the artificial intelligence sift through the data more efficiently focusing on the most relevant pieces of information while maintaining context. The potential impact here is significant. If successful, VimRAG could improve the ability of artificial intelligence to perform tasks like video analysis, image recognition, and even robotic navigation. This has implications for surveillance, autonomous vehicles, and various other areas. Efficiently handling visual data opens up new commercial possibilities and gives a potential competitive edge to those who can manage it best. Yet it's worth remembering that new artificial intelligence frameworks appear almost daily. The real test will be whether VimRAG truly delivers on its promise of improved efficiency and accuracy or whether it merely adds another layer of complexity to an already complicated field. One might even ask whether we are not simply piling on more complexity to solve a problem. that requires a more fundamental rethinking of how artificial intelligence processes visual data in the first place. It's all a bit like adding another wing to a house that's already structurally unsound, isn't it? Anyway, that's the main section done. Another week, another deluge of announcements. Sorting signal from noise seems more crucial than ever. If you'd like a concise daily briefing on the important developments, you can find it at jonathan-harris dot online. And for those wanting a more considered perspective my book "artificial intelligence in Education: Reimagining Learning for Every Student," is available at jonathan-harris dot online slash ebooks slash ai dash in dash education dash reimagining dash learning dash for dash every dash student slash buy dash now. You'll find less hype, more understanding. That's your lot for this week's Turing's Torch. If you want the daily brief, head to jonathan-harris dot online. Same time next week — try not to believe the press releases.