Turing’s Torch AI Weekly — 2026-05-01

Full Episode Transcript

Good afternoon. It's sunny in London today, which is a pleasant surprise given the usual dreariness this time of year. This week, we turn our attention to a relevant thought from Alan Turing: "There is, however, one consideration which we must not overlook." In the context of artificial intelligence, this serves as a timely reminder to separate signal from noise amidst the cacophony of claims and counterclaims surrounding advancements in the field.

With every new announcement promising groundbreaking progress, it's vital to maintain a discerning eye, sifting through the clutter to discern what truly matters. As we delve into this week's developments, let's approach the so-called breakthroughs with a critical lens, assessing their real impact and relevance. This is Turing's Torch: Artificial Intelligence Weekly — the bits that matter, minus the hype.

This week, we're seeing increased discussion of ‘agentic artificial intelligence’ – artificial intelligence systems designed to operate largely on their own, making decisions and taking actions without constant human direction. The core idea isn't entirely new; we've had systems that automate tasks for years. The difference lies in the scope and the level of independence.

Think of a self-driving car, but instead of just navigating from A to B, it also decides where A and B should be. Based on its own analysis of your schedule, traffic patterns, and your preferences gleaned from your digital footprint. Agentic artificial intelligence aims to create modular agents, each responsible for a specific aspect of a larger task, allowing for greater flexibility and efficiency. These agents can collaborate or operate independently to achieve shared goals.

The potential impact is considerable. In theory, these systems could revolutionise everything from supply chain management to scientific research, handling complex tasks with speed and efficiency beyond human capability. Imagine an artificial intelligence system that not only diagnoses diseases but also designs and implements personalised treatment plans, coordinating with various medical specialists and adjusting the plan based on real-time patient data. Or picture a financial system that not only detects fraud but also proactively manages your investments, anticipating market trends and adjusting your portfolio accordingly.

This autonomy comes with significant caveats. How much control should we relinquish to these systems? How do we ensure they align with our values and priorities? What happens when their decisions have unintended consequences? As artificial intelligence systems become more sophisticated and autonomous, the potential for errors, biases, and even malicious use increases exponentially. We've already seen examples of artificial intelligence algorithms perpetuating discrimination and making flawed judgements, so granting them greater autonomy without addressing these underlying issues could amplify these problems.

It appears we are embracing the technology first and asking questions later. The current landscape feels a bit like handing over the keys to a high-performance vehicle without first. Teaching the artificial intelligence to drive responsibly or even checking if it has a valid licence. The allure of efficiency and innovation is strong, but we must ensure that progress doesn't come at the expense of accountability and control. Perhaps a little less excitement and a little more careful consideration would be prudent at this stage.

In other news, Anthropic and OpenAI, the firms behind the Claude and GPT families of systems, almost simultaneously released new artificial intelligence models, with Anthropic releasing Opus 4. 7 and OpenAI following quickly with GPT-5. 5. The significance of point releases like these is, of course, debatable. We are told that these new versions are improvements on their predecessors, offering enhanced capabilities and features.

In practice, that usually boils down to marginal gains in specific tasks, perhaps faster processing speeds or a slightly improved ability to handle complex prompts. The marketing departments will, naturally, present these incremental steps as revolutionary leaps forward. The real impact of these releases is more about market positioning than genuine technological advancement. Both Anthropic and OpenAI are locked in a fierce battle for dominance in the artificial intelligence space.

Releasing new models, even if they are only incremental improvements, allows them to maintain a presence in the headlines, attract investment, and create the impression of relentless progress. It's a race, certainly, but one where the finish line keeps moving and the definition of "winning" is constantly being redefined. The question of which model is "better" is, inevitably, a matter of opinion. Benchmarks and performance indicators offer some objective data, but the true test lies in real-world applications.

How do these models perform when faced with the messy, unpredictable demands of actual users? And, perhaps more importantly, how do they compare in terms of cost and accessibility? A model that is marginally better but significantly more expensive may not be the best choice for many users. This constant cycle of release and upgrade raises a broader question about the sustainability of the current artificial intelligence development model.

Are we truly making meaningful progress, or are we simply chasing diminishing returns? And at what cost? The environmental impact of training these massive models is considerable, and the resources required to develop and maintain them are vast. One wonders if the benefits justify the costs, particularly when the improvements are often so marginal. The hype is, as ever, considerably ahead of the reality.

It's worth remembering that the artificial intelligence landscape is littered with models that were once hailed as groundbreaking but have since faded into obscurity. The excitement surrounding these new releases will, inevitably, subside, and the true value of these models will only become clear over time. Let's see if they can actually deliver on the promises that are now being made.

On a slightly different note, there's been chatter recently about two competing approaches to how artificial intelligence agents, or chatbots, are constructed, with some people suggesting that one must choose between them. The reality, as is so often the case, is rather more nuanced. The two methods, known as MCP and Agent Skills, are not really direct competitors at all.

Agent Skills, in essence, are pre-written prompts that can be loaded into a system as needed. Think of them as modular conversational snippets, designed to respond to specific user inputs or queries. They're reactive, in that they wait for a trigger before springing into action. MCP, on the other hand, seems to refer to a more structured and comprehensive set of functionalities. It's likely a broader framework that encompasses a wider range of tasks and capabilities.

The idea that these are somehow mutually exclusive is, frankly, a bit silly. It's akin to arguing over whether one should use a screwdriver or a hammer. They're both tools, but they serve different purposes. Conflating them leads to a misunderstanding of their potential applications. This sort of thing happens all the time in the tech world, of course.

The urge to frame everything as a battle, a zero-sum game, often trumps any real understanding of the underlying technology. This matters because it affects how we think about, and ultimately use, these artificial intelligence systems. If we're caught up in a false dichotomy, we risk overlooking the potential for combining these approaches to create more powerful and versatile agents. It also distracts from the more important questions, such as how these systems are being used, who controls them, and what impact they're having on society.

It's easy to get bogged down in the technical details, but we mustn't lose sight of the bigger picture. The real competition isn't between these two specific methods, but rather between those who seek to use artificial intelligence for the benefit of all. And those who are simply chasing the next big payday. It's tempting to see these supposed rivalries as a sign of progress, of innovation pushing the boundaries. But sometimes, it's just marketing dressed up as engineering.

And as we all know, a good salesman can sell you anything, even a solution to a problem you didn't know you had. The ability to discern substance from spectacle is becoming increasingly important.

Elsewhere, researchers have been looking at ways to shrink artificial intelligence models, specifically for use in retail environments. The idea is to make these models small enough to run on the sort of limited-power devices you might find in a shop, rather than relying on expensive cloud computing. When they talk about compressing these Long Short-Term Memory models, or LSTMs, what they're really talking about is reducing the amount of computing power and memory these things need.

Think of it like zipping a large file on your computer; you make it smaller so it's easier to store and faster to send. In this case, the 'file' is an artificial intelligence model that predicts things like product demand. For a retailer, that means trying to work out how much of a particular item they're likely to sell. So they can manage their inventory and make sure the shelves are stocked appropriately. The reason this matters is quite simple: cost.

Smaller businesses, in particular, often can't afford the infrastructure needed to run these complex models in the cloud. If you can run the model on a smaller, cheaper device in the shop itself – what they call 'edge deployment' – you cut costs and potentially improve efficiency. But there's a trade-off, because compressing these models can reduce their accuracy. If your demand forecasts are off, you end up with too much or too little stock, which hits profits.

This push to make artificial intelligence models smaller and more efficient reflects a broader trend. We're seeing a lot of effort being put into making artificial intelligence more accessible and affordable, whether that's for retail or any other sector. There's a growing recognition that the real value of artificial intelligence isn't just in the cutting-edge research, but in the practical application of these technologies in everyday settings. One might also wonder, of course, whether it's all worth it.

After all, a slightly less accurate prediction, made cheaply, is perhaps no better than a well-informed guess made by an experienced manager. Still, the pressure to appear technologically advanced is strong, and that's something to consider as we continue to integrate these systems into our lives.

In the meantime, many companies are discovering that their data infrastructure isn't up to the task of supporting artificial intelligence. It seems the impressive consumer-facing artificial intelligence applications have lulled some executives into a false sense of security. Because deploying artificial intelligence effectively inside an organisation requires a far more robust data management system than many currently possess. What this really means is that the hype around artificial intelligence has outpaced the practical groundwork needed to make it work.

The consumer applications, the chatbots and image generators, operate on carefully curated datasets, often built from readily available information. Businesses, on the other hand, need artificial intelligence to work with their own internal data, which is frequently a mess of inconsistent formats, siloed databases, and questionable accuracy. The problem isn't the artificial intelligence itself, but the quality, accessibility, and integration of the data it relies on.

The stakes are considerable, naturally. Companies hoping to gain a competitive edge through artificial intelligence-driven insights, automation, or personalised customer experiences are finding themselves blocked by their own outdated data infrastructure. This translates directly into lost opportunities, wasted investment, and a slower pace of innovation. The companies that get their data sorted will be the ones who actually benefit from artificial intelligence, while the rest will be left behind.

This situation also highlights a broader trend: the increasing importance of data governance and management in the age of artificial intelligence. As artificial intelligence systems become more powerful and pervasive, the need to ensure the quality, security, and ethical use of data becomes ever more critical. We have seen this with the concerns around transparency and bias in artificial intelligence algorithms, and it all comes back to the data they are trained on. The clamour for artificial intelligence at scale is exposing the weaknesses in our existing data practices and forcing businesses to confront the unglamorous reality of data reform.

It is tempting to see this as yet another example of technology vendors overpromising and underdelivering, but the truth is that artificial intelligence is only as good as the data it consumes. Perhaps some chief information officers should have spent less time chasing the latest artificial intelligence fads and more time cleaning up their databases. One can see how a bit of data housekeeping might just be the unsung hero of the artificial intelligence revolution.

Finally, Yolanda Gil, a professor at the University of Southern California, recently gave a talk about using artificial intelligence to make workflows better. That's to say, she discussed how artificial intelligence might improve the way tasks are organised and completed in various professional settings. When academics talk about workflows, they're really talking about the nuts and bolts of how work gets done.

It's about mapping out each step in a process, from the initial trigger to the final outcome, and then looking for ways to make it more efficient. The promise of artificial intelligence in this context is that it can automate repetitive tasks, analyse data to identify bottlenecks, and even predict potential problems before they arise. So, in theory, artificial intelligence could take over some of the more mundane aspects of a job, freeing up human workers to focus on more complex or creative work.

The real-world impact of this, should it come to pass, would be felt across numerous industries. Businesses are always looking for ways to cut costs and increase productivity, and artificial intelligence offers the tantalising possibility of doing both. However, it also raises questions about job displacement, as some roles may become obsolete if artificial intelligence can perform them more efficiently. There's also the issue of control.

Who decides which tasks are automated, and how do we ensure that these systems are fair and unbiased? These are crucial questions that need to be addressed if we're to avoid creating new problems in the pursuit of efficiency. This focus on efficiency also fits into a broader trend we've been observing: the relentless pursuit of automation in all aspects of life.

From self-checkout tills at the supermarket to algorithms that manage our finances, we are constantly told that technology can make our lives easier and more productive. Yet, this pursuit often comes at the expense of human connection and critical thinking. We must be careful not to blindly embrace automation without considering the potential consequences. One might also ask whether this drive for efficiency is really about making our lives better or simply about increasing profits for corporations.

After all, the benefits of artificial intelligence-driven workflow improvements are likely to accrue disproportionately to those at the top, while the risks are borne by everyone else. One can easily imagine a scenario where artificial intelligence is used to squeeze even more productivity out of workers, leading to increased stress and burnout. And that's before we consider the potential for artificial intelligence to be used for surveillance and control in the workplace, monitoring every keystroke and movement of employees. It seems unlikely that this particular torch will illuminate a path to any worker's paradise. For now, it's an idea worth following, but only with a healthy dose of scepticism.

Well, another week gone, another deluge of developments. Sorting signal from noise seems more vital than ever. If you would like a curated summary of the day's artificial intelligence news, do sign up for the daily briefing at jonathan-harris dot online. And while you're there, in the eBooks section, you'll find my own contribution to the debate, "Digital Diagnosis: How artificial intelligence is Revolutionising Healthcare." 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.