Good afternoon. It's partly cloudy in London, which seems fitting as we navigate the murky waters of artificial intelligence. Alan Turing once said, "I'm not telling you it's going to be easy. I'm telling you it's going to be worth it." This week we'll explore that sentiment as we sift through the endless chatter surrounding artificial intelligence—attempting to separate the signal from the noise. There's a great deal of excitement in the air yet history shows us that breakthroughs can often be overhyped leaving us to question what truly matters. As a result, let's get into the substance and see what's worth our attention. This is Turing's Torch: Artificial Intelligence Weekly — the bits that matter, minus the hype. OpenAI, the people behind ChatGPT, have put forward the idea of a four-day working week, along with a tax on robots. Now before we all start picturing long weekends paid for by levies on artificial intelligence let's have a look at what's actually being suggested. The idea as I understand it is that artificial intelligence will take over certain tasks which will raise productivity and allow for shorter working hours. Taxing artificial intelligence, meanwhile, would offset potential job losses and redistribute wealth. On the surface, it's an appealing notion, and certainly topical. Yet the implications are rather more complex. Who decides what constitutes an artificial intelligence worthy of taxation? How do we avoid stifling innovation with such a tax? And what about the inevitable unintended consequences for businesses and workers? These questions, naturally, remain largely unanswered. It's tempting to dismiss these proposals as pure speculation, the kind of thing dreamed up after one too many espressos. Yet they do point to a broader conversation we need to have about the future of work in an age of increasingly capable artificial intelligence not to mention current political narratives around taxation and inequality. Perhaps this is a genuine attempt to engage with these issues. Or perhaps it's a clever marketing ploy designed to keep OpenAI in the headlines and burnish their credentials as a thought leader. Either way, it's a reminder that even the most sophisticated algorithms can't solve fundamental questions about fairness and economic justice. And while the prospect of a three-day weekend is certainly enticing I suspect we're still some way off from robots footing the bill. That's something to consider given that the latest thinking suggests that simply bolting artificial intelligence onto existing business processes is a recipe for disappointment. The idea is that these artificial intelligence agents, unlike traditional software, learn and adapt as they go. They can handle entire workflows, interacting with data and even other agents, without constant human intervention. Yet to really make the most of them businesses need to redesign their processes from the ground up putting the artificial intelligence agent at the centre. Instead of trying to force artificial intelligence to fit into rigid outdated structures the structures themselves should be built to allow the artificial intelligence to flourish. This means creating environments where these agents can use their learning capabilities to improve efficiency and drive innovation. We're talking about organisations where processes evolve seamlessly, adapting to new data and changing circumstances without the usual bureaucratic delays. The potential benefits are considerable: improved productivity, faster decision-making, and fewer errors. Of course, it's not just about technology. It requires a fundamental shift in mindset. Leaders need to be willing to let go of old habits and assumptions. This connects to a broader trend: we're often told technology will solve our problems yet it frequently requires deeper organisational and cultural change to actually deliver on its promise. It's always tempting to think that new technology can fix broken systems, yet often it simply exposes pre-existing cracks in starker relief. One wonders though if the real problem isn't the processes themselves yet the inherent limitations of the artificial intelligence we're currently working with. Perhaps we're simply trying to build grand cathedrals with glorified spreadsheets. We shall see. One thing we do know is that artificial intelligence is no longer just about answering questions it's increasingly about making decisions and taking actions independently. This shift towards autonomous artificial intelligence agents is raising concerns about how these systems are governed and controlled. When we talk about artificial intelligence agents, we're not just referring to chatbots. These are artificial intelligence systems designed to perform specific tasks within an organisation, often with minimal human oversight. This might involve planning project timelines, making financial recommendations, or even automating customer service interactions. The key is that they operate with a degree of independence, making choices and executing actions without constant human intervention. The implications of this are significant. As artificial intelligence agents take on more responsibility, the question of accountability becomes paramount. Who is responsible when an artificial intelligence makes a poor decision that negatively impacts individuals or the organisation? How do we ensure that these systems align with human values and societal norms? These are the questions that effective artificial intelligence governance seeks to address. It's not simply about ensuring the artificial intelligence is technically accurate it's about establishing frameworks that guide their behaviour and prevent unintended consequences. This all plays into the broader discussion around artificial intelligence regulation. As artificial intelligence becomes more integrated into our daily lives, the need for oversight and control becomes more pressing. We've seen similar debates in other areas of technology, particularly around data privacy and algorithmic bias. The challenge with artificial intelligence is that it's evolving as a result rapidly that regulatory frameworks struggle to keep pace. This can lead to a situation where artificial intelligence systems are deployed without adequate safeguards, potentially creating new risks and challenges. One might observe that the rush to automate and delegate to. machines often outstrips our ability to consider the ethical and societal ramifications. It is, perhaps, a reflection of our tendency to focus on technological possibilities while overlooking the human element. And that tension will only increase in the months ahead even if Microsoft has released a set of tools designed to keep artificial intelligence systems secure while they're actually running. This is aimed at addressing the growing unease about artificial intelligence language. models that are now writing code and gaining access to company networks. What this means in practice is that as artificial intelligence systems become more autonomous and integrated into business operations the risks of misuse or malfunction increase. This toolkit is essentially a security overlay intended to ensure these artificial intelligence agents stick to predefined security protocols. It's being offered as open-source, meaning anyone can modify and improve it. The implications are significant. As artificial intelligence capabilities rapidly expand, as a result too does the need for effective governance and security. Companies can't simply hope for the best; they require frameworks to ensure artificial intelligence tools operate safely within their networks. This toolkit could be a step in the right direction yet whether it can keep pace with the rate of artificial intelligence development is another question entirely. It's worth remembering that the rush to integrate artificial intelligence into everything. from customer service to data analysis often outpaces the development of safeguards. We're essentially building faster cars without necessarily improving the brakes at the same rate. Ultimately, it's a bit like locking the stable door after the algorithm has bolted. The real test will be whether such measures can adapt quickly enough to the ever-evolving threats in the digital landscape. It's also worth remembering that Meta the social media giant recently paused its work with a company called Mercor because of a security problem linked to an open-source project known as LiteLLM. The whole episode shines a light on the often-overlooked yet crucial area of data management in artificial intelligence. What we're talking about here is the plumbing of artificial intelligence if you like: the data and the processes that feed and train these models. Companies are understandably keen to jump on the artificial intelligence bandwagon yet sometimes they don't fully check out the companies that supply the data. The Meta-Mercor situation shows that even big, well-funded organisations can run into trouble if they aren't careful about vetting their data sources. A security breach doesn't just mess up the data; it also damages the trust that people have in these companies. It's a harsh reminder that the risks of sourcing data aren't just technical; they can seriously harm a company's reputation. More generally, businesses are starting to understand that their artificial intelligence projects are only as reliable as the data they're built on. As artificial intelligence becomes more integrated into decision-making, knowing the risks associated with data providers is essential. Companies need to be more thorough when assessing their partners, making sure they have strong security measures in place. This is particularly relevant in the context of increasing regulatory scrutiny around data privacy and usage. The consequences of getting it wrong can be significant, not just in terms of fines yet also in terms of market perception. Perhaps the core issue here is a kind of gold-rush mentality where the perceived urgency of adopting artificial intelligence solutions overshadows the more mundane yet critical aspects of data governance. It is worth remembering that algorithms are only as good as the information they're fed and a flashy artificial intelligence interface can't hide the flaws in a dirty data supply chain. And that's a timely reminder to look under the bonnet of any artificial intelligence system you encounter especially given that Boomi the integration platform company is now saying that something they call "data activation" is essential for artificial intelligence to work properly. Apparently, they believe that the biggest problem in the near future won't be flawed algorithms or stupid artificial intelligence agents. Instead, it will be the poor state of the data itself. As a result, what does "data activation" actually mean? Well, in essence, it's about ensuring data is accessible, consistent, and properly labelled. Think of it as the digital equivalent of organising your sock drawer. If your data is scattered across various applications inconsistently formatted or simply incomprehensible then any artificial intelligence system trying to use it will struggle. It's like trying to bake a cake with a recipe written. in a language you don't understand and ingredients stored in unmarked containers. The end result is unlikely to be edible. The significance of this is that companies are pouring vast sums of money into artificial intelligence often without considering the underlying data infrastructure. If the data is a mess, those investments could easily turn into expensive failures. The focus tends to be on the whizzy algorithms and the promise of transformative applications yet the unglamorous reality is that artificial intelligence is only as good as the data it's trained on. This is particularly relevant in regulated sectors where data quality and provenance are critical for compliance. It also fits into a broader pattern we're seeing which is a growing recognition that the foundations of artificial intelligence are just as important if not more as a result than the headline-grabbing advancements in model design. It's a bit like building a skyscraper on a swamp: impressive to look at, yet ultimately unsustainable. Of course, there's a certain irony in a company that sells data integration solutions highlighting the importance of… well, data integration. It's a bit like a locksmith warning you about the dangers of flimsy locks. Still, the underlying point is a valid one. Unless businesses get their data in order, their artificial intelligence ambitions are likely to remain just that: ambitions. And that's worth keeping in mind as the hype cycle continues. One aspect of that hype is the ongoing attempts to improve artificial intelligence's usefulness which centre on giving these systems something resembling a memory. At present, each interaction with an artificial intelligence model is essentially a clean slate. You upload information, ask questions, receive answers, and then the entire exchange is forgotten. This becomes particularly tedious when working with large amounts of data such as extensive codebases or research archives because the artificial intelligence cannot retain context from previous sessions. The idea is that by equipping artificial intelligence with memory layers, it could remember past queries and responses, thus avoiding repetitive work. This matters because the current lack of continuity introduces significant inefficiencies. Users constantly re-upload files and repeat questions, which consumes time and resources. If artificial intelligence could build upon prior interactions, it would drastically reduce redundancy and enhance the overall efficiency of artificial intelligence applications. There is also the promise of a more intuitive and fluid user experience as the artificial intelligence could learn individual preferences and adapt its responses accordingly. As we see more integration of artificial intelligence into complex workflows, the ability to retain context becomes increasingly important. That said, the practical implications of adding memory layers are still uncertain. There is a risk of introducing new complexities and potential biases as the artificial intelligence's memory could be influenced by incomplete or skewed information. Imagine the fun we could have when these systems start misremembering things, like an unreliable colleague at the office. Nevertheless the direction seems clear: artificial intelligence is moving towards becoming more persistent and context-aware even if the latest generation of artificial intelligence currently has a memory problem or rather a lack thereof. While these systems can process vast amounts of information and even demonstrate impressive reasoning skills, they struggle to retain past interactions. In practical terms, this means that these large language models, as they're called, often start from scratch with each new conversation. They don't build on previous exchanges, which can lead to repetitive and ultimately unsatisfying experiences for users. Imagine explaining the same thing over and over again to a colleague who should really know better by now. The ability to remember past interactions is crucial for artificial intelligence to truly function as an intelligent agent. Without it, the potential for meaningful and personalized interactions is severely limited. It's not just about having access to a vast pool of knowledge it's about using that knowledge effectively and retaining relevant information over time. The race is on to equip artificial intelligence with the kind of. memory that allows it to learn and adapt in a more human-like way. This issue of memory also touches on the broader theme of artificial intelligence development and its potential impact on jobs. If artificial intelligence can't remember past interactions, then human workers will still be needed to provide continuity and context. Yet if artificial intelligence can develop a reliable memory, then it could potentially replace certain roles that require these skills. Perhaps the biggest problem with these forgetting machines is the temptation to anthropomorphise them. We are disappointed when they don't remember, because we assume they should. Yet they don't have brains, they have algorithms. And expecting an algorithm to remember is like expecting a toaster to write a poem. It's simply not what it was designed to do. And that rather fundamental limitation reminds us that we are, despite the hype, still in the early stages of this particular journey. One aspect of that journey that's worth noting is that we're seeing increased interest in open-source software for adapting large language models. Instead of building an artificial intelligence model from the ground up which is ruinously expensive developers are using pre-existing models and then tweaking them for specific purposes. The process is called fine-tuning and it's analogous to taking a general-purpose engine and then modifying it for a racing car or a tractor. The base model provides the fundamental capabilities – understanding language generating text – and fine-tuning tailors it to a particular task like drafting legal contracts or perhaps providing customer support. The practical impact is significant. Previously, only large corporations could afford to train these models from scratch. Now, smaller companies and even individual developers can participate, using these open-source libraries to fine-tune models on relatively modest computing equipment. This democratisation is driven by the collaborative nature of open-source development, where improvements are shared and refined by a community of programmers. The proliferation of these libraries also highlights a broader trend: the increasing specialisation of artificial intelligence. We're moving away from general-purpose models and towards systems designed for specific applications. This raises questions about the long-term maintainability of these bespoke systems. Will the open-source community continue to support these libraries as they become more niche? And what happens when the underlying base models are updated, potentially breaking compatibility with the fine-tuned versions? One might argue that relying on the generosity of strangers for critical infrastructure is, at best, a risky proposition. All the same, the ability to fine-tune language models is becoming a crucial skill for anyone working with artificial intelligence. Of course one of the potential applications of artificial intelligence is to make existing tasks more efficient and to that end a new piece of software has been released that aims to automate the process of optimising code for graphics processing units or GPUs. Now for those unfamiliar GPUs are essential for running complex machine learning models yet writing code that fully exploits their capabilities is a notoriously difficult and time-consuming task. This new software, called AutoKernel, uses a large language model to automatically optimise code written in PyTorch, a popular machine learning framework. In essence, it promises to take the burden of manual optimisation off the shoulders of developers. The potential impact here is considerable. If AutoKernel works as advertised, it could significantly reduce the time and resources required to develop and deploy high-performance machine learning applications. This could lead to faster innovation and wider adoption of artificial intelligence across various industries. It also touches on the ongoing debate about automation and its effect on employment. While AutoKernel might not eliminate the need for skilled programmers it could certainly change the nature of their work shifting the focus from low-level optimisation to higher-level design and problem-solving. Of course, the proof will be in the pudding. Early adopters will need to test AutoKernel rigorously to see if it lives up to its claims. And even if it does, there's no guarantee that it will be widely adopted. Developers can be a conservative bunch, often preferring familiar tools and techniques over new and untested ones. One suspects the uptake will hinge not only on its performance, yet also on how easily it integrates into existing workflows. The challenge is always getting people to change their habits. For now it's an interesting development yet let's wait and see if it truly delivers on its promise before we declare it a revolution. Yet if it does work it might accelerate the trend whereby artificial intelligence is changing how we use software moving away from clicking through menus to simply talking to the computer. Instead of navigating a program with a mouse and keyboard, you would use natural language. You would ask for what