John Pillar

Half-way trough my learning journey with Module 3 of the Oxford AI Programme now complete. This module focused on a literally “hands-on” exercise - how neural networks actually recognise images, using handwritten digit recognition as the core example. A key learning was how Convolutional Neural Networks build understanding layer by layer. In practice, when recognising handwritten numbers the model is looking for “patterns” and breaks them down into the smallest units /“tokens” i.e: • Early layers detect simple strokes - vertical or horizontal lines and diagonal edges that make up pen marks. • Intermediate layers combine those strokes into curves, corners, and loops (for example the loop in a “6” or the arc in a “3”). • Later layers learn spatial arrangements of these parts - such as distinguishing a “6” from a “9” by whether the loop sits above or below the tail. Architectural features also play a critical role: • Sparse connectivity means filters look only at small regions of the image at a time - just like our eyes focus on local shapes. • Pooling compresses information while preserving important patterns, making the model less sensitive to small shifts in where the digit appears. The result is a system that can recognise the same number even when written in slightly different styles, sizes, or positions. Scored 92.9% so I’m pretty happy given I have been battling a nasty cold (inherited now across the entire family) and in deep need of sleep! On to the next module. 🚀 #Artificialintelligence #DeepLearning #NeuralNetworks #OxfordAIProgramme

14

The more time I spend with teams experimenting with AI in financial services, the less I obsess over which model they’ve chosen – and the more I look at how it plugs into the way people actually work. The critical bit isn’t the architecture diagram. It’s how analysts describe a problem, see the impact of a change, sign it off, and defend it to a regulator six months later. When you zoom out, most AI setups fall into one of two patterns: • AI that spits out finished logic • AI that helps humans assemble and stress-test their own logic Call them autogeneration vs scaffolding if you like – but the important part is how they behave over time. • Autogeneration tends to look cheap at the start and expensive later. Every new typology, every tweak, sends you back to the AI. The real “know-how” sits outside the team. • Scaffolding often feels like more work at the start and cheaper later. Analysts see the data, the trade-offs, the impact – and they get better every time. After a while, the AI is there to accelerate thinking, not to replace it. You’re not just choosing a UX pattern. You’re choosing whether your team compounds dependency or capability over the next two years. So if you’re responsible for AI in a regulated fintech or bank, I don’t think the most useful question is: “How impressive is our AI demo?” It’s closer to: “Could my analysts, in plain language, explain why this logic exists and why these thresholds are where they are?” If I were buying (or designing) AI for a regulated workflow, I’d be probing things like: • As we use this more, do our analysts become more self-sufficient, or more reliant on the product? • In a real audit, who would you rather point to, to defend decision - an expert practitioner, or a slide about AI strategy? • Are we optimising for screenshots and slideware, or for durable expertise inside the team? At Fortify we’ve picked our side quite deliberately: use AI for the heavy lifting – mapping patterns, simulating impact, surfacing trade-offs – and keep humans in charge of the logic, thresholds and rationale. Every time we ship something in Fortify AML, we ask: “Does this make analysts stronger on their own, or more dependent on us?” If it’s the latter, it doesn’t ship. If you’re wrestling with similar choices in your own stack – or just want to see what a scaffolding-first approach looks like in practice – I’m always happy to compare notes or walk through how we’re doing this in Fortify AML.

18

Everyone has outages. Cloud providers. On premise. The Bank of England has already put a marker about business continuity. Without doubt many regulators will be considering what it means. Banks will need to run multi-cloud which will totally change the ability to run native products.

33

11 Comments

I’m a firm believer that despite all of the amazing developments we’ve seen in recent years, at the coalface of real enterprise applied AI we’re still in an evolution and not a revolution. This is good news as it means many of the hard-won learnings from the last “ML” evolution still apply. We initially experimented with ML, got disappointed that it didn’t solve all our problems, slowly learned to operationalise it (MLOps) and then saw consistent value streams open up. The same thing will happen with AI now, so start trying to think what skills and knowledge map over. I’ve been thinking about this as I start scoping out the next book and write the next season of lectures. If you can map the evolutions, you can start thinking “what do I keep, change or ditch?” And you will move so much faster. #GenAI #LLMOps #MLOps #ML #AI

63

4 Comments

It’s been really interesting watching how the whole idea of a Semantic Layer has gone from “No, we don’t need that” to “Surely AI can figure that out”, to “Wait… we definitely need this.” The organisation has basically travelled that whole arc over the last 18 months. It was a passionate bet myself, Jonathan Tierney and Michael Perratt started pushing nearly 18 months ago, so watching discussions evolve like they have, I must admit, has been tinged with a bit of pride... And now? We’re kicking off broad initiatives and having serious conversations internally and with leading vendors about what the Semantic Layer actually is, who owns which bits, and how it all fits together. Clue: Foundational Semantics and Semantic Signals are not the same thing - and understanding the nuances of that difference is where things really start to click. The most positive outcome is seeing broader realisation you need both if you want things like GenBI / Conversational Analytics to be reliable and actually useful at scale. AI isn’t some mystical creature that flawlessly just intuits your business logic. It needs the foundation in place: the solid, agreed data models and definitions - and then it can layer on the more dynamic stuff like signals: hints, context, intent, instructions, behaviour. That’s where interactions start to feel intuitive, adaptive, even personalised. And that’s why the conversations are suddenly getting very real: • Who actually owns the Foundational Semantics? • Where do Semantic Signals come from? • Who curates and trains the signals? • How do we iterate fast on signals without breaking the core definitions? • How do we serve foundations and signals through one governance layer regardless of the consuming tools? • And how on earth do we get a production-grade Semantic Layer working across teams that all speak slightly different and often validly different “semantic languages”? My view: the Semantic Layer is about to shift from being seen as a technical component to becoming an organisational capability. Something shared across data teams, analytics teams, domain experts, platform teams and the AI enabled stacks. It’s taken a bit of time to get here, but the shift seems obvious now. AI won’t replace the Semantic Layer - it stands on top of it. And the organisations that clock this early enough are the ones that will get real, scalable value from generative analytics solutions… not just cool demos. More to come as we keep shaping this, while closely watching other trends in innovation in this space, but the momentum is definitely growing…

47

6 Comments