From Six Months to Three Hours: How AI Has Shrunk the Attack Window

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: [cambridgespark.com](https://www.cambridgespark.com/?utm_campaign=6955829-Podcast%202024%20-%20Onwards&utm_source=Podcast&utm_medium=RSS) An LLM can give well-formed output 99% of the time and still break the domain's rules in most of it. You want to put an LLM into a real decision process, such as invoice approvals, pricing or clinical recommendations. The answers read well. But are they actually allowed, and do they solve the problem? Professor Alessandra Russo is Professor in Applied Computational Logic at Imperial College London and Director of the Thomson Reuters Imperial Frontier AI Research Lab. She co-directs the UKRI Centre for Doctoral Training in Safe and Trusted AI with King's College London. Her neuro-symbolic AI research combines logic with machine learning to build AI systems with functional guarantees. Jeremy Bradley and Alessandra go inside SEM-CTRL, her team's method for semantically controlled decoding in open weight models. You'll hear how a symbolic checker sits in the decoder and masks invalid tokens before they are ever chosen. Monte Carlo tree search then steers generation towards correct answers. No fine-tuning required. This episode is for data scientists, ML engineers and technical leaders responsible for enterprise AI deployment in regulated or high-stakes settings. It assumes some comfort with how LLMs generate tokens. If you want a light overview of AI trends, this one goes further. ## Chapter Markers 00:00 Why AI should be able to say "I don't know" 00:19 Introducing Professor Alessandra Russo 01:32 What organisations worry about beyond hallucination 04:30 Open weight models versus frontier APIs 08:37 The risks that come with open weights 11:16 Inside the SE-CTRL paper 12:39 Validity versus correctness in the blocks world 16:46 A pricing guardrail that needs memory 19:48 Context-sensitive grammars and answer set grammars 24:39 Where the checker sits in the decoder 28:00 Masking tokens before softmax, and temperature 30:20 Monte Carlo tree search for correct solutions 39:01 Where reward functions come from 40:18 Why a 1B model outperformed a 70B model 46:29 Who writes the grammar: ILASP and LLMs 52:47 Limits: fairness, noise and robotics 55:23 What an NHS trust should do first 59:07 Where to find the research ## Useful Links & Resources - SEM-CTRL: Semantically Controlled Decoding: https://doi.org/10.48550/arXiv.2503.01804 How LLMs and agents can be used as robust decision agents in complex decision problems, where hard safeguards can be put in place to prevent unsafe outcomes and reinforcement learning can be used to iterate towards better solutions. - Answer Set Grammars (AAAI 2019) - ILASP: The complexity and generality of learning answer set programs: https://doi.org/10.1016/j.artint.2018.03.005 How operational examples can be used to learn logical specifications for AI tools. - Thomson Reuters Imperial Frontier AI Research Lab - Cambridge Spark: https://www.cambridgespark.com ## Connect With the Show - Jeremy Bradley on LinkedIn: https://www.linkedin.com/in/jeremy-bradley/ - Cambridge Spark on LinkedIn: https://www.linkedin.com/school/cambridge-spark/ Subscribe to Inside the Algorithm so you never miss an episode. Visit cambridgespark.com to find out how Cambridge Spark can upskill your workforce in data and AI. #InsideTheAlgorithm #EnterpriseAI #NeuroSymbolicAI #MachineLearning #AISafety

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: [cambridgespark.com](https://www.cambridgespark.com/?utm_campaign=6955829-Podcast%202024%20-%20Onwards&utm_source=Podcast&utm_medium=RSS) Attackers now exploit a newly public vulnerability in under three hours. Six months used to be normal. Your developers are shipping code with Claude Code and Codex. Your security teams were trained to look for error messages, and in some companies the AI security budget now sits with a chief AI officer rather than the CISO. Ashish Rajan is CISO at TechRiot and author of AI Security Engineering, published by Wiley. He hosts the Cloud Security Podcast and the AI Security Podcast, and advises Fortune 500 companies on security. Ashish gives you a practical way to decide which AI risks to accept, where to focus first, and what skills your security teams need next. You'll hear why the confidentiality, integrity and availability principles still hold. You'll also hear why a vibe-coded tool needs the same security thinking as any application once it touches production data. This episode is for CISOs, CIOs, Chief Data and AI Officers, and the L&D leaders responsible for workforce upskilling during AI transformation. It suits anyone who wants an honest read on enterprise AI security from someone working inside it. If you want a vendor shortlist, this one isn't for you. ## Key Takeaways - An agent asking for a password throws no error code. Ashish explains why judging intent in plain English breaks the traditional security playbook. - A bank with 400+ applications can't secure them all on day one. Ashish shares how to pick the crown jewels and where compensating controls fit. - Building your own security tool in two hours with Claude Code is easy. Maintaining it after the builder leaves for a better offer is the cost few teams plan for. - A framework like NIST shows you have the basics covered. Ashish argues boards should ask their CISO a harder question: could you patch a critical flaw in under three hours? ## Chapter Markers 00:00 The three-hour exploit window 01:13 Ashish's start: a hacked Hotmail account 03:08 Why he wrote AI Security Engineering 05:28 Rewriting the book when agents arrived 07:02 Launching the AI Security Lab 08:08 First principles versus new AI risks 12:48 Who owns AI security 17:29 Advice for CEOs: protect the crown jewels 20:21 Workforce upskilling and build versus buy 23:52 Cross-functional security teams and harnesses 28:52 Personal apps versus enterprise tools 30:14 Offensive AI and the agentic SOC 33:39 Why frontier red-team models need a harness 35:57 From six months to three hours 37:00 Reasons for hope 37:34 The wake-up call that came and went 38:27 What boards should ask their CISO 41:28 Wrap-up ## Useful Links & Resources - AI Security Engineering by Ashish Rajan (Wiley) - Cloud Security Podcast and AI Security Podcast, hosted by Ashish Rajan - Cambridge Spark data and AI training: https://www.cambridgespark.com - Dr. Raoul-Gabriel Urma on LinkedIn: https://uk.linkedin.com/in/raoulurma ## Connect With the Show - LinkedIn: https://www.linkedin.com/school/cambridge-spark/ - Website: https://www.cambridgespark.com Who owns AI security in your organisation: the CISO, the CIO or a Chief AI Officer? Tell us in the comments. If your team has built its own security harness or tackled agent risk head on, we'd like to hear how it went. Visit cambridgespark.com to transform your workforce with cutting-edge data and AI training. Subscribe so you never miss an episode. #DataAndAIMastery #AISecurity #EnterpriseAI #CyberSecurity #CambridgeSpark

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com In this special compilation episode of Inside the Algorithm, host Jeremy, Chief AI Scientist at Cambridge Spark, brings together the best moments from a season of conversations with leading academics, researchers and practitioners. Guests from aviation forecasting, manufacturing, NHS health modelling and neuro-symbolic AI research take on the same core questions. Hearing them side by side shows how experts from very different fields approach the challenge of taking AI from theory to practice. The episode covers why so many AI projects stall at the prototype stage and where models break when they meet something they've never seen. It also covers how trust is built in systems that can feel like a black box, why the strongest results come from hybrid systems, and what experience still gives you as AI tools take on more of the work. Key Takeaways Production readiness means handling edge cases, unforeseen scenarios and adversarial attacks. It also means securing architecture sign-off and making sure outputs reach the people who need them.Projects succeed when teams understand how a solution will be used and what "done" looks like, and when they have champions inside the client's business.LLMs learn correlations, not the world. Their reliance on text patterns explains both their surprising capability and their strangest hallucinations.Historical data has limits. Events like the pandemic can break the relationships forecasting models depend on almost overnight. Stress testing, what-if scenario planning and human domain expertise are essential safeguards.Trust is built through involvement. Participative modelling earned deep trust with NHS clients in the Midlands, and transferring that trust to new regions proved harder than rebuilding the model itself.Hybrid systems deliver the strongest results. Pairing language models with solvers, simulations and small, fine-tuned specialist models lets each part do the job it's best at.Simplicity and production-first thinking win. A working model in production beats a more accurate one that takes months to build. Explainable approaches often serve clients better.Juniors need the chance to learn the trade. Guests warn about "rubber-stamping" AI-generated code and encourage newcomers to look beyond the hype to the wider history of AI. Useful Links & Resources Inside the Algorithm on the Data and AI Mastery podcast feedInside the Algorithm episodes on the Cambridge Spark YouTube channelCambridge Spark: cambridgespark.com Connect With the Show Cambridge Spark LinkedIn: https://www.linkedin.com/school/cambridge-spark/Cambridge Spark Instagram: https://www.instagram.com/cambridgespark/Cambridge Spark X: https://x.com/CambridgeSparkHost Jeremy Bradley on LinkedIn: https://www.linkedin.com/in/jeremy-bradley/ Which of these clips landed hardest for you: the prototype trap, the pandemic story, or the case for pairing LLMs with solvers? Tell us in the comments, and let us know which guest you'd like to hear a full episode with again. Visit cambridgespark.com to learn how Cambridge Spark can upskill your workforce in data and AI. #InsideTheAlgorithm #EnterpriseAI #DataEngineering #MachineLearning #AIDeployment

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com In financial services, getting AI wrong doesn't just mean a bad headline. It means real customers losing real money at the moments that matter most. If you're trying to balance AI innovation with regulatory scrutiny, you already know the tension: move too fast and risk a compliance failure, move too slow and watch competitors who don't hesitate pull ahead. Edmund Towers leads Advanced Analytics & Data Science at the Financial Conduct Authority, shaping the regulator's approach to AI, financial crime detection and consumer protection. Before joining the FCA, he delivered financial services transformation programmes as a manager in Accenture's Financial Services Change Practice. In this compilation episode, he's joined by Jessica Rusu, the FCA's Chief Data, Information and Intelligence Officer, alongside senior data leaders from Santander UK, Aviva and TalkTalk. You'll hear how the FCA thinks about AI governance without resorting to tick-box regulation, and how banks and insurers turn that clarity into production systems. Sarah Self at Aviva explains how a summarisation tool cut claims hold times by more than 50%. Kevin Cassar walks through building an AI triage system in health insurance that improved both customer outcomes and operational efficiency. This episode covers the FCA's AI Live Testing programme, the Consumer Duty as the policy anchor for responsible AI, and how Santander UK's Luke Pearce secured C-suite sponsorship to move faster on generative AI. It's built for data and AI leaders in regulated industries who need to make the business case for AI while keeping compliance onside, and for anyone who assumes regulation and innovation have to be at odds. Key Takeaways Edmund Towers explains why the FCA's AI Live Testing lets firms trial live products under supervision, and why it's built on guidance rather than a tick-box checklist.The Consumer Duty already covers responsible AI, according to Towers, and blaming a "black box" model won't protect a firm from accountability under the Senior Managers Regime.Sarah Self's team at Aviva cut claims-handler hold times by more than 50% with a summarisation tool, and built a medical underwriting system running at 99% accuracy.Kevin Cassar reveals the co-sponsorship model, COO paired with CDO, that got a health insurance triage system built with risk and business teams involved from day one. Chapter Markers 00:00 Why regulated industries need a different AI conversation 02:07 The FCA's role in AI adoption across financial services 04:10 Inside the FCA's AI Lab and Live Testing programme 07:49 Why Consumer Duty rules already cover responsible AI 10:37 Santander's shift from risk-averse to AI-enabled 13:05 Aviva's claim summarisation tool cuts hold times in half 16:07 Building an AI triage system in health insurance 18:39 Getting COO and CDO co-sponsorship for AI projects 20:03 The FCA's fight against fraud, scams and money laundering 23:41 Closing thoughts: compliance and innovation as one agenda Useful Links Follow Dr Raoul-Gabriel Urma on LinkedIn: https://uk.linkedin.com/in/raoulurma Visit the Cambridge Spark Website: https://cambridgespark.com/

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com For years, digital transformation and AI transformation have been used almost interchangeably. Divya Kesavan, Head of Data Science for Business and Commercial Banking at Lloyds Banking Group, explains why that conflation breaks down in practice. In this episode, Divya walks Jeremy through the shift from rule-based RPA to reasoning-driven AI and why Lloyds' enterprise AI blueprint rests on three layers: reasoning, context and control. She unpacks what it actually takes to move an agentic system from a working proof of concept to a tool trusted with real banking decisions, including a striking example of an AI agent that went badly off script. The conversation also covers governance by design in a heavily regulated environment, confidence scoring as a practical risk tool, and why agile alone cannot carry the weight of AI-led change. A grounded, experience-based look at what transformation really requires. Follow Data & AI Mastery for more conversations with the people building AI inside major enterprises. If you enjoyed this conversation, check out this special episode of Data & AI Mastery where Raoul sits down with senior leaders from Lloyds to learn more about how the team is reimagining banking through AI-powered experiences:  Apple: https://podcasts.apple.com/gb/podcast/reimagining-lloyds-banking-group-ai-transformation/id1779783413?i=1000701688423 Spotify: https://open.spotify.com/episode/06ZBLoyvemQEVftCuWytVe?si=726cf7a2d20b4b22 YouTube: https://www.youtube.com/watch?v=U2krteWX89Q Glossary Terms RPA (Robotic Process Automation): software that mimics human actions on a screen to automate repetitive, rules-based tasks.  RAG (Retrieval-Augmented Generation): a method that grounds an AI model's responses in real, relevant data retrieved at the time of the request, rather than relying solely on what the model learned during training. Human-in-the-loop: a safeguard where a person reviews or approves an AI system's output before it's acted on, typically when confidence is low or the decision carries risk. Confidence scoring: a technique used to measure how certain an AI system is in its output. Outputs above a set threshold proceed automatically, while lower-confidence results are flagged for human review. Chapter Markers: (00:00) - Cold open: the case for AI governance (00:39) - Introduction: Divya Kesavan, Lloyds Banking Group (02:23) - Where automation stops and AI begins (03:39) - Why traditional RPA breaks under ambiguity (05:04) - From RPA blueprint to AI-first processes (06:47) - Building agentic reasoning: the real technical challenge (09:29) - The rogue agent story and why permissions matter (12:03) - Lloyds' enterprise AI blueprint: reasoning, context, control (14:45) - Reliability as the hardest problem to solve (16:34) - Designing and testing AI in a regulated environment (19:25) - Embedding responsible AI thinking across the team (22:54) - Confidence scoring and managing risk in practice (24:34) - Common misconceptions about AI versus automation (26:18) - Balancing probabilistic AI with deterministic workflows (27:39) - Advice for data science leaders starting out (30:04) - The one idea to take away Useful Links: Connect with Divya Kesavan on LinkedIn: https://uk.linkedin.com/in/divya-kesavan-54800b3 For more AI insights follow Jeremy on LinkedIn: https://uk.linkedin.com/in/jeremy-bradley Explore Cambridge Spark’s AI upskilling programmes at https://www.cambridgespark.com

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com In this episode of Data & AI Mastery, host Dr Raoul-Gabriel Urma is joined by Simona Paravani-Mellinghoff, Chief Investment Officer at BlackRock, to explore why data quality and accessibility sit at the heart of every serious AI strategy. Simona explains what good data infrastructure looks like when overseeing hundreds of billions in mandates and why relevant data, not just volume, drives better investment insights. She shares her Ferrari and Fiat Punto analogy for deciding when to use large generative AI models versus smaller, specialised ones and outlines a three-layer skill pyramid, prompting, critical thinking and creative thinking, that leaders and their teams will need to thrive in an AI-driven world. The conversation also covers AI's role in expanding financial inclusion, why Simona funds scholarships and teaches at Cambridge, and her advice for anyone starting a career in data or finance today. If you enjoyed this episode, follow Data & AI Mastery and share it with a colleague thinking through their own data and AI strategy. Chapter Markers (00:00) - Introduction: no data, no AI party (01:16) - Simona's journey from Italy to global CIO (03:40) - What academia and industry learn from each other (05:21) - Building good data infrastructure at scale (08:56) - Frontier models vs specialised models: the Ferrari and Fiat Punto analogy (11:42) - Why education and scholarships matter to Simona (15:03) - The three-layer skill pyramid for the AI era (17:19) - Will AI threaten jobs or create new ones (21:04) - AI's role in advancing financial inclusion (23:22) - Advice for the next generation (24:06) - Quickfire round (27:22) - Closing thoughts and key takeaways Useful Links Connect with Simona Paravani-Mellinghoff on LinkedIn: https://uk.linkedin.com/in/simona-paravani-cfa-5129524 Follow Dr Raoul-Gabriel Urma on LinkedIn: https://uk.linkedin.com/in/raoulurma Visit the Cambridge Spark Website: https://cambridgespark.com/

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com What does it actually take to ship machine learning inside one of the UK's largest insurers? Jeremy Bradley sits down with Alberto Romero, director of AI engineering at Aviva, to trace his path from InsurTech founder to enterprise AI leader. Alberto explains why prototypes are so often mistaken for finished products and what production readiness really demands once edge cases, drift and adversarial behaviour enter the picture. The conversation covers how to get genuine explainability out of large language models rather than plausible-sounding justification, when fine-tuning earns its place in a regulated stack, and why Aviva built its own internal platform to govern AI use cases at scale. Alberto also shares his take on fraud detection as an adversarial ML problem and the one failure mode he sees engineering teams repeat most often. Follow Data & AI Mastery so you never miss an episode, and share it with a colleague working through similar production challenges. If you enjoyed this conversation, you might also like this episode featuring Sarah Self. She joined us on Data and AI Mastery to explore what most organisations get wrong when deploying AI. Apple: https://podcasts.apple.com/gb/podcast/from-cybersecurity-to-ai-director-sarah-self-on-leading/id1779783413?i=1000764247007 Spotify: https://open.spotify.com/episode/0BpSq5X1ZP8ctIYTxWVAJT?si=1264586e79f3446f YouTube: https://www.youtube.com/watch?v=2jgM095SYG0 Glossary Terms RAG: Retrieval-Augmented Generation is an AI methodology that enhances Large Language Models by pulling factual context from external knowledge bases. GAN: Generative Adversarial Network is a deep learning architecture in which two neural networks compete against each other to create highly realistic synthetic data from a training dataset Non-deterministic: describes a process, algorithm, or system whose outcome is inherently unpredictable and cannot be guaranteed to repeat exactly, even when it starts from the exact same initial conditions ReAct (Reasoning + Acting) approach: a prompting technique that enables AI models to solve complex problems by alternating between thinking and taking action Chapter Markers (00:00) - Cold open: why prototypes get mistaken for production (02:53) - Avoiding common AI adoption pitfalls in regulated sectors (05:44) - Real explainability versus post-hoc justification in LLMs (09:18) - From startup founder to enterprise: the mindset shift (11:38) - Managing AI across 70+ use cases at Aviva (13:31) - Standards first, technology second (17:19) - Where fine-tuning earns its place (20:33) - Building Aviva's own governed AI platform (23:51) - Fraud detection as an adversarial ML problem (28:34) - Quick fire: the most common AI failure mode (29:37) - What deserves more attention as AI scales Useful Links Connect with Alberto Romero on LinkedIn: https://uk.linkedin.com/in/albertoromero-uk For more AI insights follow Jeremy on LinkedIn: https://uk.linkedin.com/in/jeremy-bradley Explore Cambridge Spark’s AI upskilling programmes at https://www.cambridgespark.com

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com There's a real difference between an organisation experimenting with AI and one that has genuinely thought through what it means for its strategy, its accountability, and the people it serves. That gap is where a lot of the frustration lives. In this episode, we bring together some of the sharpest thinking from across the show on AI strategy and governance: the decisions that happen before AI gets deployed and the frameworks that make sure it keeps working once it does. You'll hear from Luke Pearce, Chief Data Officer at Santander UK, on presenting AI governance at the board level. From Miryem Salah at VodafoneThree, on building an AI strategy from first principles. And from senior leaders at Aviva, Oxford Saïd Business School, and Intact Insurance, all grappling with the same fundamental question: how do you move fast enough to stay relevant, but carefully enough to stay trusted? Follow Data and AI Mastery to stay ahead of the conversations shaping the future of data and AI. These are just the highlights. Find links to every full episode below. Miryem Salah: Apple - Spotify - YouTube Sarah Self: Apple - Spotify - YouTube Luke Pearce: Apple - Spotify - YouTube Mark Bramwell: Apple - Spotify - YouTube Indhira Mani: Apple - Spotify - YouTube Conny Ploth: Apple - Spotify - YouTube Nick Edwards: Apple - Spotify - YouTube Chapter Markers (00:00) Episode Introduction  (02:01) Miryem Salah, VodafoneThree: building an AI strategy from first principles (05:40) Sarah Self, Aviva: why guardrails and customer outcomes are not a trade-off (09:47) Miryem Salah, VodafoneThree: a governed, agile approach to change (12:28) Luke Pearce, Santander UK: moving the board past fear of hallucination (16:22) Mark Bramwell, Oxford Saïd Business School: a portfolio approach to AI tools (20:13) Luke Pearce, Santander UK: a domain-led approach to use cases (23:44) Indhira Mani, Intact Insurance: what success looks like a year from now (27:24) Conny Ploth: why you can't manage what you don't measure (29:53) Nick Edwards, The AA: curated, not controlled (30:53) Closing reflections Useful Links Follow Dr Raoul-Gabriel Urma on LinkedIn: https://uk.linkedin.com/in/raoulurma Visit the Cambridge Spark Website: https://cambridgespark.com/

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com Large language models are surprisingly good at producing fluent, plausible text. So why do they still confidently get simple things wrong? In this episode, Dr Jeremy Bradley is joined by Dr Vaishak Belle, Reader at the University of Edinburgh's School of Informatics, Alan Turing Institute Faculty Fellow and Director of Research and Innovation at the Bayes Centre. Vaishak has spent 16 years working at the intersection of logic, probability and machine learning and brings that lens to one of AI's most persistent problems: hallucination. The conversation traces why scaling alone will not solve reliability, what neurosymbolic AI actually is and why tools like Claude Code quietly depend on it, how theory of mind is being engineered into language models, and where reinforcement learning fits into the future of AI reasoning. If you work at the frontier of AI research or engineering, this is a grounded, technically rich conversation worth your time. Follow Data & AI Mastery so you never miss an episode. If you enjoyed this conversation, you might also like this episode featuring Dr Petar Veličković. Petar joined us on Data and AI Mastery to explore how graph neural networks bring structured reasoning into systems like Google Maps and how AI is being used as a genuine discovery partner in mathematics. Apple: https://podcasts.apple.com/gb/podcast/bridging-ai-research-and-real-world-impact-dr-petar/id1779783413?i=1000734000576 Spotify: https://open.spotify.com/episode/7qA0AY9MlLS2L9PANqlXNi?si=37d8a8fb43c14cc9 YouTube: https://www.youtube.com/watch?v=GwMUSNidnvE Glossary Terms Neurosymbolic AI: an emerging field that merges the intuitive pattern recognition of neural networks with the logical, rule-based reasoning of symbolic AI. Theory of Mind: refers to an AI’s capacity to attribute mental states to humans or other agents and understand that these states may differ from its own. Confabulation: In AI, it is the generation of factually incorrect, distorted, or entirely fabricated information presented as absolute truth. Delegation Module: a software component that allows users or systems to assign tasks, roles, or access rights to others. Retrieval Augmented Graphs: an advanced AI framework that enhances large language models by grounding their responses in interconnected data networks, such as knowledge graphs. Reinforcement Learning: a machine learning method where an AI agent learns to make decisions through trial and error. Dynamic Pathway Analysis: a computational method used in systems biology and bioinformatics to model and simulate how biological processes change over time. Chapter Markers (00:00) - Why LLM hallucinations happen (02:12) - The biggest shift in AI over the last 16 years (07:42) - How logic and probability shaped Vaishak's path into AI (10:11) - Introducing neurosymbolic AI (11:27) - Claude Code, algebraic delegation and the theory of mind problem (20:01) - Theory of mind in robotics and human computer interaction (21:54) - Confabulation versus hallucination (26:42) - Why AI errors are not the same as human dishonesty (27:24) - Reinforcement learning, reward signals and learned behaviour (32:36) - Syntax checks and the engineering behind reliable code (37:52) - What happens to the software engineer's role (40:49) - Advice for early career AI thinkers Useful Links Connect with Dr Vaishak Belle on LinkedIn: https://uk.linkedin.com/in/vaishakbelle Learn more about Vaishak’s work here: https://www.vaishakbelle.org/ Read Vaishak's report on The Future of Neuro-Symbolic AI: https://ojs.aaai.org/index.php/AAAI/article/view/42130 For more AI insights follow Jeremy on LinkedIn: https://uk.linkedin.com/in/jeremy-bradley Explore Cambridge Spark’s AI upskilling programmes at https://www.cambridgespark.com

👉 Discover how Cambridge Spark helps organisations build the data and AI capabilities needed to turn strategy into measurable impact: cambridgespark.com The technology is the easy part. Getting people ready for it, that's the hard part. In this special compilation episode, Dr Raoul-Gabriel Urma steps back from the individual conversations and zooms out to one of the most persistent themes across the show in 2026: AI skills and literacy. Because no matter the industry, the company size, or the stage of the AI journey, the same challenge keeps surfacing. How do you get an entire workforce genuinely ready? You'll hear from Jessica Rusu, Chief Data, Information and Intelligence Officer at the FCA, on building T-shaped skills for an uncertain future. From Nick Edwards and Harj Johal at the AA on the combination of tooling, training, and space to practise that has driven rapid, organisation-wide progress. From Sarah Self at Aviva on why democratising AI is as much a cultural imperative as a commercial one. And from senior data and AI leaders at Santander UK, Oxford Saïd Business School, and TransUnion on curiosity, fear, and what genuinely high-performing teams look like in an era of AI. A compelling, cross-industry episode for anyone thinking about the people side of AI transformation. These, however, are just the highlights. Find links to every full episode below: Jessica Rusu: Apple - Spotify - YouTube Nick Edwards & Harj Johal: Apple - Spotify - YouTube Sarah Self: Apple - Spotify - YouTube Mark Bramwell: Apple - Spotify - YouTube Luke Pearce: Apple - Spotify - YouTube Kevin Cassar: Apple - Spotify - YouTube Conny Ploth: Apple - Spotify - YouTube Sudhish Mohan: Apple - Spotify - YouTube Follow Data and AI Mastery to stay ahead of the conversations shaping the future of data and AI. Chapter Markers (00:00) Introduction and welcome (01:31) Harj Johal, The AA: curiosity as the defining skill for AI adoption (03:14) Nick Edwards, The AA: tooling, upskilling and watching the water level rise (05:05) Sarah Self, Aviva: democratising AI and the case for broad access (08:01) Mark Bramwell, Oxford Saïd Business School: leading through uncertainty and creating a culture of FOMO (10:03) Luke Pearce, Santander UK: removing fear and augmenting expertise (11:14) Jessica Rusu, FCA: T-shaped skills and building flexibility for the future (13:31) Kevin Cassar, TalkTalk: the value of continued learning and the apprenticeship experience (16:13) Conny Ploth: reimagining upskilling for the 21st century (19:14) Sudhish Mohan, TransUnion: curiosity over credentials (21:47) Closing reflections Useful Links Follow Dr Raoul-Gabriel Urma on LinkedIn: https://uk.linkedin.com/in/raoulurma Visit the Cambridge Spark Website: https://cambridgespark.com/