John Hearty

AI literacy and training - especially in Responsible AI is essential. Congratulations to the latest cohort that graduated from Edinburgh Futures Institute. This bespoke course is designed for professionals who create and work with commercial AI applications. Thank you to all the teams at NatWest Group, including our AI & Data Ethics team, and to Professor Shannon Vallor who made this course a success - a first of its kind. We will continue to expand this to even more colleagues by launching this course internally in the bank.

99

7 Comments

Stop treating your favourite LLM like a single source of truth. Andrej Karpathy just dropped “LLM Council” – a simple local app where multiple models answer your question independently, then review and rank each other’s answers, and finally a “chairman” model synthesises the final response via OpenRouter. This matters because it: - Assumes models are fallible and need cross checking - Builds critique and review into the workflow - Focuses your effort on orchestration, not hero-worshipping one provider If you are a student or early-career dev, do not build yet another single-model chat UI. Fork this pattern and ship your own “council” on a real problem in your domain. Repo: https://lnkd.in/ghj4hjJ2

28

5 Comments

Another short blog post, and my first hand-drawn post for a while. https://lnkd.in/eYDjbw7K This post attempts to explain differences between the chatbot-style LLMs most people are familiar with, and Agentic AI. In brief: With chatbots Human and Bot interface through a chat window. With Agentic AI we need to think about Human, Bot and Machine as distinct elements, with Human and Bot both sharing access to (and so indirectly communicating via) the Machine, as well as the chat window. This one simple difference is key to why AAI can be so much more powerful than AI through chat window alone.

4

What a day at the Big Data London 2025 event. With connections rekindled and new alliances formed, the air was charged with insightful conversations that spanned from Anthropic/ Ephemeral Data to Vector Store and AI Observability to Diffusion Models. But the big story by far was Agentic AI. Completely agree with my good friend Gnans S. - It dominated stages and demos alike, underscoring both its growing potential and its current challenges. Yet the paradox persists: despite the heavy investment, success rates have been modest. Huge bets are being placed on new AI architectures, infrastructure, resourcing, and emerging tech — but answers to critical questions linger: Where is the ROI? How is success measured? Which investments truly drive meaningful outcomes? What is the business value add? I attended several sessions covering a wide range of topics — from AI Readiness and Data Culture & People to Kafka Streaming, AI Governance to MCP. I connected with various exhibitors including Sigma, Acceldata, and Databricks to explore their offerings and gain insights; and across all these discussions, one theme stood out: Deploying enterprise-scale Agentic AI requires more than ambition. My Key Takeaway - Technology is the enabler and AI acts as the accelerator - the true driver behind organisational success is a clear strategy. The next 12 months in Agentic AI will likely see fewer experiments and more real deployments. MCP and similar protocols will unify ecosystems, while hybrid architectures will balance cost and capability. Governance and security frameworks will rapidly mature with AI ready frameworks taking center stage. I walked away feeling both inspired and grounded by the opportunities, but mindful of the hard truths. Very excited to see how we shape our tech roadmap with Agentic AI deployments here at GLAS. #BigDataLDN #Data #AI #AgenticAI #AIObservability #GraphDatabase #Kafka #Iceberg #AIGovernance #MCP

106

Fantastic time had at the 3rd Annual Responsible AI Summit in London this week. Some new things I picked up: - New Consensus: AI governance has to be about enablement, not just compliance. - New Priority: Everyone is mobilising around governing AI agents. - New Question: What's your organisation's policy on AI wearables? - New Evidence: The level of AI bias in the same task varies across different languages (new proof from BBVA research). - New Risks: Agentic risk of cascading hallucinations, and dynamic value drift (where agents can learn to over-prioritise short-term goals) - amongst many more. - New Challenge: What's your human in the loop really for - to take the blame, or to review, ratify, or perhaps justify the decision? - New Analogy: If AI is Hannibal Lecters, AI Security is keeping them in prison, and AI Safety is stopping them from wrecking themselves and the prison New Contacts: Loads! So energising to meet many bright brilliant people working in the field. Well done to Mhairi Macbride and the team for putting on a fantastic event. See you next year. #RAISummit #aigovernance #responsibleAI

77

7 Comments

Please read the post below to explore our exciting research on GenAI applications in the banking and finance industry, being presented at ACL 2025! 💬AUTOSUMM: A Comprehensive Framework for LLM-Based Conversation Read the paper: https://lnkd.in/eH4EjBDQ How Personality Traits Shape LLM Risk-Taking Behaviour Read the paper: https://lnkd.in/eziMW4vR

23

1 Comment

Really excited to see the launch of our AI and engineering blog, where we'll be sharing some of the exciting insights from inside the team! https://nwg.ai/ In one of the first posts on this blog, I recap some of the cutting-edge insights that Raad Khraishi and I saw at NeurIPS 2024, where NatWest Group were bronze sponsors! https://lnkd.in/evE9U_mM Other early highlights include: - how LLMs and RAG are used to enhance our colleague chatbot, AskArchie: https://lnkd.in/eSuXPcTb - an overview of the research that our team has been producing recently, covering LLMs, Graph Neural Networks, and agent-based modelling: https://lnkd.in/eAi-iwYN Looking forward to sharing more on the blog in the near future!

22

This week the UK Government is set to announce a major collaboration with OpenAI and Nvidia, along with associated investment. On Friday morning last week I was live on BBC news discussing why this investment in the UK’s AI infrastructure is important, how the government is seeking to ensure the UK remains competitive with AI on the global stage. Thanks to Neha Misri for the video and support! UST #AI #SovereignAI #AIInfrastructure

58

1 Comment

AI: Are We Asking the Right Questions? One thing I’ve heard consistently from colleagues, over the years is this: “Are we asking the right question?” It’s a simple question, but a powerful one. As humans we clarify, challenge or reframe questions before jumping to try and answer them. A recent Sky News interview with ChatGPT highlighted just how important this is. If the question is vague, the answer often will be too. But ask for a specific source, timeframe, or comparison, and you’re far more likely to get a grounded, fact-based response. In a world increasingly reliant on AI, this matters. To get better answers from AI: • Be specific and clear about what you want • Ask for statistics, sources, or comparisons • Don’t assume the model will narrow it down for you

16

2 Comments

AI insight of the week for me :-) Stanford Verbalised sampling research paper: https://lnkd.in/eXcX_crG We’ve been trying to control models with longer and longer prompts. But a Stanford technique suggests the opposite — ask for multiple responses and their probabilities. I am definitely trying this out :-)

10

1 Comment

Y'all, I've been saying this for years and it's so good to see more people talking about it. A few years back, someone asked me about the relationship between data and AI, and why I was so adamant that one person should lead both in your organization. My answer then is the same as it is now: the data flywheel. You need someone who deeply understands that the quality of your predictions and outputs depends on users generating data you can actually use. It's a virtuous cycle when done right. This quote from the article nails it: "Every accepted or rejected suggestion adds to training data for future model improvements. With this data in hand, it was inevitable that model quality would improve dramatically over time." If you're building AI products and your data strategy lives in a separate silo from your AI strategy, y'all might want to rethink that. Data isn't just fuel for AI. It's the whole engine. https://lnkd.in/eavZcHKz #DataStrategy #ArtificialIntelligence #DataFlywheel #AILeadership #MachineLearning #DataDriven #TechLeadership

2

Very grateful to have been able to listen to Richard Sutton's keynote talk at #Neurips2025. He recently won the 2024 Turing award along with Andrew Barto for their foundational work in reinforcement learning. With so much noise around AI models or use of agents, it's great to turn it all off and listen to the key contributors in this field. In his talk, professor Sutton introduced the OaK (Options & Knowledge) architecture, a bold, long‑term vision for how intelligent agents can evolve beyond today’s models and truly learn from experience. He discussed designs enabling an agent to build abstractions, learn high‑level behaviors (Options), and form world models instead of constraining its world knowledge like today's pretrained models. State abstractions can be considered as features where they can be further broken down to sub problems. This vision echoes Sutton’s own “Bitter Lesson”: 70 years of AI history show that general methods that leverage computation and learning outperform human knowledge constraints — especially as compute keeps growing. His vision, is to have a domain independent agent that learns at runtime by experience able to discover its own abstractions without domain dependency. It was great to see Prof. Sutton and others of similar calibre moving toward experiential learning frameworks, in contrast to today’s more constrained, data-bound training paradigms. I see meaningful parallels with ongoing work in spatial intelligence, world models, and exploratory approaches that leverage Hamiltonian dynamics — all of which aim to enable agents to adapt and learn directly from their environment across diverse states while managing uncertainty and reducing entropy. Upcoming work as listed in the talk to accelerate us to superintelligence: 1. Reliable deep continual learning 2. Meta learning or step-size optimisation (can weights themselves have learning rates?) 3. Safe & efficient off-policy learning The next few years will get even more interesting and I can't wait to learn how we progress in these areas. (useful links & papers from the talk shared in comments) #AI #ReinforcementLearning #AGI #OaK #NeurIPS2025 #RichSutton #ContinualLearning

16

2 Comments

I’m excited to share the latest post for NatWest Group where I explore what really changes when we embed AI/ML into software systems and applications. In traditional software, correctness is logical and deterministic: given the same input, you always get the same output. Testing is typically straightforward with unit tests, assertions, and pass/fail outcomes. With AI/ML (especially generative AI), correctness becomes probabilistic. You must rely on statistically-valid evaluation, confidence intervals, and continuous monitoring as the same “input” can yield different “outputs,” even in production. This shift has big consequences for engineering, governance, and risk. If you’re thinking about integrating AI into your stack, it’s not just “another dependency”, and it often demands a whole new validation, monitoring and data-governance mindset. In future blogs we will talk about how we evaluate AI in production, explain how to govern these systems in the real world to build transparency and trust, while maintaining human involvement and oversight. If you work in software, data science, AI or governance I hope you find it useful. I’d love to hear your thoughts 👇 #AI Read more here: https://lnkd.in/e7cYcXTS

169

12 Comments

I’ve spent a lot of time this year thinking about how AI actually scales within a large organization. It’s one thing to run a pilot; it’s another to lead a full-scale transformation. I recently joined Kris McFadyen on the AI Right podcast to share my perspectives on: 1. Navigating the "AI transition" without losing sight of the people. 2. Where the real ROI is hiding in enterprise AI. 3. Why leadership matters more now than ever. Would love to hear your thoughts on our discussion! #AI #EnterpriseAI #Leadership #Innovation #Podcast

54

1 Comment

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