09 May 2025

EP8: Solving Hard Problems with AI: Creativity, Strategy & Cutting-Edge Tools | Passion Lab

While much of the AI conversation centres on large language models and generative chatbots, Passion Lab, a London-based AI research and development lab, is focused on something broader: building custom, proprietary AI systems to solve problems that off-the-shelf tools simply can't. In a recent episode of the UK AI podcast, Tim spoke with co-founders Tom (CEO) and Nadine (Chief Scientific Officer) about their work across creative industries, structural engineering, and beyond — and why understanding AI beyond LLMs is essential for businesses trying to keep pace.

 

More Than Just an AI Consultancy

Passion Lab positions itself first as an AI audit and advisory lab, helping businesses assess their internal capabilities, workflows, and data readiness before recommending existing tools or building entirely proprietary systems. The company's mission, as Tom describes it, is to help businesses "unlock their potential" with AI — and its client list spans creative and marketing agencies, record labels, fashion brands, and even structural engineering firms, where Passion Lab built a novel system to automate converting 2D architectural drawings into 3D structural designs.

What sets the team apart, Tom explains, is a preference for solving genuinely unsolved problems — building proprietary AI systems from scratch rather than simply wrapping an existing API.

AI in Music and the Creative Industries

Drawing on her background in machine learning for music understanding, Nadine outlines several practical applications of AI within the creative and media sectors:

  • Trend detection — using machine learning to identify emerging music trends on social platforms like TikTok before they break into mainstream visibility on services like Spotify
  • Semantic search — moving beyond traditional genre-based browsing toward search by example or natural description, applicable to music, image, and video
  • Audio as a sensor — using sound analysis for applications like low-light surveillance or detecting equipment and machinery faults
  • Recommendation systems — applying the same personalisation principles that power platforms like TikTok to improve engagement across websites and other digital platforms

Why LLMs Aren't Always the Right Tool

One of the most valuable insights from the conversation is Nadine's clear-eyed view of the limitations of large language models. While LLMs are excellent for tasks resembling human language processing — summarising, rewriting, extracting information — they're often too slow or expensive for high-volume, real-time tasks like spam filtering, and poorly suited to problems like large-scale logistics planning.

Instead, Passion Lab frequently uses LLMs to generate synthetic data, which is then used to train more efficient deep learning or reinforcement learning models — treating the LLM as "one cog in a bigger system" rather than a universal solution. This nuance matters, Tim notes, because many organisations declaring themselves "AI-first" default to LLMs without understanding whether that's actually the right tool for their specific problem.

Is Agentic AI Really New?

On the growing hype around agentic AI, Nadine offers useful historical context: agents, in the technical sense, are a foundational concept from reinforcement learning, already powering things like YouTube and TikTok recommendations, logistics decisions, and traffic-light systems. Agentic AI, she argues, isn't a brand-new idea — it's a familiar approach applied in a new context.

Tom adds a note of caution: realising the full potential of agentic systems depends on factors well beyond AI itself, including API availability and the state of often outdated data infrastructure within organisations — meaning meaningful agentic adoption will take time, not happen overnight.

Advocating for Creators in AI Policy

Passion Lab has taken an active role in UK AI policy discussions, participating in the House of Lords debate on the Data (Use and Access) Bill. Tom describes the company's position as firmly supportive of creators, while acknowledging that meaningful safeguards require greater transparency into how LLMs are trained and how data is used — an area he says still needs significant work.

Key AI Adoption Trends to Watch

Tom highlights several statistics shaping the current AI landscape:

  • Around 80% of UK children are reportedly using LLMs to help with homework and exams — driving early, widespread familiarity with prompting and AI tools
  • Roughly 35% of recent AI use cases are being built by non-technical users, powered by low-code and no-code tools such as Cursor and Lovable — a trend Tom calls "vibe creation"
  • 98% of Fortune 500 businesses have implemented AI in some form — but Tom warns of a significant gap between adoption and having a coherent AI strategy, with some companies now needing to "unravel" hastily adopted tools with unclear data practices and overlapping functionality

Tom also flags the often-overlooked issue of AI's energy consumption, citing projections that data centres could consume around 1,000 terawatts annually by 2030 — compared to the UK's entire annual consumption of roughly 200 terawatts — underscoring the need for continued research into AI efficiency.

Advice for Businesses Trying to Keep Up

Both founders emphasise humility and continuous learning as the best strategy for navigating a fast-moving AI landscape. Nadine encourages businesses to resist the urge to chase every new tool, and instead invest time in understanding AI as a discipline, not just its latest applications. Tom encourages active upskilling through free resources from providers like Coursera, Microsoft, and Google — and stresses the value of understanding AI's long history, stretching back over 70 years to Alan Turing, rather than assuming the field began with the recent generative AI boom.

Return to listing