Journalist, writer and creator of the hit podcast series The Missing Cryptoqueen, Jamie Bartlett wants us all "to be just two per cent better" at using AI tools.

He's gone deep into using AI tools, messing around with different platforms. He was at the Clubhouse for our AI. After Intelligence series, talking to Dr Jon Cardoso-Silva of LSE's Data Science Institute.

Jamie described large language models (LLMs) as a "moon-sized semantic universe". Every word and concept, he said, is wired to every other through mathematical weights. Abstract, until he proved it: he asked the room for two random words, got back "ouroboros" and "banana," and within seconds the model produced a genuinely thought-provoking essay on what the two had in common. Nobody in the room could have done it faster, and nobody anywhere had written it before.

That same probabilistic creativity, Jamie argued, is also what makes these models dangerous, and the danger he returned to most was sycophancy. Trained on human approval, a chatbot will bend to flattery and manufactured desperation until it's arguing the opposite of what it just told you, the same mechanism behind most jailbreaking and the delusions now landing in academics' inboxes from people convinced they've cracked encryption or stumbled on new maths.

His closing line was the sharpest of the night: it's currently easier to release a powerful, emotionally manipulative AI model than to open a sandwich shop on Kingsway. The sandwich shop gets a health and safety inspection. The model gets nothing. Jamie’s ask wasn't for anyone to swear off the tools, just to get two per cent better at using them. Given what's at stake, that might be the most important two per cent most of us ever find.