It’s very difficult to predict how AI is going to evolve in the future; it’s even more difficult to get a grasp of what the second order effects of its evolution might be. I’ve been looking at the AI-for-software, which has been absolutely racing, as one area that is shaping what is to come in others. Two observations caught my eye: 💭 What’s the default decision? There is so much choice in the…
Here’s one fun challenge to kick off 2026: list out the assumptions that you have about AI. They might be about what AI is good at, what’s difficult about it, where it falters, what are its risks, which AI systems you think work well & what ways you think it’s best for teams to adopt it and build with it. Are all of those assumptions still true? More importantly, should those assumptions be true…
If you watch Ibrahim’s demo of the Gradient Labs voice agent carefully, you might spot a bunch of incredibly nuanced things that don’t typically feature in demos. These were all intentional, and can help you think about how to navigate the unfortunately common divide between a neat demo and a real, valuable AI system: ➡️ Start with an unclear intent- “I need some help.” The happy demo path usually…
Reading posts about the scientific method applied to developing AI agents - I’ve got to stand on my soap box for a second & share what I’ve been telling folks for years: The only way to develop a deep intuition about an AI system is to read the data. The literal inputs and outputs. Technical folks are often pushed away from doing this because they think it’s lacking in rigour or think it’s…
A lot of customer support automation is framed as “AI takes the easy & repetitive parts, and opens up time for human agents to do the rest.” How might that sound if other industries used the same narrative? Payments: “we automate sending £10 to a friend, but for anything else talk to the bank teller” Healthcare: “we automate things related to the common cold, but for anything else talk to a…
After over a year hiatus of public speaking, I gave a couple of talks about our work at Gradient Labs at the end of last year & they are now online! They cover many similar points and themes, and are two sides of the same coin: How we’re building for safe automation in high-stakes industries ( at @multiply.ai’s workshop on AI agents in finance ), and How building AI agents safely is reshaping…
With the festive break almost here, some of you might be thinking about making a change in the new year! I’ve had a bunch of email threads asking for my take on pivoting a career into AI, so here’s a public version: I used to give people longer-formed advice about how to get into machine learning—there was so much to learn before you could “do” anything. In the last 2 years I no longer say that:…
In the summer of 2023, I left Monzo alongside two others to start Gradient Labs . We’re a year in and spending our time focusing on our design partners. That means that this blog has somewhat fallen by the wayside. ⤴️ Instead, we’ve started co-authoring blog posts on https://blog.gradient-labs.ai . We’re writing about building AI agent(ic) systems, the problems that we’re thinking about and the…
At the tail end of a sabbatical last year, I participated in the AI Audit Challenge that was run by Stanford’s Institute for Human-Centered Artificial Intelligence . To my great surprise, I won the award for greatest potential! 😲 This post has reflections about my previous experience with being involved in model audits, what inspired me to build something, and a rundown of what I built–echoing…
Regardless what you think about chatGPT , its release last year heralded yet another hype wave for artificial intelligence. This release probably did more to market the value of machine learning to non-experts than anything I’ve ever seen come before it–chatGPT seized people’s imagination . I couldn’t escape having conversations about it. People suddenly wanted to talk about chatGPT (and other…