GAEA Talks

Forrester technology and innovation forums (Austin, London, New York). Promo code "GAEATECH" for 10% off. Visit https://forrester.com/events/#tech to book your place.This week on GAEA Talks, Graeme Scott sits down with Darko Fabijan, co-founder of Semaphore and now of Superplane, for a conversation about what happens to software, to moats and to human roles when the cost of building something collapses to almost nothing.Darko started writing code at eleven or twelve on Visual Basic and Windows 3.11, moved into Linux around 1996, and studied computer science with a focus on the low-level end of the stack, device drivers and how things actually work underneath. He moved into Ruby on Rails around 2008, ran product for several Austin-based startups, built a consultancy of around twenty people, and then co-founded Semaphore, the continuous integration and delivery platform he ran for twelve years, with customers including Confluent, Superhuman and Reply. About a year ago he started Superplane.Filmed in our London studio with Darko joining remotely, this is one of the most practical conversations we have recorded on what AI is actually doing to the craft of building things.His starting observation is that for the last two decades software engineers had abundance in compute, storage and bandwidth. Now a second variable has changed. Intellectual resource is close to unlimited. Two people can run a twelve hour hackathon, burn tokens, and produce something that would previously have taken a team months. But almost every other variable in the system has stayed exactly where it was. Can you come up with good ideas that fast? Can you validate them on customers that fast? Can you get in front of people that fast? No. And that mismatch is where most of the current confusion lives.The consequence he is most direct about is mediocrity at scale. If you can build it in a weekend, so can everyone else, and largely the same way. He draws the parallel with desktop publishing in the late nineties, when everyone suddenly had the tools of a graphic designer and almost nobody had the principles. Software, pitch decks and LinkedIn posts are all converging on the same predictable output for the same reason.He is equally clear about what that does to moats. Building an application used to take millions of dollars of developer time, and that expenditure was itself the moat. Copying is now close to free, because someone else already thought through the UX and the details. The moat has to come from somewhere else.

Martin has been writing about this since 2008, when he was running a small software company in Silicon Valley and started noticing the trend lines. His first book on the subject came out in 2009, more than a decade before ChatGPT. In 2018 he published a book of interviews with more than twenty of the most significant people in the field, including Demis Hassabis, Yann LeCun and Rodney Brooks. He released an updated edition of Rise of the Robots with a substantial new chapter last year. Very few people have watched this question for as long, or from as consistent a position.His central argument in this episode runs directly against the prevailing consensus.The conventional wisdom, promoted heavily by think tanks close to Silicon Valley and taken seriously in publications including The Economist, is that advanced AI will turbocharge growth, potentially taking a developed economy to twenty or thirty percent annual growth. Martin thinks the opposite is at least as plausible. Consumer spending is around seventy percent of the US economy. Every recession in recorded history follows the same self-reinforcing cycle, where people lose work or fear losing it, cut spending, businesses see falling demand and cut more jobs. His concern is that AI-driven job losses would be perceived as permanent rather than cyclical, which makes that cycle worse, not better.He also lays out two risks that most commentary treats separately and which he argues are intertwined. One is AI automating large parts of the workforce. The other is the AI bubble bursting because the frontier labs cannot generate the revenue to justify the capital being deployed. Neither excludes the other. And historically, economic downturns are precisely when companies turn to labour-saving technology.What you will take from this conversation:• Why the frontier labs were selling to investors rather than to consumers, and what that did to the narrative• Why Martin thought Dario Amodei's white collar jobs prediction was over the top, despite broadly agreeing with the direction• The data centre backlash and where it came from• Why open weight models from China may undermine the frontier lab business model entirely• The railroad and fibre optic bubbles, and why AI infrastructure may not age as well as either• Continual learning as the single missing capability holding AI back from real workforce impact• Why a graduate is useless on day one and proficient in six months, and why models cannot do that• The S-curve argument - propeller planes to jets, and whether LLMs are near their ceiling• Rodney Brooks and the one dollar litter picker that beats a hundred thousand dollar robot• Why electricians and plumbers are currently the safest jobs in the economy• Martin's scepticism about humanoid robots and the Optimus value proposition• Why universal basic income is necessary but nowhere near sufficient• The education incentive problem UBI creates, and how he would fix it• Why the "live experience economy" is not a solution at scale, and the indigenous craft economies that prove it• What happens when a high wage country becomes a low wage country, and why it would be catastrophic

This week on GAEA Talks, Graeme Scott sits down with Manos Koukoumidis, co-founder and CEO of Oumi, joining from Seattle. Manos spent his career building the technology that became Gemini, then left Google because he became convinced it was the wrong answer for enterprise.At Google Cloud, Manos led science and engineering for natural language AI services. The model his teams built shipped as Google Cloud PaLM and later became Gemini. Before Google he was at Meta and at Microsoft, where in 2016 he built Zo.ai, an open-ended multimodal chatbot, six years before ChatGPT. He has been working in AI for close to twenty years, and he was pushing Google leadership to prioritise text-to-text generative models a full year before ChatGPT launched.Then he walked away from it. His reasoning is the spine of this episode. A handful of companies owning and controlling the most critical technology of the century is a terrible idea, and history is fairly clear on what happens when that much power concentrates in that few hands. But he also makes a colder, more practical argument. If AI is genuinely critical to your enterprise, why would you rent a generic model built for everybody and optimised for no one?His analogy is the sharpest we have had on the show. If you were performing surgery, would you rent the biggest Swiss Army knife available, one that happens to have a blade, and one the owner could take back mid-operation? Or would you use a scalpel that you own?Oumi exists to make the scalpel easy to build. The product launched a couple of months ago and Manos describes it as a frontier AI engineer, or Claude Code for AI development. You start with a prompt describing the model you want. It builds your evaluations, curates data against the gaps it identifies, selects the training strategy, trains, evaluates and iterates until it has the best model it can produce, then deploys it and keeps improving it in production. Human effort measured in minutes rather than months.

This week on GAEA Talks, Graeme Scott sits down with Jeff Bullas, one of the most widely read voices in the world on social media, digital marketing and now AI. Jeff joined us from Australia.Jeff's route into technology was not conventional. He trained as a high school teacher and spent six years in the classroom, teaching fifteen year olds about history and wisdom at the age of twenty one, before deciding the curriculum was letting them down and the profession was burning people out. He ran an experiment over one summer holiday, trying real estate, life insurance and technology, and chose technology. That was 1984, Jobs versus Gates, the PC wars. He never left.In December 2008 he joined Twitter, when it had around five million users. His first tweet was "watching the cricket", which confused a great many Americans. He started jeffbullas.com the same year because he believed social media was about to change the world. Eighteen years later he writes on his blog, on Substack, on LinkedIn and on X, and has built one of the largest independent audiences in the space.Filmed in our London studio with Jeff joining remotely, this conversation is about what social media taught us and whether we are about to make the same mistakes with AI. Jeff's mission, which he says took him fifty two years to find, is helping people use AI to amplify their humanity rather than be trapped by it.He is clear-eyed about the mechanics. Social media changed when Facebook went public and had to answer to shareholders, at which point the algorithm was redesigned to serve the platform rather than the human. Neuroscience and psychology were brought in to keep people on it. His argument is that AI is now built on the same incentive. Success is measured in time on platform, and sycophancy is a feature not a bug. The answer, in his view, is not to reject the technology but to be awake to how the game is played.The most useful thing in the episode is what he does about it. Jeff runs Claude, ChatGPT, Gemini and DeepSeek. He uploads a new story from his own life every day, plus book summaries and his own history, and then asks the models what patterns they can see in him. What energises him. What drains him. His view is that these systems are super pattern recognition machines, and that they can find the signal in the noise of a human life better than the human can, because we are too close to ourselves.

This week on GAEA Talks, Graeme Scott sits down with Vivek Natarajan, Research Scientist at Google DeepMind, where he works at the intersection of AI, science and medicine. Vivek is one of the people most directly responsible for bringing large language models into healthcare, and this is one of the most genuinely hopeful conversations we have recorded on the podcast.Vivek grew up in Tamil Nadu, India, where he watched people walk thirty or forty miles in extreme heat, give up a day's wages or go without food to see a doctor. His uncles ran eye camps in nearby villages, sending out flyers a year in advance because that was the only reliable way to get people to come and be examined. That experience never left him. As an undergraduate he and a few friends tried to build an app called Ask The Doctor Anytime, Anywhere. The technology was not ready. He came to the US for graduate school, joined Facebook AI Research in 2014 in the early days of deep learning, and then moved to Google to join the newly formed Medical Brain team under Greg Corrado, co-founder of Google Brain. He has now been at Google and Google DeepMind for seven and a half years.Filmed in our new London studio, this conversation covers the full arc of medical AI. Vivek walks Graeme through the early specialised vision models his team built for detecting skin conditions and breast cancer from mammograms, why those supervised approaches kept breaking the moment they left the hospital they were trained in, and why the arrival of large language models changed everything. He tells the story of the moonshot proposal he and Dr Alan Karthikesalingam wrote over dinner in 2022, which fifty colleagues signed up to within a week, and which became Med-PaLM, one of the first specialised medical LLMs. Within months it was achieving expert-level scores on US medical licensing exam questions. When the paper went out over the Christmas break of 2023, the heads of many of the world's top health systems contacted Google asking for access immediately.