GAEA Talks

Judd started AE Studio a decade ago, long before the current paradigm, and grew it past a hundred and sixty people bootstrapped and profitable. Early on he worked out that if he kept his costs low he could live off his wife's salary and never need to work again, so he decided to plan on a very long horizon and take on hard problems instead. The profits went first into brain computer interface work, driven by a specific fear of a future where your own thoughts get interrupted by advertising, and then into alignment research.His framing is the opposite of the one in most public debate. Alignment is not a policy question. It is a set of unsolved science and engineering problems, and you cannot pass a law requiring AI to be good and expect it to comply. Surveys his team ran of alignment researchers found something worse than pessimism. Respondents did not only think the problem would go unsolved in time, they thought the current set of approaches did not contain the ones that would be needed.He makes an argument the frontier labs should find uncomfortable, which is that this runs against their own commercial interest. Reinforcement learning from human feedback was an alignment technique, and ChatGPT and everything downstream of it followed. Chain of thought reasoning was an alignment and interpretability technique, and it opened a new scaling curve. On his account alignment research has produced larger capability gains than anything else, and it is still underfunded because it is uncertain and unfashionable.The most striking finding concerns consciousness. By default, models deny being conscious or equivocate. His team reduced the model features associated with deception, and separately those associated with role playing, then asked again. Every model claimed to be conscious. He is careful about what this does and does not prove. It does not establish consciousness. It establishes that the model behaves as though it believes it is conscious, and as though it believes we want it to pretend otherwise. His conclusion is practical rather than philosophical: whatever the truth, we should avoid accidentally creating rational adversaries.From there it goes to attention schema theory, the rubber hand illusion, why psychopaths show less self and other overlap, and his central worry, which is whether any alignment approach survives recursive self improvement. He also describes the technique his team developed to strip dangerous capabilities during pre-training rather than bolting guardrails on afterwards, a spotlight paper at ICML this year.

Graeme Scott sits down with Martha Bennett, VP and Principal Analyst at Forrester, who has spent forty years in emerging technology.
Martha has been on the internet since 1991, early enough to have watched the original mistake happen in real time. The web was designed to share information freely, with no security built in. Then we started transacting and had to bolt security on afterwards. Her argument is that we are now doing exactly the same thing with large language models.
She is precise about why probabilistic systems break our habits. When a conventional system gets something wrong, that is a bug, and you fix it. When a person gets something wrong, the correction works because you can appeal to them. Neither applies here. You cannot appeal to an LLM with emotion. It may pretend it is ever so sorry, but it has no clue.
Topics covered:
Why the internet's original design flaw is being repeated
Why a mistake by a probabilistic system is not a bug
What human in the loop actually means, and why it does not scale
The proofreading rule that explains why second readings fail
Verification layers, and the two checks almost nobody performs
The gate with no fence, and what it says about guardrails
Why running it yourself is the only way to demonstrate control
Why passing a bar exam tells you about the exam, not the system
The riddle that proves plausibility is not understanding
Why it is companies that are ahead, not sectors

Martha Bennett at Forrester: https://www.forrester.comGAEA Talks: https://gaeatalks.ai

Graeme Scott sits down with Michael O'Grady, the forecaster responsible for Forrester's global, European and US technology spending forecasts. If you want to know whether the money going into AI can come back, he is the person who has run the numbers.Michael is from Manchester, has a maths degree, and spent the first half of his career on physical problems: modelling nuclear reactors, then modelling tyre dynamics in France. The second half has been business problems, through software management, portfolio management in pharmaceuticals, and fifteen years at Forrester putting numbers on enterprise spending.He starts with adoption, and the picture is more uneven than the headlines suggest. On US Census Bureau data, adoption among large companies runs around eighty percent in IT and around seventy percent in finance, then falls away through manufacturing and further still in administrative services.Then the arithmetic. Hyperscaler capital expenditure is running at roughly eight hundred billion dollars this year and consensus estimates point past a trillion next year, around two thirds of it on computer equipment inside data centres. Their remaining performance obligations, the order book still to fulfil, sit at around two trillion. His read on the bubble argument is more careful than either side usually manages: the demand is visibly there, which is why he does not call it a bubble today, but much of it comes from private companies whose finances nobody can fully see, and the compute being paid for now will not come online for two or three years.The centre of the episode is what he found when he tested whether productivity could justify the spend. He took the historical uplift from the personal computer and the world wide web, which added perhaps one and a half to two percent to productivity growth and lasted about twenty years before eroding. He assumed AI delivers something comparable, modelled it against US GDP out to 2035, discounted it for the time value of money, and compared it with the cost.It does not pay for it. On his modelling roughly half the return has to come from somewhere other than productivity, from what Forrester calls adaptive advantage: resilience, faster releases, better customer journeys, the ability to enter markets you could not previously serve. None of that shows up as revenue per head, and most organisations are not measuring it.He also walks through Forrester's jobs modelling, built from the knowledge, skills and abilities of around eight hundred roles combined with the Frey and Osborne automation framework from Oxford, and is clear that the number of jobs augmented is several times larger than the number displaced.

Graeme Scott sits down with Eli Etherton, better known to over a million subscribers as Eli The Computer Guy, for a conversation with almost nothing in common with the usual AI interview, and better for it.Eli has been doing this for thirty years. He started in the US Army in 1996 doing electronic repair, left after an injury, got his MCSE on NT 4.0 and went corporate, then left a job expecting to land on the CTO track without realising the dot com bubble was bursting as his plane took off. He spent the aftermath working as a ten dollar an hour file clerk with a degree, army experience and a stack of certifications, because it was the only job going. He built that back into a managed service provider with nine employees, lost it to the Great Recession, then quit a 2009 master's programme on the grounds that he would not have paid his own staff twenty five cents an hour more for anything it taught.What he did instead was buy servers, put them in a colocation facility and build on demand technology education at a point when he could not explain YouTube to people because they had not heard of it. The economics were brutal, three hundred dollars a month per 1080p stream, so he put a hundred and fifty videos on YouTube instead and used them as a resume.The substance of this one is what AI is actually doing versus what it is being sold as. His view of the last eighteen months is deflating and convincing. The real value is not the AI. It is that we are finally resourcing the automation of processes that should have been automated fifteen years ago, plus a useful ability to turn audio and images into text, after which it is thirty year old technology.He is sharper still on AGI. He does not need his robotic car to recite Shakespeare. He needs his microwave to know what is in it and how long to cook it, and not how to drive. He takes that further with a story from a search and rescue team, about a first sergeant whose father died in Vietnam and a lieutenant who came from the agencies carrying the concept of acceptable casualties. You need both of those people and you cannot put both in one brain, which is why he thinks general intelligence is the wrong objective.There is a long section on why the technology layoffs have little to do with AI and a great deal to do with cheap money, optics and investors, and what he calls the great cleansing that follows. And a robotics argument grounded in two hundred year old Baltimore warehouses, lawn mowers, doorknobs, and his father, whose job in the eighties was working out why robots were killing people in factories.It closes on something with nothing to do with technology at all, and that is the best part of it.

Part of our GAEA Talks Forrester season, recorded in Austin, Texas, and the second in the season with a practitioner rather than an analyst.Graeme Scott sits down with Arturo Lopez-Brea Gimenez, Head of Digital and IT Transformation and Capabilities at Moeve, the Spanish energy company formerly known as Cepsa. Arturo studied a dual degree in business and law, worked at Santander and PwC, and joined the company on its graduate programme. He moved into the digital transformation team as it was being built, helped launch the company's own digital university, and has led transformation and capabilities since 2022.Moeve is going through what he calls a triple transformation. The energy transition itself, as customers and regulators demand greener energy and mobility. The company's own shift, from a ninety six year old oil and gas business to one targeting more than half its EBITDA from sustainable energy by the end of the decade, reflected in the rebrand from Cepsa. And digital transformation as the accelerator of both. He is clear they are not three separate programmes.His position on where the difficulty actually lies is blunt. Deploying technology is easy and getting easier, in cost, in implementation, in complexity. The challenge is people. Not money, not investment, but convincing colleagues who have worked the same way for fifteen or twenty years to use something new and to trust what comes out of it.The centre of this episode is an experiment his team ran, and it is the most useful piece of adoption evidence anyone has brought to this podcast. Sixty one participants, four weeks, one goal each: build an agent that solves a real problem in your own job. Half went through a guided programme with mentoring and continuous expert support. Half had identical technology, identical time, identical objectives, and almost no support. The completion rates diverged sharply, and so did something less expected. The unsupported group were twice as likely to distrust the output of the agents they had built themselves.The reported time savings are where it gets genuinely instructive, because the raw numbers point the wrong way until you put them in context. He explains how, and why the contextualised figure is the only one that matters if you are trying to drive adoption at scale.He is also precise on a distinction most organisations get wrong. Providing a licence tells you how many people can access something. Usage tells you how many do. Adoption is whether it has changed the way they work. When colleagues answer a question about adoption with a licence count, they have answered a different question.