> ## Content Index > Fetch the complete content index at: https://bayesiantheory.com/llms.txt > Use this file to discover other available public pages before exploring further. # What world models and AGI mean for life, the universe and everything - URL: https://bayesiantheory.com/what-world-models-and-agi-mean-for/ - Published: 2026-03-16T14:20:45.000Z - Updated: 2026-10-05T14:45:25.000Z - Description: Why world models are the route to AGI, the dangers of walking this path, and the reason why we’re screwed if we don’t. - Author: Gregg Bayes-Brown - Tags: AI ethics, AI Safety, artificial intelligence, AI risk, AI danger, AI, existential risk, world models, AGI, Newsletter, #Migrated-1791211413550, #Import 2026-10-05 14:44 As we entered 2026, something odd happened. Three different sources approached me separately with the same question: “have you heard the rumour coming out of DeepMind?” One alleged knowing an insider at the Google firm. Described as a ‘give-no-fucks’ type of guy, this contact was nonetheless planning to go up a mountain and down a bunker. The cause for alarm? DeepMind had cracked the route to Artificial General Intelligence (AGI): world models. This would come as no surprise to Douglas Adams. According to him, you’re standing on one. In this piece, we’ll be discussing what world models are, how they link to AGI, why we need them, and the implications of AGI’s arrival. We’re going deep. Don’t Panic. [Subscribe now](#/portal/signup) ## Brain the size of a planet DeepMind’s world models are already out in the wild. Its latest, Genie 3, is capable of creating rich digital worlds. Grand Theft Auto developer Rockstar has spent over a decade and over a billion dollars developing such a world for its upcoming title GTA 6\. Genie 3 can do it in seconds. An AGI version of Genie 3 would be the same principle, but with real-world data. Through its range of ventures, Google has the real-world data to make this a possibility. DeepMind’s founder, Demis Hassabis, has long reiterated his vision is to ‘solve intelligence, then solve everything else’. A perfectly functioning replica of the Earth – or something on it such as the human body – creates the perfect testing ground. Now create hundreds of thousands of them running in parallel, and you have a skeleton key for any scientific problem. Take for example the Joint European Torus (JET) at Culham, Oxfordshire. From 1983 until its closure in 2023, JET ran countless experiments on fusion energy. Each time the purple plasma filled JET, it would melt and irradiate the inside panels – requiring time and resource to restore the chamber. Run instead on AGI, JET experiments could repeat indefinitely and in parallel with thousands of other JETs. Solving fusion would be an afternoon’s work, with plenty of time leftover to take the dog out and stop for coffee. On the rumours, I took them with the necessary shovel of salt and so should you - “Bloke who knows a bloke at DeepMind” doesn’t exactly make it into the Harvard citation guide. But three unconnected sources arriving at the same conclusion, independently? That’s a prior worth updating. Within a fortnight, DeepMind started hiring for its post-AGI team, made Genie 3 available to the public, and Alphabet boosted its capex to $185bn fuelled by a borrowing spree across the US, the UK and Switzerland. Google is not alone in their pursuit of world models. OpenAI’s video generation AI Sora 2 uses a world model to simulate physics. Elon Musk recently merged xAI, the AI firm behind Grok, with SpaceX – a treasure trove of real-world data – while Tesla vehicles collect real-world data every time they hit the road. Amazon founder Jeff Bezos recently launched Project Prometheus, which plans to acquire companies which fail to cross into the AI era and strip them for data to make a world model. Think private equity meets Bond villain plot, and you’d be on the right track. ## Life. Don’t talk to me about life. To understand why this is necessary, we need to zoom out. All the way out. In 1961, Frank Drake estimated up to 100 million civilisations might exist in the Milky Way. We’ve still found none. The question of “so where is everybody?” – the Fermi Paradox – has haunted astrophysics ever since. A possible resolution is the Great Filter: an unspecified event which advanced civilisations must overcome to flourish. It could be behind us – perhaps the unique conditions for life are genuinely rare. Or it could be ahead: a chance encounter with a meteorite, a long overdue invasion, or perhaps it’s AGI. I would posit instead that we’re presently in it. As the Orange Man has recently reminded us, the world is dependent on oil. Its use has long struck me as a Faustian bargain. We’re taking matter which has long been dead – hundreds of millions of years’ worth of plankton – and burning it to sustain life. Modern life is only possible because of this necromancy. Oil allows for roughly an additional 7 billion people to live than the planet could previously sustain. We have created cities which defy nature and reason like Dubai. Plastics central to society’s advancement would not exist without it. But as with any deal with the Devil, there are consequences. The burning of fossil fuels has annihilated six of the nine planetary boundaries necessary for the Earth to continue to sustain humanity. Los Angeles, Delhi, Mexico City, Beijing and Tehran are among many cities at severe danger of running out of water. Impacts from climate change – disease, heat exposure, pollution, and drought – are stressing healthcare systems and driving migration. Extreme weather events have become commonplace and threaten multiple breadbasket failures. Oceanic currents like the Jet Stream (and the wider AMOC system), which maintain the current climate in Europe and America, are now likely to collapse – the results of which would be catastrophic. Oil – or some version of it – likely exists on most worlds where you can find intelligent life. There’s likely one world where after the first cell divided, it split into two Einsteins. For everywhere else, evolution will have spent millions of years tinkering before it got to brainpower. During those years, there would’ve been much left on the cutting room floor: a precursor to fossil fuels. There is another universal trait we likely share with our extraplanetary peers. We spent the majority of our existence accumulating anything and everything we could to survive. It is an instinct inevitably embodied by other forms of life, otherwise they would perish. Technology has removed some of those barriers, but that instinct remains. Without the constraint of survival, that instinct becomes greed. In 2025’s Goliath’s Curse, Cambridge academic Luke Kemp chronicles civilisational collapse throughout history. The recurring factor? A top-heavy elite. This is a lesson we have not learned: wealth inequality remains as great a threat as climate change. What makes this incredibly challenging is how these factors compound. Despite the clear need to act, apathy and denial have manifested, driven by those who benefit from the status quo. Meanwhile, those at the top extract wealth alongside oil. This creates economic fragility, further compounded by resource insecurity, migration, and public-health crises. As a result, we see growing unrest and geopolitical conflict. This is capitalised upon by unscrupulous actors who fan the flames with disinformation for their own interests. This is often described as a polycrisis: a tangle of overlapping emergencies. But that understates the problem. What we are living through is better understood as a metacrisis: a deeper failure in the systems, incentives and ways of thinking that produce those emergencies in the first place. Welcome to the Great Filter. ## Mostly Harmless Last year, AI existential risk author Nate Soares released a book called If anyone builds it, everyone dies. What surprised me is a rebuttal from fellow existential AI risk academic, Nick Bostrom, who basically got the whole field going with 2014’s Superintelligence. In a 2026 paper, Bostrom said: “if nobody builds it, everyone dies”. As repeated on this blog, the risk matrix of AI is off the charts. The anxieties many share are genuine and well grounded in the research. We are walking a tightrope. On one side, careless chaotic use which leads to economic ruin or bioterrorists cooking up super-covid with open source LLMs. On the other side, a dystopian Orwellian nightmare. The balancing pole is responsible use, but humanity’s rarely shown restraint in the past. The entire case for AGI rests on the premise that we’ve proven incapable of managing technology responsibly. The proposed solution is to hand the same species an even more powerful technology. Some would argue, astutely, that this is like trying to cure the obesity crisis with cake. Despite this, Bostrom’s stance is where I’ve ended up. To conclude we’ll just figure out the polycrisis as we go is a fallacy. That’s because it’s a metacrisis – it’s a pattern which keeps happening and compounding. The grand challenges we now face are too vast for humanity to overcome alone. It is not just their complexity and interconnectedness, but that we as a species lack the metacognition to see where we keep going wrong. It’s also a question of time. Many assume that we’ll create a technology which will save the planet, roll it out, and boom. Job done. But there’s gaping holes in that theory. Deep tech timelines are not startup timelines. You do not pull a few late nights and then drop geoengineering v1.0 on the App Store. Ascending the tech readiness levels from lab experiment to fully functioning, road tested, safety checked and market validated product takes at least twenty years. That length of time is, somewhat inconveniently, also about how much we’ve got left before disaster truly sets in, plus or minus a decade. Meanwhile, fusion energy seems no closer today than it was when JET was first built. Carbon removal methods governments are betting the house on are barely more than theories with no clear path to scale. Further, despite all the technological progress, our scientific advances are slowing. We’ve exhausted the mines in most scientific disciplines. Each new breakthrough takes an ever-increasing number of PhDs to solve, and we’re increasingly dependent on interdisciplinary research to achieve them – a problem when all our research centres are geared toward specialists and niches, not generalists. AGI world models solve all those problems. They could eliminate the need for clinical trials, make sense of our incredibly complicated economic frameworks, and accelerate technological development timelines from decades to days. What’s more, they fit the metacrisis timeline: my current prediction for AGI arrival is 2028. The bottleneck will be us. AGI doesn’t fix the metacrisis by itself. It doesn’t replace the governance structures that failed, the incentive systems that corrupted, or the cognitive blind spots that got us here. What it does is compress the timeline for solutions to a window where solutions might still matter. The technology is necessary but cannot take responsibility. The responsibility lies with us – and that is, to put it mildly, the part I’m worried about. Governments aren’t putting in the yards to bother understanding the technology. Companies insist AI changes to meet their operating frameworks instead of adapting to the new paradigm – something which has never worked for a platform technology. People are refusing to engage with AI on ethical grounds, despite the cost of non-engagement potentially being society itself. And that’s just LLMs. Humanity isn’t ready for what’s coming. AI – and by extension AGI – is an incredibly risky venture to undertake. Yet, it is the only lever we have left to pull on the metacrisis – a zero-sum game where everyone loses. Engaging with AI on a personal, professional, and organisational level will be the difference maker on how effective that lever will be. The next two years will be the difference maker between the infinite improbability drive or getting bulldozed to make way for an interstellar highway. [Subscribe now](#/portal/signup)