In 2015, Demis Hassabis looked back at a decision he had made more than a decade earlier. By 2004 he was already thinking about what would eventually become DeepMind, but concluded that the technology was not ready. As he later put it, “we still didn’t have enough of the components.” Deep learning had not yet taken off, computing power was inadequate and many of the systems needed to build something approaching general intelligence simply did not exist.
By 2026, Hassabis was describing a fuller set of ingredients: Gemini as a world model, AlphaGo-style search and planning, specialist tools, memory, reasoning and systems capable of acting in the world. The possibility is that Google has not recently changed its theory of artificial general intelligence. It may finally have enough of the pieces to execute a theory DeepMind has pursued for years.
That matters beyond the contest among AI labs. Falling AI costs make individual tasks cheaper to automate, but cheap task execution does not necessarily replace a worker. Jobs are bundles of tasks. Planning, memory, simulation, tool use and action let an AI system move across that bundle. If Google can combine them reliably, the shift from AI as a tool to AI as a worker could happen before anyone agrees that AGI has arrived.
DeepMind was never really building a chatbot
DeepMind’s early history looks very different from modern large language models. Its first major breakthrough was not language generation but an agent learning how to act. In 2013, DeepMind showed that a neural network could look at raw pixels from Atari games, decide what action to take, observe the result and improve from the reward. Hassabis later described the system as moving from “pixels to action”.
Planning became another strand of the same programme. AlphaGo combined neural networks with search, examining possible future positions before deciding what to do. In 2016 Hassabis said one purpose was to see how neural networks could be “meshed with planning approaches”. A year later DeepMind was experimenting with imagination-augmented agents that simulated possible futures before acting.
DeepMind was also working on memory and continual learning. The projects differed technically, but together they reveal a consistent view of intelligence: a capable system would need to perceive, remember, plan, imagine, act and learn over time. What these systems lacked was generality. AlphaGo could defeat the world’s best Go players while being useless at almost everything else.
Also Read: Does the iPhone show us which industries AI will destroy?
Then the LLM arrived
Modern large language models changed that part of the problem. Google Brain’s Transformer architecture, published in 2017, became the basis for systems trained across enormous quantities of language, code, images and other information. As they scaled, these models developed broad competence across subjects and tasks that earlier DeepMind agents could not approach.
Former DeepMind researcher Misha Laskin has described the distinction neatly: language models provide “generality”, while reinforcement learning provides “capability”. DeepMind had systems that could act effectively but only inside narrow worlds. LLMs were almost the reverse: remarkably broad, but unreliable when required to pursue objectives autonomously across long sequences of actions.
In 2023 Google combined the two research traditions institutionally when it merged Google Brain and DeepMind into Google DeepMind. Gemini appeared soon afterwards. From the beginning, Google described it as multimodal, designed to work with tools and intended eventually to support memory and planning. Hassabis went further, describing Gemini as combining some strengths of AlphaGo-type systems with the language capabilities of large models.
That makes Gemini easier to understand as more than Google’s answer to GPT. The large language model may instead be the general-purpose cognitive core around which DeepMind’s older work on search, planning, memory and action can finally be assembled.
The old architecture is reappearing
Google’s recent projects increasingly resemble components DeepMind once had to investigate separately. Hassabis has said the ambition is a “world model, not just a language model”. Google’s Genie project explores that direction by generating interactive environments in which actions alter subsequent events. SIMA supplies the agent side: Google has combined Gemini with SIMA so an AI can interpret objectives, reason about them and act inside 3D environments.
Deep Think attacks another weakness of conventional language models by exploring several possible reasoning paths before committing to an answer. Robotics extends the same structure into the physical world. Gemini Robotics ER 2 uses Gemini 3.5 Flash as a high-level reasoning system that can interpret a situation, plan multiple steps and coordinate lower-level models that produce physical actions.
That modularity may be crucial. Human workers do not personally contain every capability they use. They understand an objective, choose a tool, interpret its output and decide what to do next. An artificial worker could operate similarly. This also explains Hassabis’s emphasis on specialist tools: the general model need not contain the best protein-folding system or robotics controller. It needs to know when to call one and what to do with the result.
Also Read: The AI boom won’t help you scale. Your unit economics will
The economic threshold is below AGI
The important economic threshold is much lower than full human-level intelligence. Consider an analyst. The job may require understanding an objective, finding information, deciding what matters, operating software, creating an output, checking it, recognising mistakes, revising the work and deciding when it is finished. An AI system that can draft one paragraph cheaply automates a task. One that can perform most of that sequence reliably begins to compete with the role.
That is why falling cost and expanding capability matter together. Inference has been getting cheaper. At the same time, search, planning, world models, tools and memory expand the number of adjacent tasks AI can perform. More of the job therefore becomes technically automatable just as the price of automating it falls.
Better reasoning often requires more computation: several candidate solutions, verification and repeated tool calls. If inference remains expensive, those techniques can make autonomous AI uneconomic. As inference prices fall, more computation can be spent making the system reliable while the total cost of completing the work still declines. Different occupations can therefore cross the threshold at different times, before any universal AGI moment.
Google may still fail
None of this means Google has solved the problem. World models can be wrong. Search can amplify errors. Reinforcement learning works best where success can be measured clearly, whereas many real jobs involve ambiguous objectives and human judgement. Long autonomous workflows accumulate errors, and persistent memory can preserve bad information as easily as useful experience.
Google itself is also showing signs of execution difficulty. Gemini 3.5 Pro missed its expected release window. In August Hassabis handed day-to-day leadership of Google DeepMind to Koray Kavukcuoglu and became Alphabet’s Chief Scientist. Google says the change allows Hassabis to concentrate more heavily on long-term AGI research. A less favourable interpretation is that the long-horizon research programme and commercial Gemini roadmap are becoming more separate.
The architecture may never converge in the way the historical pattern suggests. Genie, SIMA, Deep Think, robotics and memory could remain distinct research programmes while Gemini itself improves mainly through better training and more compute. If AI becomes cheaper and better at individual tasks but remains unreliable when asked to complete long workflows without supervision, the thesis will have failed its clearest test.
But the history makes Google’s current direction more credible than recent announcements alone would suggest. DeepMind was pursuing reinforcement learning, memory, planning, simulated futures and action before modern LLMs existed. Those systems lacked broad generality. Transformer-based models supplied much of that missing capability. Google then merged the two research traditions and began building world models, agents, reasoning systems and robotics around Gemini.
That does not prove Google has discovered the route to AGI. It does suggest that its current work is not merely a collection of fashionable additions to a chatbot. Google may be assembling the machine DeepMind wanted to build before the components existed. If it works, the economically important date will probably not be the day Google announces AGI. It will be the quieter point when an AI system becomes cheap, reliable and broadly capable enough to do the whole job.
For more information on this proprietary research framework analysing AI-driven economic change, please visit here.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post Google may already be showing us its route to AGI appeared first on e27.