AGI isn’t coming, it is already here: Google’s Blaise Aguera y Arcas

“We already have AGI”, the definitive verdict of Blaise Agüera y Arcas, VP, Fellow, and CTO of Technology & Society at Google, where he leads the Paradigms of Intelligence (Pi) team on research in artificial intelligence (AI). In an exclusive conversation with HT, Blaise marks a significant departure from the industry’s fixation on artificial general intelligence (AGI) timelines, which has seen many tech and AI executives express different timelines for when they believe AGI will become a reality.

Blaise believes in-context learning as a method may provide a real breakthrough for large models of the future, to learn rapidly. (Official photo)

Even from his vantage point within the AI space, Blaise admits he was as surprised as anybody, that unsupervised learning produces general intelligence. Recognised as one of the inventors of federated learning, having first introduced the concept in 2016, Blaise believes in-context learning as a method may provide a real breakthrough for large models of the future, to learn rapidly.

“The reason I say this is because the ‘G’ in AGI means general. The moment we began to train models like LLMs that are not artificial, narrow intelligence like face recognition, handwriting recognition, speech recognition or image synthesis, and train them in an unsupervised way, we got rid of the narrow. That is when they do become general,” he tells us, insisting this destination is already behind us.

Tesla and X chief Elon Musk, for instance, believes AGI will be prominent by next year. Nvidia’s CEO Jensen Huang believes AGI could be achieved by the year 2029, while OpenAI’s Sam Altman has offered varying timelines in the past, with the latest estimate pegged for human-level AI intelligence by 2026. Microsoft’s AI chief Mustafa Suleyman anticipates an AGI timeline of around 10 more years. The reason Blaise believes most people don’t realise that we’ve already crossed into that territory, is due to the lack of a “moment” that would signify that transition.

In his latest book titled What is Intelligence?, Blaise argues the “Winograd Schema challenge,” introduced in 2011 by Canadian computer scientist Hector (this was positioned as an alternative to the Turing Test for AI benchmarking) began to be decisively defeated sequence models by the year 2019. “Its defeat roughly coincided with the arrival of “real”AI, or Artificial General Intelligence (AGI), as one would expect; since then, we have simply been moving the goalposts,” he says.

There was no cinematic flip from ‘not AGI’ to ‘AGI,’ — only a steady ramp from unreliable early systems like LaMDA (2021) to far more capable frontier models today. By his historical yardstick, if a system like Google’s Gemini model as it is today was taken back to the year 2000, when AGI was coined to distinguish ‘real AI’ from narrow recognisers, few would doubt the label.

He concedes gaps, such as the lack of durable long-term memory, occasionally surprising errors, architectures unlike brains—but says those don’t negate generality. “I’m more expecting a continuation of the ramps that we’re already on,” his vision of things to come.

Training and responsible guardrails

Blaise believes it is important to be “bold and responsible in innovation”. While not wanting draw into comparisons between Gemini and its rivals from AI companies including OpenAI, Meta and Anthropic, he admits Google’s sense of responsibility often defines the pace of product releases, which is often “frustrating even to me, as as a technologist and somebody who always wants people to see and experience the newest thing.”

Guardrails in AI, Blaise says, are essential precisely because of the technology’s scale. When billions of people interact with models daily, even a tiny fraction of problematic outputs translates into real harm. Missteps, including disturbing conversations with vulnerable users or inappropriate responses, are inevitable at such scale. The responsibility, then, lies in how AI companies detect, correct, and learn from them. “The best that a company can do is be aware, responsive, and thoughtful,” he says.

Recently, an investigation was opened into how Meta’s AI chatbots were allowed to have ‘sensual’ chats with children, after a leaked internal document detailed instances. The internal Meta Platforms policy document also details how the social media giant’s AI is prone to detailing incorrect medical information, and indulge in provocative interactions on topics including sex, race and celebrities.

Diminishing returns from models and data sets is a reality AI companies are struggling with now, after years of a generative AI race that relied on bigger models trained on ever-larger datasets. OpenAI, Meta, Anthropic, and Google themselves, have poured a lot of resources into scaling with more data, more parameters in the hope of better reasoning capabilities. The plateau-ing has been acknowledged by AI companies recently.

Is the age of giant AI models already over? Blaise believes the key for large models of the future is in-context learning. “I think it is not obvious to a lot of people how powerful in context learning is. When you take a pre-trained model and then you put something in the context window, it can learn extraordinarily efficiently from what is in the context window,” he says, having helped pioneer privacy-preserving learning through Federated Learning. That means, the model can learn without retraining or altering its parameters.

Some of the examples he sights include something simpler such as telling an AI model about a word as well as its meaning in a conversation and it’d learn that even though the weights haven’t changed, or a more complex illustration, drawing from an experiment they did was to take a translated version of the New Testament in a low resource language, and the model began to converse in that language reasonably well.

Blaise identifies a paradox in the current scaling paradigm. When models are pre-trained on billions of tokens randomly sampled from the internet, their first exposures teach them the most — every sentence is new. But as training continues, repetition dominates; the model encounters fewer novel patterns and begins to learn less efficiently.

“As the progress gets slower, it seems like you need a ton of data. To contrast with in-context learning, where a tiny amount of data goes a long way. This is why very highly curated data sets have actually turned out to be super important for reinforcement learning,” he says. Blaise believes diminishing returns from training does not mean a model is getting worse and worse at learning. On the contrary, it’s that the data is more repetitive.

The future he envisions departs sharply from the brute-force scaling era. Instead of endlessly retraining on “everything everyone has ever written,” he imagines models that learn actively and selectively, much like humans choosing what to read or study. That’s also why Google’s evolving Gemini and DeepMind’s research into reasoning agents represent more than product iterations. It signifies a philosophical shift, one that moves away from an idea that intelligence comes from sheer size, towards one where intelligence comes from efficiency.