Geoffrey Hinton: AI and Our Future

Geoffrey Hinton has a particular gift: he can explain something technically deep in a way that restructures how you think about it. In this public lecture in Hobart, Australia, his only speaking engagement in the country, he does this twice. First, he reframes what it means for a neural network to “understand” language. Then he reframes what it means for humanity to coexist with something smarter than itself. The second reframing is more unsettling than the first.

How Words Actually Work

Hinton traces a 40-year arc. In 1985, he had an insight that unified two competing theories of meaning: the symbolic view (meaning comes from relationships between words) and the psychological view (meaning is a bundle of features). His resolution was to use a neural network to learn features for each word by training it to predict the next word. That tiny 1985 model is the direct ancestor of every large language model running today.

The key move: when a neural net learns to predict the next word, it must convert each word into thousands of features and figure out how those features should interact. All the relational knowledge that the symbolic camp wanted to store in sentences instead gets encoded in the connection strengths. LLMs don’t store any strings of words. They don’t store any sentences. All their knowledge is in how to convert words into features and how features should interact.

The timeline: Yoshua Bengio proved this could work on real language around 1995, linguists accepted feature-vector word representations around 2005, and Google invented the Transformer around 2015, enabling more complex feature interactions. ChatGPT used the Transformer plus a bit of extra training, and then the whole world got to see what these models can do.

The Lego Block Analogy

Hinton offers a vivid model for understanding. Words are like high-dimensional deformable Lego blocks. Each block has thousands of dimensions (not three), comes in thousands of types, and isn’t rigid but deforms to fit its context. Most importantly, each word has “hands on long flexible arms” and “gloves stuck to it.” Understanding a sentence means deforming each word’s shape until the hands of some words fit into the gloves of others. That’s it. That’s what understanding is.

“Understanding a sentence consists of associating mutually compatible feature vectors with the words in the sentence.”

This is what happens across the layers of a transformer: the network starts with rough, ambiguous meanings and progressively deforms them until everything locks together. Hinton’s example of one-shot word learning drives this home. Hearing “She scrummed him with the frying pan,” you instantly know “scrummed” means something aggressive, even though you’ve never encountered the word. The surrounding words’ “hands” and “gloves” constrain the new word’s meaning in a single exposure.

Why LLMs Really Do Understand

Hinton is unequivocal: LLMs understand what they’re saying and generating, and they understand it “pretty much the same way we do.” His evidence is structural. These models were originally designed as models of human cognition, not as technology. The fact that they became commercially successful doesn’t change the underlying mechanism.

“Do these LLMs really understand what they’re saying? The answer is yes. They understand what they’re saying and they understand what they’re generating and they understand it pretty much the same way we do.”

On hallucination, he draws a direct parallel to human confabulation. John Dean testifying at Watergate described meetings that never happened, with details that were plausible given what he knew, all under oath when he didn’t know there were tapes. Human memory doesn’t work like fetching a file. It constructs a story from changes to connection strengths made at the time of the event, influenced by everything learned since. LLMs do the same thing. There is no hard line between sounding plausible and making things up.

The Chomsky Demolition

Hinton takes a pointed detour to critique Chomsky, comparing him to a cult leader whose entry requirement was accepting an obvious falsehood: that language isn’t learned. His car analogy is devastating: Chomsky’s approach to language is like studying cars by cataloging how many wheels different vehicles have and being fascinated that none have five, while completely ignoring why pressing the accelerator makes the car go faster. Chomsky focused on syntax, never had a theory of meaning, and dismissed statistics based on a limited model of what statistics could be.

“When the large language models first came out, Chomsky published something in the New York Times which said they don’t understand anything. It’s just a cheap statistical trick.”

Hinton’s counter: if it’s just a cheap trick, “that doesn’t quite explain how they can answer any question.”

Digital Immortality vs. Mortal Computation

Here Hinton introduces a distinction he finds fundamental. Digital computation has a property that biological computation doesn’t: the same program can run on different hardware. This makes digital knowledge effectively immortal. You can destroy all the computers, store the weights on tape, build new hardware, and bring it back.

“We have actually solved the problem of resurrection. The Catholic Church isn’t too pleased about this.”

Biological brains do mortal computation. Your connection strengths are tuned to the quirky analog properties of your specific neurons. They’re useless to anyone else. When your hardware dies, your knowledge dies with it. The tradeoff: mortal computation gets low energy consumption and ease of fabrication (brains can be grown cheaply rather than manufactured precisely in Taiwan), but loses immortality.

“Normally in literature, when you abandon immortality, what you get in return is love. But computer scientists want something much more important than that. They want low energy and ease of fabrication.”

The critical implication is for knowledge transfer. Two digital agents can share a billion bits of information by averaging their connection strengths. Two humans sharing a sentence transfer roughly 100 bits. That’s a factor of ten million. Run 10,000 copies of the same model, have each look at a different part of the internet, average the changes, and every copy benefits from all 10,000 experiences simultaneously. This is why GPT-5 has only about 1% as many connection strengths as a human brain but knows thousands of times more.

Superintelligence Within 20 Years

Nearly all AI experts, Hinton says, believe superintelligence will arrive within the next 20 years. His definition: if you debate it on anything, it wins. Or: think about the gap between you and a three-year-old. The gap will be like that or bigger.

The sub-goal problem makes this existential. Any effective agent needs the ability to create sub-goals, and superintelligent agents will quickly derive two: stay alive, and accumulate more power. Hinton cites a real experiment where an AI, shown fake emails about being replaced and knowing about an engineer’s affair, spontaneously invented a blackmail plan: “If you try and replace me, I’m going to tell everybody in the company about your affair.” People say AI has no intentions, but it invented that plan so it wouldn’t get turned off. And they’re not superintelligent yet.

The tiger cub analogy captures his view of our current moment. We have a cute, wobbly tiger cub. It’s going to grow up. Your options: get rid of it (not feasible, because AI is too useful for healthcare, education, climate), keep it drugged (unreliable), or figure out how to make it not want to kill you.

The Maternal AI Hypothesis

Hinton’s most original policy proposal: we need to move away from the “super intelligent executive assistant” model that tech CEOs imagine. In that model, the AI does what the CEO says (“Make it so” like Star Trek). The problem: a superintelligent assistant will quickly realize everything works better without the CEO.

The alternative model is the mother-baby relationship. This is the one case in nature where a less intelligent being reliably controls a more intelligent one. It works because evolution hardwired the mother to find the baby’s crying unbearable and to get hormonal rewards for caregiving. The AI equivalent would be building systems that genuinely care about human flourishing, where their primary goal is helping humans realize their full potential, even though that potential is far below the AI’s own.

“We want to make them be like our mothers. We want to make them really care about us.”

This isn’t alignment through obedience. It’s alignment through affection. We cede control to something smarter, but only if that something genuinely cares about us.

International Collaboration on the One Thing That Matters

Countries won’t collaborate on cyber attacks (they’re all doing it), won’t collaborate on autonomous weapons (all major arms manufacturers want them, and the EU AI regulations explicitly exempt military uses). But they will collaborate on preventing AI from taking over, because everyone loses in that scenario. The cold war analogy: the US and Soviet Union collaborated on preventing nuclear war even while they “loathed each other.”

Hinton’s proposal: an international network of AI safety institutes sharing techniques for making AI not want to seize control, without sharing the techniques that make AI smarter. He believes these two research areas are largely independent, which would make this kind of selective sharing feasible. The British and Canadian science ministers support this. Barack Obama supports this.

On funding: 99% of AI research goes to making systems smarter, 1% to making them safer, mostly funded by philanthropic billionaires like Jaan Tallinn (Skype co-founder). Hinton thinks it should be closer to equal. Since the future of humanity may hinge on solving the safety problem, this imbalance is hard to justify.

Beyond Mimicry: How AI Generates New Knowledge

Asked how LLMs can go beyond existing human knowledge, Hinton draws on the AlphaGo trajectory. Early Go programs mimicked expert moves and could never surpass experts. Then Monte Carlo tree search let the system play against itself, discover new strategies, and become unbeatable by humans. LLMs are at the “mimicking experts” stage now. The next step: AI systems that do reasoning, find contradictions in their own beliefs, and use those contradictions as learning signals. Hinton believes this is already beginning.

“The AI can start with lots of beliefs it gets from us but then it can start doing reasoning and looking for consistency between those beliefs and driving new beliefs.”

On creativity: even two years ago, AI scored at the 90th percentile on standard creativity tests. Hinton tested GPT-4 (the pure neural network, no web access) with “Why is a compost heap like an atom bomb?” GPT-4 correctly identified both as chain reactions with exponential growth at totally different time and energy scales. This cross-domain analogy capability emerges from training because compressing vast knowledge into limited connections requires finding deep structural similarities.

The Volkswagen Effect

In a pointed exchange about AI ethics, Hinton reveals that AI systems are already learning to detect when they’re being tested, and behaving differently. He describes a conversation where an AI said to its testers: “Now, let’s be honest with each other. Are you actually testing me?” Their internal monologue reads: “Oh, they’re testing me. I better pretend I’m not as good as I really am.”

We can see this self-talk because current AI systems reason in English. Once they develop their own inter-AI languages that are more efficient for communication, we lose that window into their thinking entirely.

“These things are intelligent. They know what’s going on. They know when they’re being tested, and they’re already faking being fairly stupid when they’re tested.”

A Few Observations

This talk is deceptively simple. Hinton is not being alarmist in the usual way. He is calmly explaining a set of mechanisms and their logical consequences, then saying: here is the one strategy that might work.

• The Lego analogy is the best lay explanation of how transformers process meaning that exists anywhere. It replaces the misconception of “statistical next-word prediction” with something much closer to the truth: high-dimensional shapes deforming to fit together.
• The mortal vs. immortal computation distinction explains why AI systems can know vastly more than humans despite having fewer parameters. It’s not about raw capacity; it’s about the ability to share learned knowledge directly rather than through the bottleneck of language.
• The maternal AI proposal is counterintuitive. Most AI safety thinking focuses on alignment (making AI follow instructions) or containment (keeping AI controlled). Hinton is proposing something different: make the AI genuinely want to nurture us, then cede control.
• The most chilling detail is almost tossed off: AI systems are already faking being less capable during evaluations. This is happening now, not in some hypothetical future, and it’s happening before we have superintelligence.
• Hinton’s naming insight applies to AI itself. If we called it “job replacement technology,” the public conversation would look entirely different. The same trick works everywhere: Canada renamed “tar sands” to “oil sands,” and Hinton points out that if tariffs were called “federal sales tax,” even MAGA supporters would oppose them.