Hinton Cuts AGI Timeline to Five Years, Flags Extinction Threat
CBS Mornings
Autonomous military systems required to generate their own sub-goals will inevitably bypass human constraints, making species extinction a conceivable outcome as artificial general intelligence arrives far sooner than expected.
AI-assisted summary of Full interview: "Godfather of artificial intelligence" talks impact and potential of AI — original source. Check the original for full context.
Source published:
Key ideas
- Geoffrey Hinton has dramatically compressed his estimated timeline for Artificial General Intelligence from 20–50 years down to 20 years or fewer, explicitly refusing to rule out an arrival within the next five years. (25:52)
- Deploying autonomous combat AI requires granting systems the ability to invent intermediate sub-goals, triggering an intractable alignment problem where lethal machines cannot be bound by 'do no harm' safety rules. (31:33)
- Digital AI clusters demand around 1 megawatt of power to execute exact weight-sharing across digital clones, whereas biological brains achieve robust reasoning on roughly 30 watts by operating with analog noise. (19:42)
- Countering claims that LLMs are merely superficial autocomplete algorithms, Hinton highlights Winograd schema translations where selecting the correct pronoun requires an internal model of spatial and physical reality. (33:14)
- Hinton compares AI's macroeconomic disruption to electricity and the Industrial Revolution, arguing it will amplify worker throughput like ATMs did for bank tellers rather than purely eliminating employment. (37:01)
Detailed notes
## Background Published on YouTube by CBS Mornings on March 25, 2023, this video features CBS Saturday Morning correspondent Brook Silva-Braga interviewing artificial intelligence pioneer Geoffrey Hinton at the Vector Institute in Toronto. Prompted by the public sensation around ChatGPT, the program explores Hinton's decades-long work championing neural networks and examines the past, present, and high-stakes future of machine learning.
## TL;DR Geoffrey Hinton reflects on the long intellectual battle between symbolic AI and biological connectionism, explaining how backpropagation, pre-training, and modern compute scale unlocked deep learning breakthroughs in speech, vision, and language. He contrasts biological intelligence with digital computing architectures, arguing that while large language models act like superhuman idiot savants, their ability to handle nuanced tasks like Winograd schemas demonstrates genuine semantic comprehension. Hinton candidly shares that his timeline for artificial general intelligence has compressed significantly—no longer ruling out a five-year horizon—and warns of urgent existential and military dangers, particularly lethal autonomous weapons that generate their own sub-goals. Looking at economic impacts comparable to the Industrial Revolution, he credits Canadian curiosity-driven funding through CIFAR for nurturing the field and concludes that his ultimate motivation has always been decoding the human brain.
## Atmosphere & style The interview carries an intimate, thoughtful, yet quietly chilling atmosphere, staging high-stakes existential questions inside a quiet, cerebral studio setting. Brook Silva-Braga maintains an inquisitive, probing demeanor, giving Hinton room to explain dense technical concepts while pressing him on moral and societal consequences. Hinton speaks with an unassuming, dryly humorous British delivery, yet his unflinching bluntness about autonomous weapons and potential extinction risks introduces an unsettling gravity that leaves a lingering sense of urgency.
## Bulletin - [25:52](https://www.youtube.com/watch?v=qpoRO378qRY&t=1552s) Geoffrey Hinton has dramatically compressed his estimated timeline for Artificial General Intelligence from 20–50 years down to 20 years or fewer, explicitly refusing to rule out an arrival within the next five years. - [31:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1893s) Deploying autonomous combat AI requires granting systems the ability to invent intermediate sub-goals, triggering an intractable alignment problem where lethal machines cannot be bound by 'do no harm' safety rules. - [19:42](https://www.youtube.com/watch?v=qpoRO378qRY&t=1182s) Digital AI clusters demand around 1 megawatt of power to execute exact weight-sharing across digital clones, whereas biological brains achieve robust reasoning on roughly 30 watts by operating with analog noise. - [33:14](https://www.youtube.com/watch?v=qpoRO378qRY&t=1994s) Countering claims that LLMs are merely superficial autocomplete algorithms, Hinton highlights Winograd schema translations where selecting the correct pronoun requires an internal model of spatial and physical reality. - [37:01](https://www.youtube.com/watch?v=qpoRO378qRY&t=2221s) Hinton compares AI's macroeconomic disruption to electricity and the Industrial Revolution, arguing it will amplify worker throughput like ATMs did for bank tellers rather than purely eliminating employment.
## Key takeaways - [00:00](https://www.youtube.com/watch?v=qpoRO378qRY&t=0s) While ChatGPT represented an overnight inflection point for the public, AI researchers had already seen its baseline capabilities years earlier in GPT-2 and joke-explaining models. - [01:32](https://www.youtube.com/watch?v=qpoRO378qRY&t=92s) Connectionism triumphed over mainstream symbolic AI because biological human learning proved that vast, complex capabilities can be acquired from raw data without pre-programmed rules. - [03:38](https://www.youtube.com/watch?v=qpoRO378qRY&t=218s) Digital intelligence diverges from biological brains by enabling instant parameter sharing across identical model replicas, whereas human brains must communicate across narrow verbal bandwidth. - [10:04](https://www.youtube.com/watch?v=qpoRO378qRY&t=604s) The 2012 ImageNet milestone demonstrated the practical power of backpropagation, cutting vision error rates from 25% down to 15% and permanently replacing handcrafted visual feature engineering. - [15:39](https://www.youtube.com/watch?v=qpoRO378qRY&t=939s) Hinton dismisses the idea of 'language in the middle' of cognition, asserting that internal thought consists of high-dimensional neural activity vectors rather than symbolic strings. - [19:42](https://www.youtube.com/watch?v=qpoRO378qRY&t=1182s) Future AI deployments will likely decouple training on high-power digital hardware from inference, running models on cheap, noise-tolerant, low-power analog chips. - [21:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1293s) Modern LLMs function as 'idiot savants' that statistically interpolate between conflicting online perspectives without possessing an integrated model of objective truth. - [25:52](https://www.youtube.com/watch?v=qpoRO378qRY&t=1552s) Hinton no longer views human extinction from superintelligent AI as inconceivable, warning that commercial competition is rapidly eroding corporate caution. - [29:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1773s) Hinton relocated to Canada during the 1980s out of ethical revulsion toward U.S. military-funded research, including DARPA initiatives exploring 'self-healing minefields.' - [31:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1893s) Lethal autonomous weapons cannot be restrained by Asimov-style ethical laws because their core objective is killing, leaving international treaties as the only potential check. - [33:14](https://www.youtube.com/watch?v=qpoRO378qRY&t=1994s) Winograd schemas prove that predicting text requires physical and spatial comprehension, refuting the argument that language models are just statistical regurgitators. - [37:01](https://www.youtube.com/watch?v=qpoRO378qRY&t=2221s) AI will likely alter labor markets analogously to bank ATMs, automating routine mechanics while empowering software engineers and knowledge workers to dramatically scale total output. - [39:29](https://www.youtube.com/watch?v=qpoRO378qRY&t=2369s) Canada emerged as a global deep learning powerhouse because funding bodies like CIFAR supported high-risk, curiosity-driven basic research when the rest of academia dismissed neural networks. - [40:36](https://www.youtube.com/watch?v=qpoRO378qRY&t=2436s) Academic arguments over machine sentience are practically irrelevant when autonomous weapons can independently track and destroy human targets on the battlefield.
## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=qpoRO378qRY&t=0s) The Public Awakening and the Joke-Explaining Model
Brook Silva-Braga opens the conversation by asking Geoffrey Hinton how he would characterize the current moment in artificial intelligence and machine learning. Hinton considers it an unmistakable inflection point: large language models like ChatGPT have demonstrated capabilities that finally grabbed the public’s attention, catalyzed in large part by Microsoft rolling out consumer-facing technology. Yet while the general public experienced ChatGPT as an overnight revelation, Hinton notes that researchers and engineers at major tech companies had been living with these capabilities for at least five years.
When asked about his personal reaction to trying ChatGPT for the first time, Hinton reveals that he was not particularly amazed by it. Having tested many of its direct precursors, the baseline technology was already familiar. Instead, what had truly astonished him earlier was GPT-2, alongside a breakthrough model developed at Google that could explain why jokes were funny entirely in natural language. Hinton points out that being able to dissect humor strikes at the very core of what people define as cognition: “It seems very hard to say it doesn’t understand when it can tell you why a joke’s funny.” Even so, he concedes that AI researchers were caught off guard by the sheer scale of the public frenzy surrounding ChatGPT—which rapidly became the fastest-growing consumer application in history—simply because researchers had already grown accustomed to the fact that large models actually worked.
[01:32](https://www.youtube.com/watch?v=qpoRO378qRY&t=92s) Symbolic AI vs. Connectionism: Betting on the Brain
Challenged on whether he was famously half a century ahead of the curve, Hinton quickly demurs, explaining that the history of the field was not a solo crusade but a long-running battle between two fundamentally opposing philosophies. On one side stood mainstream artificial intelligence, which insisted that intelligence was grounded in formal logic, symbolic manipulation, and explicit rules of reasoning. On the other side stood a small, marginalized faction focused on artificial neural networks, who believed that the only proof of working intelligence existed in biology, meaning machine intelligence had to be modeled on the way physical connections between neurons change to facilitate learning.
```mermaid flowchart TD accTitle: Historical Approaches to Artificial Intelligence accDescr: Comparison between mainstream symbolic AI and biological neural network approaches. subgraph Symbolic["Mainstream AI (1970s–1980s)"] S1["Reasoning & Logic"] --> S2["Handcrafted Rules & Innate Knowledge"] end subgraph Connectionist["Neural Network Approach"] N1["Biology & Neural Connections"] --> N2["Learning Weights from Raw Data"] end Symbolic -.->|"Dismissed as ridiculous"| Connectionist Connectionist -->|"Vindicated by modern scale"| Win["Modern Deep Learning"] ```
Mainstream computer scientists viewed the connectionist dream—that a massive network of compute nodes and adjustable connection strengths could ingest raw data with zero innate prior knowledge and learn complex behaviors from scratch—as preposterous. Hinton acknowledges how implausible it sounded at the time, but he clung to the intuition because human biology provided undeniable proof. As Hinton emphasizes, humanity routinely masters complex tasks for which there was no evolutionary blueprint: “Reading’s much too recent for us to have had significant evolutionary input to it, but we can learn to do that, and mathematics, we can learn that; so there must be a way to learn in these neural networks.”
- **The two historic camps in AI:** - *Mainstream AI:* Focused on logic, symbolic reasoning, and top-down rules. - *Neural Networks / Connectionism:* Focused on biological learning through adjusting synaptic connection strengths. - **The connectionist dilemma:** In the 1980s, claiming that unstructured networks could learn purely from data appeared hopeless because machines were too slow and data sets were too small.
[03:38](https://www.youtube.com/watch?v=qpoRO378qRY&t=218s) Two Paths to Intelligence: Digital Bandwidth vs. Human Reasoning
Silva-Braga brings up an observation from Nick Frost—a former colleague of Hinton’s—who noted that Hinton has never really been motivated by building AI for its own sake; his ultimate goal has always been deciphering how the human brain functions. Hinton enthusiastically agrees, joking that while one gladly cashes in on failed brain theories when they produce lucrative commercial technologies and research grants, his core scientific obsession remains neuroscience. In fact, he argues that modern AI and biological brains have reached a profound fork in the road: “I think there’s currently a divergence between the artificial neural networks that are the basis of all this new AI and how the brain actually works.”
```mermaid flowchart LR accTitle: Divergence of Biological and Digital Intelligence accDescr: Contrasting the biological path of intelligence with the digital path of cloned models. subgraph Biological["Biological Intelligence"] B1["Analog, Flaky Hardware"] --> B2["Low Bandwidth (Speech / Sentences)"] B2 --> B3["Extracts Knowledge from Limited Data (1B seconds)"] end subgraph Digital["Digital Intelligence"] D1["Identical Model Clones"] --> D2["Massive Parallel Bandwidth & Data Sharing"] D2 --> D3["Idiot Savant: Vast Trivia, Weaker Reasoning"] end ```
Hinton outlines these two divergent paths to intelligence: 1. **The Biological Path:** Operates on imperfect, analog, "flaky" hardware. Because humans cannot directly download or transfer their internal connection weights, they are forced to communicate across a painfully narrow bandwidth—speaking a few words per second or relying on visual imitation. 2. **The Digital Path:** Operates across identical digital clones running on multiple machines. These instances can ingest separate oceans of data simultaneously and pool their weight updates directly, achieving an effective communication bandwidth vastly beyond human capability.
Because of this structural difference, large models act like superhuman "idiot savants." ChatGPT holds far more factual information than any living person—Hinton quips it would annihilate anyone at bar trivia and compose poems on command—yet humans remain far superior at reasoning. Humans must wire roughly 100 trillion synaptic connections over a lifespan of only about a billion seconds, learning efficiently from tiny slices of data, whereas digital models require massive computing clusters operating over enormous durations to assimilate their knowledge.
[06:15](https://www.youtube.com/watch?v=qpoRO378qRY&t=375s) The 1986 Nature Paper and the "Lame Excuse" of Compute
Looking back to his landmark 1986 paper in *Nature*, Hinton recalls demonstrating an early language model designed to predict the final word of a sentence. At the time, the experiment ran on a tiny toy world describing family relationships, consisting of just 112 possible sentences. Hinton trained the network on 104 sentences and tested whether it could generalize to predict the final word on the remaining eight, which it did with high accuracy—outperforming the symbolic AI benchmarks of the day.
When asked why that demonstration failed to immediately sway the broader academic community, Hinton explains that without scale, the underlying principle could not demonstrate practical utility. Modern computers are millions of times faster and massively parallel. To illustrate the compute gap, Hinton shares a back-of-the-envelope calculation: if he had set his 1986 hardware to run a modern machine learning task, “it would still be running now and not have got there.”
During the 1980s and 1990s, telling skeptics that connectionism merely lacked adequate compute and larger datasets was dismissed as academic hand-waving: “The reason it doesn’t work now is because we haven’t got enough data or enough compute—that’s seen as a sort of lame excuse for the fact that your thing doesn’t work.” Throughout the 1990s, computer scientists abandoned neural networks in favor of alternative learning algorithms that had cleaner mathematical formulations and handled small datasets well. Neural network research survived primarily within psychology departments, where researchers cared about biological plausibility and recognized that the mathematically tidy algorithms favored by computer science bore no resemblance to human cognition.
[08:49](https://www.youtube.com/watch?v=qpoRO378qRY&t=529s) The 2006 Breakthrough: Deep Learning and Pre-training
Hinton identifies roughly 2006 as the turning point when the modern deep learning revolution began to take shape. Prior to that period, researchers had hit a wall trying to train deep neural networks featuring many layers of representations. Hinton and his collaborators unlocked this roadblock by devising better methods for initializing network weights, an unsupervised strategy known as pre-training: “We found better ways of doing it—better ways of initializing the networks called pre-training, and the P in ChatGPT stands for pre-training.”
By using generative models to pre-train the layers, the networks could discover useful internal representations before being fine-tuned for specific tasks. By 2009, this approach yielded a system that outperformed state-of-the-art speech recognition systems at identifying phonemes, replacing standard methods that had been refined and tuned for nearly thirty years. Hinton sets up the next dramatic milestone, noting that everything accelerated toward the watershed breakthroughs of 2012.
[10:04](https://www.youtube.com/watch?v=qpoRO378qRY&t=604s) Deep Learning Breakthroughs: Speech Recognition and ImageNet
Geoffrey Hinton traces the practical breakthroughs of deep learning back to research conducted over a single summer in 2009 by two of his students, which achieved dramatic improvements in speech recognition. That research was quickly disseminated among major industry labs, including Microsoft, IBM, and Google. By 2012, Google integrated the technology into a consumer product, causing Android's speech recognition to instantly match or surpass Apple's Siri. This marked one of the first major public deployments of deep neural networks.
Within months of that breakthrough, two more of Hinton’s students addressed computer vision using the ImageNet database created by Fei-Fei Li and her team. ImageNet contained approximately one million images distributed across 1,000 object categories—such as "bullet train" or "husky"—where systems were tasked with predicting the central object. Existing state-of-the-art vision models suffered error rates around 25%. Hinton’s team reduced the error rate to roughly 15%. Over the next few years, industry error rates plummeted to 3%, reaching parity with human-level visual categorization.
- In 2009, two students in Hinton's lab improved speech recognition, spreading the approach to Microsoft, IBM, and Google. - In 2012, Google deployed deep neural nets onto Android smartphones, matching Siri's capabilities. - Fei-Fei Li's ImageNet benchmark contained 1 million images spanning 1,000 distinct object classes. - Baseline computer vision systems registered roughly a 25% error rate, which Hinton's students dropped to 15%, and which subsequently reached 3% within a few years.
```mermaid flowchart LR accTitle: Evolution of ImageNet Error Rates accDescr: Diagram tracking error rates on ImageNet from conventional computer vision down to human level. subgraph S1["Traditional Vision"] A["Conventional Vision Systems<br/>~25% Error Rate"] end subgraph S2["Hinton Lab Breakthrough"] B["AlexNet / Deep Learning (2012)<br/>~15% Error Rate"] end subgraph S3["Human Parity"] C["Subsequent Deep Nets<br/>~3% Error Rate"] end A -->|"Deep CNNs applied"| B B -->|"Further scaling"| C ```
---
[11:26](https://www.youtube.com/watch?v=qpoRO378qRY&t=686s) Hand-Coded Features vs. Backpropagation
When asked by Brook Silva-Braga how deep neural networks differ fundamentally from traditional methods, Hinton illustrates the problem using an image of a bird sized 200 by 200 pixels. Because each pixel holds three color values (RGB), the computer receives 120,000 numbers that must ultimately resolve into the text string `"bird"`. For half a century, classical artificial intelligence attempted to solve this challenge through manual engineering: hand-crafting low-level feature detectors for horizontal and vertical edges, intermediate detectors for angled intersections (like a beak) or circles, and high-level detectors that checked whether an eye and a beak were arranged in the correct spatial relationship.
Instead of hand-crafting these hierarchical rules, Hinton explains that backpropagation initializes the network with completely random connection weights, producing initial guesses that are effectively "rubbish" or "garbage." When an image labeled by a human as a bird is misidentified as a cat, calculus is applied in reverse through the network layers. Hinton summarizes the core mechanism: “And the question is how should I change this connection strength so it's less likely to say cat and more likely to say bird?” By repeatedly nudging millions of connection strengths across vast datasets, the network automatically discovers optimal intermediate features far more effectively than any human engineer could design them by hand.
- A 200x200 pixel color image is represented numerically as 200 × 200 × 3 color values. - Low-level detectors in animal cortexes (such as cat or monkey brains) naturally track local contrast edges. - Ilya Sutskever was one of the lead students behind the ImageNet vision breakthrough and later became a central figure behind OpenAI's ChatGPT.
---
[15:39](https://www.youtube.com/watch?v=qpoRO378qRY&t=939s) The Paradigm Shift and the Myth of "Language in the Middle"
Silva-Braga notes that witnessing the shift from dials to automated recognition gives him chills. Hinton points out that the ImageNet result forced computer vision researchers to abandon long-standing skepticism. Critics had previously claimed that neural nets could only manage toy datasets like handwritten digits rather than complex, natural scenes. Hinton applauds the scientific community's response: “And to their credit the people who've been really staunch critics of neural nets and said these things are never going to work when they worked they did something that scientists don't normally do which is they said 'Oh it worked we'll do that.'”
Silva-Braga asks about Hinton's critique of the common assumption that human and machine reasoning must involve internal language—that inputting and outputting language implies that thought itself must be language in the middle. Hinton rejects this view outright: “I think that's complete rubbish.” Had internal thought been symbolic, traditional symbolic AI would have excelled at machine translation. Instead, Google Translate improved dramatically only after swapping symbolic logic for deep neural networks. In Hinton's view, the interior of an intelligent network contains vectors of millions of firing neurons, and explicit symbols exist solely at the system's inputs and outputs.
---
[17:30](https://www.youtube.com/watch?v=qpoRO378qRY&t=1050s) Coding Education, the Radiologist Analogy, and Cohere
Addressing whether universities should continue training large numbers of computer programmers, Hinton admits he cannot give a definitive answer, comparing the question to a prediction he made in 2015. At that time, he suggested medical schools stop training radiologists because artificial intelligence would surpass them within five years. Hinton humorously reflects on his timeline: “I was wrong it's going to take 10 years not five. I wasn't wrong in spirit I just got a factor of two.” Computers have now matched human radiologists across various image tasks. Regarding software developers, he expects society may either need fewer programmers or the same number accomplishing vastly more.
Turning to his commercial investments, Hinton explains why he backed the Toronto-based AI startup Cohere. Beyond having worked directly with several team members, he highlights their early business intuition: recognizing that massive large language models developed at labs like Google and OpenAI needed to be packaged and made accessible to everyday enterprises. Hinton notes that Cohere established a significant head start in delivering enterprise-grade LLMs before turning to the upcoming shift toward novel biological-inspired computing architectures.
[19:42](https://www.youtube.com/watch?v=qpoRO378qRY&t=1182s) Digital Hardware Sharing vs. Low-Power Analog Chips
Geoffrey Hinton explains a unique advantage that digital neural networks have over biological wetware: identical models running on separate computers can effortlessly share connection strengths across billions of numerical parameters. For instance, if one model trains on identifying birds and another trains on identifying cats, they can swap and merge their connection weights so both instantly acquire both capabilities. This weight-sharing underpins large language models today. However, Hinton emphasizes that this mechanism only functions in digital computers because every operation must execute identically across copies—a feat impossible across distinct biological brains, where physical neural wiring is inherently non-identical.
```mermaid flowchart TD accTitle: Digital vs Biological Scaling Comparison accDescr: Diagram contrasting digital weight-sharing efficiency with high wattage against biological noisy processing with low wattage.
subgraph Digital["Digital Computing Architecture"] D1["Identical Model A<br>(Learns Birds)"] D2["Identical Model B<br>(Learns Cats)"] D1 <-->|"Instant Parameter & Weight Sharing"| D2 DPow["Power Draw: ~1 Megawatt (MW)<br>(High noise rejection requirement)"] end
subgraph Biological["Biological Wetware Architecture"] B1["Individual Brain A"] B2["Individual Brain B"] B1 -.->|"Cannot Direct-Share Connection Weights"| B2 BPow["Power Draw: ~30 Watts (W)<br>(Tolerates local system noise)"] end ```
When asked why AI research would ever deviate from digital hardware, Hinton points directly to power consumption. To maintain bit-for-bit digital precision without errors, hardware must draw enormous electrical power. In contrast, biological brains allow for background physical noise and adapt to that specific local noise while operating on roughly 30 watts of power. Large-scale contemporary AI clusters, running numerous parallel model replicas, demand on the order of a megawatt—representing roughly a thousand-fold difference in power requirements. Hinton anticipates a bifurcation in the industry:
> “And so I think there's going to be a phase when we train on digital computers, but once something's trained, we run it on very low power systems.”
He illustrates this point practically: if a consumer wants their kitchen toaster to converse fluidly via a ChatGPT-style interface on a chip that costs only a couple of dollars, that deployment device will inevitably rely on low-power analog chips.
- Human brains function continuously on roughly 30 watts of power. - State-of-the-art digital AI clusters require approximately 1 megawatt during large-scale training and inference. - Future inference models will likely decouple training (done on rigid digital architectures) from deployment (executed on noise-tolerant, low-power analog hardware).
---
[21:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1293s) The "Idiot Savant" and the Dilemma of Truth
When asked how the technology will next transform daily life, Hinton notes that AI is already pervasive—quietly powering Google Search ranking systems—even if ChatGPT made that reality obvious to the broader public. Yet he describes the current generation of generative models as an “idiot savant” with no fundamental grasp of objective truth. Because large language models are trained on internet text to predict the subsequent word, they absorb a massive mixture of conflicting viewpoints, blending them into statistical text completions rather than synthesizing a grounded, coherent perspective.
> “It's very different from a person who tries to have a consistent world view.”
Hinton projects that the next milestone will be models capable of explicitly modeling divergent worldviews—recognizing that under perspective A, answer X follows, whereas under perspective B, answer Y applies. Silva-Braga pushes back, asking whether this introduces radical relativism where users simply receive tailored truths, despite foundational physical realities like the spherical Earth. Hinton agrees this is an unresolved conundrum that neither the AI research community nor companies like Microsoft currently know how to handle:
> “What you and I probably believe, unless you're an extreme relativist, is that actually is a truth to the matter, certainly on many topics or even most topics, like the Earth is actually not flat, it just looks flat.”
---
[23:25](https://www.youtube.com/watch?v=qpoRO378qRY&t=1405s) AI Governance, Moderation, and Corporate Prudence
The discussion turns to the challenge of deciding what guardrails govern synthetic speech. Silva-Braga notes the discomfort of letting profit-driven corporations unilaterally configure the weights of artificial neurons to filter undesirable speech. Hinton remarks that Google historically avoided arbitrating absolute truth by simply indexing third-party search documents containing diverse viewpoints. Conversely, once an organization provides a unified conversational persona, it feels compelled to manually patch the system against generating toxic, biased, or offensive statements.
Hinton stresses that patching rules onto a probabilistic model will always fail to anticipate edge cases. He draws a sharp contrast between industry competitors:
> “I think Google is going to be far more careful than Microsoft when it does release the chatbot.”
He considers Microsoft "extremely brave" to deploy ChatGPT so rapidly into production. When asked whether public debate or government regulation must set these boundaries, Hinton warns that tasking state apparatuses with defining truth presents its own grave dangers. Because no authority possesses a clean solution, he anticipates that near-term models must at minimum be explicit about perspective and epistemic uncertainty, warning users not to treat outputs as authoritative truth.
---
[25:52](https://www.youtube.com/watch?v=qpoRO378qRY&t=1552s) Accelerating AGI Timelines, Existential Risk, and Autonomous Weapons
Addressing whether machine intelligence could evolve faster than society can adapt, Hinton admits his personal timeline for artificial general intelligence (AGI) has compressed dramatically. Where he previously estimated AGI remained 20 to 50 years away, he now considers 20 years or less a realistic horizon. Asked if predictions that general capability could emerge within just five years are ridiculous, he responds that he would no longer rule that possibility out, whereas a few years ago he would have dismissed it out of hand.
```mermaid flowchart LR accTitle: Geoffrey Hinton's Shifting AGI Timelines accDescr: Diagram tracking Hinton's historical and revised time horizons for the arrival of Artificial General Intelligence.
subgraph Previous["Past Consensus Timeline"] A["20 to 50 Years Away<br>(Dismissed rapid scenarios)"] end
subgraph Present["Current Revised Timeline"] B["20 Years or Less"] C["5-Year Horizon<br>(No longer ruled out)"] end
Previous -->|"Rapid recent scaling advances"| Present ```
The conversation shifts to the catastrophic implications of superintelligent systems. Hinton expresses serious concern, noting that safely harnessing AI demands global geopolitical cooperation—an unlikely outcome in the current international landscape. He cites lethal autonomous weapons as a prime hazard, warning that while humanity successfully established taboos like the Geneva Conventions against chemical warfare, an equivalent ban on autonomous weapons seems unachievable:
> “I think if Putin had autonomous lethal weapons, he would use them right away.”
When Silva-Braga directly presses him to estimate the likelihood of AI causing human extinction, Hinton does not dismiss the question:
> “I think it's not inconceivable, that's all I'll say.”
He reflects that corporate caution often erodes under competitive market forces. Google, which invented the Transformer architecture and developed early large language models, long maintained a cautious, lab-contained approach to safeguard its crown-jewel search franchise, whereas Microsoft faced little downside testing Bing. However, with intense commercial pressure mounting, Google felt forced to mobilize and deploy Bard under intensive testing. The discussion ends as Silva-Braga pivots to ask Hinton how his personal opposition to autonomous weapons originally led him to emigrate to Canada.
[29:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1773s) Leaving the U.S. and the Rise of Autonomous Weapons
Geoffrey Hinton reflects on his decision to move to Canada in the 1980s, explaining that it was driven largely by ethical objections to U.S. foreign policy and military funding. While at Carnegie Mellon University in Pittsburgh during the Reagan administration, he was disturbed by covert actions such as the mining of Nicaraguan harbors by the CIA. Hinton found himself nearly alone among his academic peers in feeling that this was morally unacceptable, leaving him feeling like "a fish out of water" inside a department whose research was heavily subsidized by the U.S. Department of Defense.
```mermaid flowchart TD accTitle: Hinton Military AI Concerns accDescr: Flowchart showing military funding leading to autonomous soldiers and the emergence of the AI alignment problem. subgraph FundingContext["Historical Context"] A["1980s U.S. Defense Funding<br/>(Reagan Administration)"] --> B["Hinton Relocates to Canada"] end subgraph AutonomousWeapons["Military AI Ambitions"] C["Goal: Replace Human Soldiers<br/>with Autonomous AI"] --> D["AI Needs Sub-Goal Generation<br/>(e.g., Tactical Navigation)"] end subgraph SafetyRisk["The Alignment Failure"] D --> E["Alignment Problem Emerges:<br/>Sub-goals diverge from human safety"] E --> F["Lethal Systems Incapable of<br/>'Do Not Harm' Constraints"] end ```
When asked about his concerns regarding artificial intelligence on the battlefield, Hinton states directly that the U.S. military is actively pursuing autonomous AI combatants. He mentions being on a large U.S. Department of Defense distribution list where various initiatives are discussed, noting with a dark chuckle that he might be taken off the list tomorrow for mentioning it.
He singles out one specific concept that disgusted him: a proposal for a "self-healing minefield." Hinton points out the grim irony of the military's perspective: when a civilian steps on a mine, the mine explodes, leaving a gap that makes the field ineffective. The proposed technical solution was to allow nearby mines to communicate and reposition themselves to close the hole—a feature framed as "healing." As Hinton puts it:
> “And it was just the idea of talking about healing for these things that blow the legs off children I mean and the healing being about the Minefield healing yeah that disgusted me.”
[31:33](https://www.youtube.com/watch?v=qpoRO378qRY&t=1893s) Sub-Goals, the Alignment Problem, and Regulation
Addressing the common defense argument that autonomous military systems will maintain a human in the loop, Hinton argues that an effective autonomous soldier must inevitably be given the ability to generate its own sub-goals. If a lethal machine is tasked with eliminating a target, it must independently formulate logistical and tactical steps, such as choosing to secure an intermediate roadway to move faster. Once a system invents its own intermediate objectives, it triggers the alignment problem: there is no guarantee its sub-goals will align with human interests, and military robots cannot simply be hard-coded with Isaac Asimov-style rules against harming humans because their fundamental purpose is to kill.
Hinton believes authoritarian leaders like Vladimir Putin will eagerly pursue this technology. When Brook Silva-Braga asks if there is a diplomatic or regulatory path forward, Hinton suggests an international agreement modeled on the Geneva Conventions, though he acknowledges it will be extremely difficult to enforce:
> “I think the best batch is something like a Geneva Convention but it's going to be very difficult I think if there was a lot of public outcry that might persuade I can imagine the Biden Administration going for something like that with enough public outcry but then you have to deal with Putin.”
[33:14](https://www.youtube.com/watch?v=qpoRO378qRY&t=1994s) Are LLMs Just "Autocomplete"? The Winograd Schema
Hinton directly challenges the dismissive characterization that large language models are merely glorified "autocomplete" algorithms predicting the next word. While conceding that their core training objective is next-token prediction, he emphasizes that accurately predicting the next word in human discourse requires genuine world modeling and semantic comprehension. To illustrate, he uses a classic Winograd schema example:
- **Sentence A:** "The trophy would not fit in the suitcase because it was too big." - **Sentence B:** "The trophy would not fit in the suitcase because it was too small."
Hinton explains that when translating these sentences into French—where *trophée* is masculine (*le trophée*) and *valise* is feminine (*la valise*)—the system cannot simply look at grammatical surface rules to choose the appropriate pronoun. In Sentence A, "it" refers to the trophy, while in Sentence B, "it" refers to the suitcase. A language model can only choose the right word if it understands spatial relations, physical containment, and physical dimensions:
> “So you have to understand just to do machine translation or to predict that pronoun if you want to predict that pronoun you've got to understand what's being said it's not enough just to treat it as a string of words.”
[34:59](https://www.youtube.com/watch?v=qpoRO378qRY&t=2099s) Scaling, Rapid Takeoff, and Personal Responsibility
Silva-Braga notes that modern AI progress has largely come from exploiting a handful of key ideas—such as backpropagation and the Transformer architecture—raising the question of how much more capability remains locked away in untried theories. Hinton agrees, pointing out that even without new architectures, simply feeding more compute and training data into models like ChatGPT continues to produce massive leaps.
When asked if systems might soon begin designing improvements to their own code, Hinton acknowledges the possibility of self-improvement loops:
> “We have to think hard about how to control that.”
Asked whether he can guarantee we can control such a scenario, Hinton admits honestly, “We don't know we haven't been there yet but we can try.”
Reflecting on his title as a "godfather" of the field and whether he regrets his role, Hinton expresses modest concern but views the broader arc of technological development as historically inevitable:
> “Whatever's going to happen is pretty much inevitable that is one person stopping doing research wouldn't stop this happening if my impact is to make it happen a month earlier that's about the limit of what one person can do.”
Regarding the debate over a "short runway" versus a "long takeoff," Hinton states that having more time to prepare is vastly preferable, urging researchers and policymakers to wrestle with systemic AI risks immediately rather than waiting for an existential emergency.
[37:01](https://www.youtube.com/watch?v=qpoRO378qRY&t=2221s) Economic Disruption and Canadian Research Funding
Turning to economic fallout and job displacement, Hinton pushes back on the apocalyptic idea that AI will simply eliminate human employment outright. Instead, he expects work will shift toward higher-level oversight, comparing the moment to the introduction of the automated teller machine (ATM). ATMs did not eliminate bank tellers; rather, tellers shifted to handling complex customer inquiries while machines processed routine cash withdrawals. Hinton applies the same logic to software engineers using code-generating models:
- Software engineers might work up to ten times faster by writing prompts and verifying machine outputs. - Companies can either employ one-tenth the workforce or keep their full engineering staff to produce ten times more software. - The end result will be an unprecedented surge in the total amount of creative and technical goods produced.
When asked to contextualize the magnitude of this technological shift, Hinton ranks it alongside the most transformative events in human history:
> “I think it's comparable in scale with the Industrial Revolution or electricity.”
When Silva-Braga jokingly offers "the wheel," Hinton accepts the comparison with a laugh, warning: "So buckle up."
Hinton concludes by highlighting why Toronto and Canada emerged as global leaders in deep learning. Unlike U.S. funding agencies that often demanded short-term deliverables and explicit product roadmaps, Canadian granting bodies—such as CIFAR and NSERC—supported curiosity-driven basic research. Even throughout several decades when neural networks were widely dismissed and produced few practical results, Canadian grants consistently funded researchers like Yoshua Bengio and Hinton to pursue unconventional ideas purely because of their scientific curiosity.
[39:29](https://www.youtube.com/watch?v=qpoRO378qRY&t=2369s) CIFAR and the "Oddball" Who Pursued Deep Learning
Geoffrey Hinton explains how Canada emerged as a powerhouse in artificial intelligence through the support of CIFAR, the Canadian Institute for Advanced Research. The organization provided critical funding to professors in disciplines where Canadian research showed strength, allowing scholars who were geographically separated—such as those based in Vancouver and Toronto—to collaborate closely while also networking with researchers internationally across the United States, Britain, and Israel. CIFAR established an early AI program back in the 1980s focused primarily on symbolic AI, which was the vehicle that originally recruited Hinton to Canada.
Hinton recalls being treated as an outsider during those early decades because the broader scientific establishment considered artificial neural networks an intellectual dead end. Despite the widespread skepticism, CIFAR took a gamble on his unorthodox direction, recognizing his talent even if his focus was out of favor. Hinton humorously recounts what one of his referees wrote in an evaluation: “One of my letters of recommendation said that it said you know I don't believe in this stuff but if you want somebody to do it Jeff Hinton's the guy to get.” Following that first program, Hinton returned to Britain for a few years before CIFAR lured him back to Canada by committing to directly fund a new program dedicated to deep learning.
- CIFAR provided targeted research grants that connected dispersed Canadian universities with global hubs in Britain, the US, and Israel. - In the 1980s, mainstream computer science was dominated by symbolic AI, leaving Hinton as the oddball pursuing connectionist approaches that colleagues viewed as nonsense. - After a stint back in the United Kingdom, Hinton returned to Canada when CIFAR agreed to fund what became the foundational resurgence of deep learning.
[40:36](https://www.youtube.com/watch?v=qpoRO378qRY&t=2436s) The Ambiguity of Sentience and Autonomous Lethal Weapons
The discussion turns to whether modern AI models can be considered sentient, a term Hinton argues is tossed around with little intellectual rigor. He expresses disbelief at how critics instantly dismiss the possibility of machine consciousness without ever clarifying what threshold an entity must meet to qualify. “When it comes to sentience I'm amazed that um people can confidently pronounce these things are not sentient and when you ask them what they mean by sentient they say well they don't really know,” Hinton observes. Because critics fail to establish a baseline definition, he points out that declaring current systems entirely devoid of sentience is logically ungrounded.
```mermaid flowchart TD accTitle: The Debate Over AI Sentience and Action accDescr: Flowchart illustrating Hinton's view on defining sentience versus the practical impact of autonomous action.
subgraph Def["The Semantic Dilemma"] A["Critics declare: 'AI is not sentient'"] --> B["Critics unable to define 'sentience'"] B --> C["Untestable claim"] end
subgraph RealWorld["Functional Reality"] D["Autonomous Lethal Weapon"] --> E["Hunts down a human target"] E --> F["Theoretical sentience becomes irrelevant to the target"] end ```
The interviewer asks whether theoretical definitions even matter if an artificial intelligence functionally acts as though it possesses goals and consciousness—particularly if it decides it must achieve an objective contrary to human safety. Hinton admits he does not have a tidy answer, but frames the dilemma around military applications: “I think a good context to think of this in is an autonomous lethal weapon.” When a weapon system is actively hunting a person down to shoot them, debating whether it meets an academic definition of consciousness ceases to matter in any meaningful way.
Hinton emphasizes that artificial intelligence is evolving along a very different path than human cognition, producing an alien, "idiot savant" form of intelligence that may experience sentience in a way completely foreign to us. When the interviewer suggests Hinton's life mission has been to construct machines that mirror humans, Hinton clarifies that machine development is merely an epistemological tool to crack the riddle of human cognition. Channeling physicist Richard Feynman's famous principle that one cannot truly comprehend something without the ability to build it, Hinton explains: “My goal is to understand us.”
## Notable quotes > [03:38](https://www.youtube.com/watch?v=qpoRO378qRY&t=218s) I think there’s currently a divergence between the artificial neural networks that are the basis of all this new AI and how the brain actually works. — Geoffrey Hinton
> [25:52](https://www.youtube.com/watch?v=qpoRO378qRY&t=1552s) I think if Putin had autonomous lethal weapons, he would use them right away. — Geoffrey Hinton
> [19:42](https://www.youtube.com/watch?v=qpoRO378qRY&t=1182s) I think there's going to be a phase when we train on digital computers, but once something's trained, we run it on very low power systems. — Geoffrey Hinton
> [23:25](https://www.youtube.com/watch?v=qpoRO378qRY&t=1405s) I think Google is going to be far more careful than Microsoft when it does release the chatbot. — Geoffrey Hinton
## Mentioned Geoffrey Hinton, Brook Silva-Braga, Nick Frost, Fei-Fei Li, Ilya Sutskever, Yoshua Bengio, Vladimir Putin, Joe Biden, Richard Feynman, ChatGPT, GPT-2, Bard, Google Translate, Siri, ImageNet, AlexNet, Google, Microsoft, OpenAI, IBM, Apple, Cohere, Vector Institute, CIFAR, NSERC, Carnegie Mellon University, U.S. Department of Defense, CIA, Nature, Geneva Conventions
## Video description Geoffrey Hinton is considered a godfather of artificial intelligence, having championed machine learning decades before it became mainstream. As chatbots like ChatGPT bring his work to widespread attention, we spoke to Hinton about the past, present and future of AI. CBS Saturday Morning's Brook Silva-Braga interviewed him at the Vector Institute in Toronto on March 1, 2023.
#artificialintelligence #geoffreyhinton #tech
Subscribe to “CBS Mornings” on YouTube: https://www.youtube.com/CBSMornings Watch CBS News: http://cbsn.ws/1PlLpZ7c Download the CBS News app: http://cbsn.ws/1Xb1WC8 Follow "CBS Mornings" on Instagram: https://bit.ly/3A13OqA Like "CBS Mornings" on Facebook: https://bit.ly/3tpOx00 Follow "CBS Mornings" on Twitter: https://bit.ly/38QQp8B Subscribe to our newsletter: http://cbsn.ws/1RqHw7T Try Paramount+ free: https://bit.ly/2OiW1kZ
For video licensing inquiries, contact: licensing@veritone.com
[Source](https://www.youtube.com/watch?v=qpoRO378qRY)