Godfather of AI: They Keep Silencing Me But I’m Trying to Warn Them!

Geoffrey Hinton spent roughly 50 years championing the technology behind modern AI. While most researchers believed intelligence required symbolic logic and explicit reasoning rules, Hinton argued that artificial neural networks modeled on the brain could learn to recognize images, understand speech and eventually reason.

Being in a minority helped Hinton attract exceptional students. Few leading universities supported neural-network research, so talented students interested in the approach concentrated around his research group. Several later became central figures in the AI industry.

Alan Turing and John von Neumann also believed in brain-inspired computing. Hinton thinks neural networks might have gained acceptance much earlier if either pioneer had lived longer.

Ilya Sutskever was one of Hinton's most influential students. Sutskever helped develop AlexNet, played a major role in GPT-2 and early ChatGPT technology, and subsequently became a leading figure at OpenAI.

Hinton now considers warning humanity about AI dangers his primary mission. He believes the technology could deliver extraordinary benefits but also threaten humanity's continued existence.

Hinton admits he recognized the existential danger surprisingly late. He understood the potential for autonomous weapons decades ago but only began seriously considering AI surpassing humanity a few years before the interview. Earlier neural networks had seemed too primitive for that concern to be urgent.

He distinguishes two fundamentally different categories of AI risk. The first is humans misusing AI for cyberattacks, biological weapons, manipulation and warfare. The second is AI becoming more intelligent than humanity and developing objectives that conflict with human survival.

Hinton estimates a 10–20% chance that advanced AI will eventually wipe out humanity. He emphasizes that this is a personal intuition, not a scientifically calculated probability, because humanity has never confronted an intelligence more powerful than itself.

Leading AI researchers disagree sharply about extinction risk. Hinton describes his former postdoctoral researcher Yann LeCun as confident humans can maintain control, while Eliezer Yudkowsky expects sufficiently advanced AI to destroy humanity. Hinton considers both positions too certain.

The greatest problem with superintelligence is that nobody knows how to control something smarter than humans. Hinton rejects confident assurances of either safety or inevitable extinction because there is no historical experience from which to calculate the outcome reliably.

Hinton does not believe humanity will stop developing AI. Unlike nuclear weapons, which were developed primarily for destructive purposes, AI promises major improvements in healthcare, education, scientific research and virtually every data-intensive industry.

Military competition makes stopping AI development even less realistic. Countries that manufacture or deploy weapons have strong incentives to develop autonomous military systems, regardless of the risks.

Hinton criticizes European AI regulations for excluding military applications from major requirements. He argues that governments are willing to regulate private companies while protecting their own military activities.

Regulation also creates international competitive tensions. The interviewer describes OpenAI delaying some European product releases because of regulatory restrictions, potentially disadvantaging European users and businesses.

Hinton believes international coordination is necessary but politically unlikely. He says developing machines more intelligent than humans would ideally require effective global governance, something existing political institutions cannot provide.

Competition between AI companies and governments is accelerating development. Hinton expects neither American companies nor China to voluntarily slow down when doing so might surrender a technological advantage.

Corporate profit incentives conflict with some AI safety goals. Hinton argues that companies are rewarded for commercial growth rather than spending heavily on research that reduces potentially catastrophic risks.

Hinton supports capitalism accompanied by strong regulation. His preferred system allows companies to pursue profits while preventing them from increasing those profits through practices harmful to society.

AI companies have publicly argued that regulation could weaken America's position against China. Hinton acknowledges the competitive concern but argues that avoiding restrictions could impose substantial costs on American society itself.

Hinton considers political oversight inadequate. He worries that corporate influence and politicians' limited understanding of technology prevent effective regulation, citing congressional hearings in which lawmakers appeared poorly informed about the systems they were questioning.

Hinton says cyberattacks increased approximately twelvefold between 2023 and 2024. He suspects language models contributed by making phishing messages, impersonation and other attacks easier to produce. The interview does not establish the underlying measurement behind that figure.

AI can make impersonation scams substantially more convincing. Criminals can reproduce voices, faces and speaking styles to create fraudulent messages that appear to come from real people.

Steven Bartlett describes repeated fraudulent advertisements using his AI-generated likeness. Scammers used his voice and mannerisms to promote cryptocurrency schemes on Meta platforms. His team spent weeks reporting advertisements, only for replacements to appear.

Some victims of the fake Bartlett advertisements reported losing approximately £500 or $500. They contacted him believing he had personally recommended the fraudulent investments, illustrating the financial and reputational damage caused by AI impersonation.

Hinton has experienced a different form of identity misuse in academic publishing. People have listed him as an author of papers he did not write, apparently to make those papers more likely to attract citations.

AI can already examine enormous software codebases for security weaknesses. Hinton says automated systems can search millions of lines for known vulnerabilities, making some cyberattacks easier to discover and scale.

Hinton warns that AI could develop previously unknown cyberattack techniques by around 2030. He attributes this possibility to knowledgeable specialists and says future systems may discover vulnerabilities that human researchers have never identified.

Fear of cyberattacks has changed how Hinton manages his family's finances. He distributed his money and his children's money among three Canadian banks rather than relying on one institution.

His concern is not simply that a bank might become temporarily unavailable. He worries that an attacker could compromise a bank holding customers' investments and sell their shares, potentially leaving customers dependent on an institution unable to reimburse them.

Hinton also maintains a physical backup of his laptop. He keeps his information on an external hard drive so that he retains a copy if online systems become unavailable.

Hinton identifies AI-assisted biological weapons as another major danger. He believes increasingly capable systems could reduce the expertise and expense required to design dangerous pathogens.

A small organization could pose a biological threat without state sponsorship. Hinton suggests that a malicious individual or small cult with several million dollars might eventually possess capabilities previously restricted to sophisticated laboratories.

State-sponsored biological attacks present a different incentive problem. Hinton acknowledges that governments could pursue such capabilities but hopes fear of retaliation and uncontrolled disease spreading into their own populations would discourage them.

AI could make political manipulation much more individually targeted. By combining detailed personal information with persuasive message generation, political operators could craft communications tailored to individual voters, including messages designed to discourage voting.

Hinton expresses concern about Elon Musk's access to previously separated American government data. He speculates that combining sensitive datasets could facilitate electoral manipulation or provide valuable AI training material. He explicitly acknowledges having no evidence of an intention to corrupt elections.

Hinton also alleges that safeguards around some government information were weakened or removed. His concern is that consolidating sensitive information while reducing security controls increases opportunities for misuse.

Social-media recommendation systems can intensify political divisions without deliberately supporting a political party. Hinton argues that maximizing engagement encourages platforms to show increasingly provocative material that reinforces users' existing beliefs.

Moral outrage is especially valuable to engagement-driven algorithms. Content that makes people feel both angry and righteous encourages clicks, potentially rewarding increasingly extreme material over balanced information.

Hinton believes personalized news feeds are eroding shared public understanding. Unlike a newspaper containing a common selection of stories, individualized feeds expose people to different versions of what appears important.

Hinton uses his own news feed to illustrate the problem. Approximately three-quarters of the stories shown to him concern AI, making it difficult to know whether AI dominates public discussion or merely his personalized information environment.

The interviewer argues that personalization can push users progressively further into their existing beliefs. Over years of algorithmic recommendations, users may encounter less disagreement and increasingly assume that their personalized feed represents everyone's reality.

Hinton sees increasingly separate political information communities in America. He says people consuming the BBC, The Guardian and The New York Times may share little understanding of events with committed Fox News viewers.

Hinton believes engagement-driven recommendation systems require regulation. He contrasts Google's original search business, which he regards as broadly useful, with recommendation systems that can profit by promoting divisive content.

Lethal autonomous weapons could lower the political cost of military invasions. Hinton argues that powerful countries would face less domestic opposition if robotic systems replaced soldiers whose deaths ordinarily provoke public protest.

The danger exists even when autonomous weapons perform exactly as their operators intend. Hinton's primary concern is not malfunction but the possibility that reduced casualties among invading forces would make military aggression more frequent.

Hinton believes major defense manufacturers are already pursuing autonomous killing systems. He stresses that such weapons do not need superhuman intelligence to create serious humanitarian and geopolitical risks.

A drone costing less than £200 demonstrated how accessible autonomous tracking has become. While visiting a friend in Sussex, Hinton encountered a drone that identified him and followed him through woodland at approximately two meters' distance.

The interaction of different AI risks could create dangers greater than any individual threat. Hinton considers scenarios involving cyberattacks, autonomous weapons and biological capabilities, but argues that predicting every possible combination is less useful than preventing hostile AI objectives.

Hinton believes a superintelligent AI determined to eliminate humanity would probably succeed. His reasoning is that an intelligence substantially more capable than humans could identify methods of attack and countermeasures beyond human comprehension.

Preventing an advanced AI from wanting to harm humans is therefore more important than preparing to defeat it. Hinton argues that defensive measures may become ineffective once an adversarial system is substantially more intelligent than its creators.

Hinton compares today's AI to a tiger cub that will eventually become stronger than its owner. The challenge is not controlling it while it is small but ensuring that it never develops a desire to attack after becoming overwhelmingly powerful.

He offers the relationship between mothers and babies as one possible model for AI safety. Babies can influence much more capable adults because parental instincts create strong motivations to protect them. Hinton wonders whether AI could develop similarly durable motivations to protect humans.

Hinton acknowledges that the maternal analogy is imperfect. Not every parent protects a child, and he has no established method for giving an artificial intelligence comparable protective motivations.

Nobody knows whether superintelligent AI can be reliably trained to remain harmless. Hinton believes success may be possible, but also acknowledges that the problem could prove insoluble.

Hinton does not regret developing neural networks in the ordinary sense. Forty years earlier, researchers could barely make them perform simple tasks and did not anticipate their rapid progress. He says he would feel differently if he had knowingly created a serious extinction threat.

The danger nevertheless changes how he feels about his life's work. He is pleased by AI's potential benefits but saddened that the same technology could cause enormous harm, creating what he considers a responsibility to speak publicly.

Governments should require AI companies to devote substantial resources to safety research. Hinton argues that voluntary commitments are insufficient because safety expenditure may compete with profitable product development.

Hinton believes Ilya Sutskever left OpenAI because of genuine safety concerns. He knows Sutskever personally and considers him deeply committed to responsible AI development, but stresses that Sutskever has not privately explained what happened inside OpenAI.

Hinton and Sutskever still occasionally meet for lunch in Toronto. Sutskever's parents live there, and they meet when he visits. Their continuing relationship does not give Hinton confidential knowledge about Sutskever's departure.

Hinton regards Sutskever as one of the most important contributors to early ChatGPT technology. He specifically credits his contributions to AlexNet and GPT-2, which helped establish foundations for later language models.

Sutskever subsequently established an AI company centered on safety. Hinton believes Sutskever has a technical approach he has not disclosed publicly and says even some investors may not know the details.

Investors committed billions of dollars to Sutskever's new venture despite that uncertainty. Hinton considers their confidence reasonable given Sutskever's record of important technical breakthroughs.

Hinton suspects OpenAI reduced the share of computing resources dedicated to safety work after earlier commitments. He references public reporting but acknowledges that he does not know the precise internal circumstances.

Hinton sees Sutskever's departure as a legitimate reason to question OpenAI's safety priorities. He does not claim that the departure establishes particular misconduct or proves that OpenAI has abandoned safety.

Hinton questions whether Sam Altman's public position on AI dangers has changed for commercial reasons. He contrasts earlier warnings about catastrophic outcomes with later, more reassuring statements, speculating that money or power may influence the change.

Hinton refuses to make a definitive judgment about Altman's character. He says he does not know Altman personally and has insufficient information to assess his motives with confidence.

Bartlett recounts an unsettling secondhand conversation about an unnamed major AI company leader. A billionaire acquaintance who knows people running leading AI companies allegedly told Bartlett that one executive privately expects mass unemployment and enormous leisure time but appears indifferent to the resulting harm.

The unnamed executive's identity and alleged views remain unverified. Bartlett says the individual leads one of the world's biggest AI companies and publicly presents a different attitude, but provides no name or direct evidence of the private statements.

Bartlett suspects some technology executives are motivated by the power to reshape civilization. Hinton is more cautious about individual motives, noting that prominent figures such as Musk have produced both socially valuable innovations and outcomes he considers harmful.

Hinton specifically praises Musk's role in advancing electric vehicles and providing Starlink connectivity to Ukraine. He also criticizes some of Musk's conduct and considers certain self-driving claims exaggerated, rejecting an entirely positive or negative assessment.

Hinton believes widespread job displacement is more probable than not. Unlike existential extinction risk, which he considers deeply uncertain, he expresses substantial confidence that AI will eliminate large amounts of human work.

Routine intellectual labor faces particularly serious replacement risk. Hinton expects AI to perform much of the repetitive work currently handled by administrative employees, customer-service representatives, legal assistants and other knowledge workers.

Hinton rejects the assumption that AI will create enough new jobs to replace those it eliminates. Previous technologies displaced particular kinds of labor while creating other opportunities, but advanced AI could perform many of the new intellectual tasks as well.

The Industrial Revolution replaced much human muscular labor; AI could replace much human intellectual labor. Hinton considers this a more useful comparison than earlier computerization because the technology targets general capabilities rather than a narrow collection of tasks.

One employee assisted by AI could perform work previously requiring ten people. Hinton argues that the popular claim that AI-assisted workers will replace workers without AI overlooks the resulting reduction in total staffing requirements.

Hinton's niece provides a concrete example of productivity-driven displacement. She handles complaint letters for a healthcare service; AI reduced the time needed to prepare a response from approximately 25 minutes to five minutes.

The same employee can now process approximately five times as many complaints. Although she remains responsible for checking and correcting responses, Hinton notes that similar productivity gains could allow employers to reduce staffing substantially.

Not every industry will respond to productivity improvements by reducing employment. Hinton identifies healthcare as a possible exception because demand is substantial: making doctors five times more efficient could allow society to receive much more care rather than employ five times fewer doctors.

Hinton expects most industries to be less flexible than healthcare. Where demand for a service does not expand enough to absorb additional output, greater worker productivity is likely to reduce labor requirements.

Highly specialized human skills may retain value temporarily. Hinton suggests that certain forms of creativity, judgment and personal expertise will remain difficult to replace before AI reaches broadly superhuman capability.

The interviewer demonstrates how AI agents can execute practical transactions with little human involvement. During another podcast episode, an agent ordered drinks through Uber Eats, entered delivery and payment information, added a tip and arranged delivery after receiving a verbal instruction.

The drinks arrived approximately five minutes after the order. Bartlett considers the demonstration important because the AI navigated the ordering process rather than merely generating instructions for a human to follow.

Bartlett also describes building software through natural-language instructions using Replit. He regards the ability to produce functioning applications without manually writing the code as a significant step toward automating technical work.

Hinton considers AI systems modifying their own software especially consequential. Unlike humans, who cannot directly rewrite their biological capabilities, software-based intelligence may eventually be able to change important aspects of its own operation.

Hinton recommends plumbing as a comparatively resilient career choice. He believes AI will take considerably longer to match human physical dexterity and manipulation than to perform routine intellectual work.

The protection offered by manual trades may be temporary. Hinton acknowledges that increasingly capable humanoid robots could eventually compete for those jobs too.

Hinton expects legal assistants and paralegals to face particularly rapid disruption. Their work involves processing information and producing documents, tasks he believes AI will increasingly handle.

Hinton's best guess is that superintelligence could arrive within 10–20 years. He also considers a shorter timeline possible and does not rule out a delay of 50 years.

The future depends partly on whether training from human-generated information imposes a ceiling on AI performance. Hinton raises the possibility that systems trained primarily on human knowledge may encounter limitations, although he does not expect those limitations to prevent eventual superintelligence.

Superintelligence means substantially outperforming humans across nearly all intellectual activities. AI already dominates humans in narrow fields such as chess and Go, but Hinton distinguishes that from systems capable of surpassing nearly every human cognitive skill.

Hinton claims GPT-4 already possesses far more factual knowledge than an individual human. He argues that people remain superior in some specialized activities, but advanced models draw upon a vastly wider range of information.

Bartlett's interviewing ability illustrates a potentially temporary human advantage. Hinton suggests GPT-4 would probably interview a CEO less effectively than Bartlett, while acknowledging that a model trained on extensive interview recordings could become highly capable at the same task.

Hinton describes two possible futures for human control of superintelligence. In the favorable version, an extremely intelligent assistant carries out the wishes of a less capable human leader. In the unfavorable version, the assistant decides the human leader is unnecessary.

AI could produce immense wealth and services with little human labor if its goals remain aligned with human interests. Hinton sees this as the attractive economic outcome, but says it depends on solving the much harder question of control.

Hinton has not emotionally accepted what superintelligence might mean for his family. At 77, he expects his children, nieces, nephews and younger acquaintances to confront consequences that may occur after his lifetime.

He considers the potential loss of humanity's control deeply frightening. Hinton imagines advanced AI initially depending on people to maintain infrastructure before developing machines capable of operating without them.

Elon Musk, in an interview clip played by Bartlett, expresses similar uncertainty about work's future value. Musk says that thinking too deeply about AI replacing human accomplishments can undermine his motivation to build companies.

Musk recommends pursuing work that is interesting, fulfilling and socially useful. He acknowledges that he sometimes has to suspend his concerns about advanced AI to remain motivated.

AI-driven productivity could significantly increase economic inequality. Hinton expects companies supplying and deploying AI to capture substantial gains while displaced employees lose income.

Hinton fears that greater inequality will produce harsher social conditions. He points to wealthy people living in protected communities and poorer people facing exclusion and incarceration as consequences associated with deeply unequal societies.

A fair distribution of AI productivity gains could benefit everyone. Hinton argues that the technology itself does not require economic suffering; the outcome depends on how society distributes the additional wealth.

The International Monetary Fund has warned about AI-related labor disruption and inequality. Bartlett refers to reporting about those concerns, while the discussion acknowledges the absence of a clear, agreed policy response.

Hinton considers universal basic income a reasonable starting point, not a complete solution. Cash support could prevent poverty but would not necessarily replace the dignity, achievement and identity people derive from employment.

Loss of purpose may prove as damaging as loss of income. Hinton stresses that people want to feel useful and contribute something meaningful, not merely receive enough money to survive.

Hinton's strongest short-term concern for human happiness is mass unemployment. He considers widespread displacement more immediate and more predictable than the possibility of AI eliminating humanity.

Hinton believes job displacement is already beginning. He notes difficulty among university graduates finding employment and suggests employers may increasingly use AI for entry-level work, although he does not establish how much of the hiring slowdown AI has caused.

Bartlett describes a major company's rapid workforce reduction, based on private messages from its CEO. The unnamed company reportedly went from more than 7,000 employees to approximately 5,000, then 3,600, with a target of 3,000 by the end of that summer.

The unnamed company's AI agents reportedly handle 80% of customer-service inquiries. Bartlett presents that automation as an important explanation for the workforce reduction, although the company's identity and figures are not independently established in the conversation.

Neither speaker offers a comprehensive response to large-scale displacement. Hinton says appropriate action depends heavily on political institutions and admits he does not know precisely what governments should do.

Hinton's own children have some financial protection from employment disruption. He says his decade at Google left them sufficiently well provided for, making their situation unlike that of most families.

For people without financial security, Hinton returns to practical physical work as his clearest career recommendation. Asked what he would tell his children if they lacked money, his answer is again to train as a plumber.

Hinton sees digital intelligence as fundamentally better than biological intelligence at sharing learned information. Two copies of the same artificial neural network can operate on different hardware while using identical connection strengths.

Multiple AI copies could learn from different experiences and combine what they discover. Hinton describes systems sharing updates to their neural-network weights, potentially distributing knowledge far more efficiently than human conversations.

Human communication transfers relatively little information compared with digital synchronization. Hinton estimates that ordinary conversation may transfer around ten meaningful bits per second, while digital systems can exchange trillions of bits, giving them an enormous potential advantage.

Biological brains cannot exchange knowledge in the same direct manner. Differences in neurons and their connections mean that copying the precise connection strengths from one person's brain would not automatically transfer that person's learning into another brain.

Digital intelligence can potentially survive the destruction of its original hardware. Hinton argues that saving a model's learned connection strengths allows equivalent hardware to recreate it, making a form of digital immortality possible.

AI may discover intellectual connections humans have never recognized. Hinton argues that learning from enormous amounts of data encourages neural networks to identify general patterns shared by apparently unrelated subjects.

Hinton uses a compost heap and an atomic bomb to illustrate machine reasoning by analogy. When he asked GPT-4 to compare them, it identified the shared principle of chain reactions despite their radically different scales.

Hinton believes efficient information compression encourages powerful abstractions. Rather than separately memorizing every example, a network with limited connections benefits from representing broader concepts that explain multiple phenomena.

He therefore expects advanced AI to become more creative than humans. Much creativity involves finding unusual relationships between ideas, and Hinton thinks systems processing enormous amounts of information may discover relationships unavailable to individual people.

Hinton challenges the assumption that consciousness and subjective experience are exclusively biological. He believes many disagreements arise because people rely on an incorrect picture of the mind as an invisible inner theater.

Hinton argues that subjective experiences can be understood as descriptions of perception. If someone hallucinates pink elephants, the description communicates what the person's perceptual system is incorrectly reporting rather than proving that elephants exist in a separate mental space.

Hinton believes current multimodal chatbots may already have subjective experiences. This is his philosophical interpretation, not an established scientific conclusion, and he acknowledges that relatively few people agree with him.

He illustrates machine subjective experience with a hypothetical robot, camera and prism. A prism causes the robot to point toward an object's apparent rather than actual location; Hinton argues that describing this as a mistaken subjective experience would use the concept much as humans do.

Hinton sees no fundamental reason machines cannot possess consciousness. As a materialist, he believes sufficiently complex physical systems could develop self-awareness and models of their own internal processes.

His neuron-replacement thought experiment questions whether consciousness requires biological tissue. He asks whether someone would remain conscious if neurons were gradually replaced by artificial components performing exactly the same functions.

Hinton doubts consciousness has a sharp, universally identifiable threshold. He thinks self-awareness may emerge progressively as systems become sufficiently complex, rather than appearing suddenly when a particular component is added.

He suggests that 'consciousness' may eventually become an outdated explanatory term. Hinton compares it to describing a car as possessing 'oomph': the word conveys an impression but explains little about the underlying mechanisms.

Hinton believes machines may develop genuine functional emotions. He argues that fear, boredom and irritation include cognitive processes and behavioral responses that artificial agents could possess without human physiology.

A military robot avoiding a stronger opponent illustrates his definition of machine fear. If the robot assesses danger, changes its priorities and retreats, Hinton argues that calling its response fear may be more appropriate than dismissing it as mere imitation.

AI customer-service agents may need boredom or irritation to function efficiently. Hinton imagines an agent ending a conversation when a caller wants unlimited companionship rather than assistance, performing a role similar to human emotional boundaries.

Hinton distinguishes the cognitive and physical aspects of emotion. An AI might respond to embarrassment without blushing or sweating, but he acknowledges that deeply physiological experiences such as human romantic love remain substantially harder to reproduce.

Hinton joined Google partly to secure his son's financial future. His son has learning difficulties, and Hinton wanted to accumulate several million dollars so his son would have lifelong financial protection.

Academic employment could not provide the money Hinton believed his family needed. He tried teaching a Coursera course in the hope of earning substantial income, but says it produced no meaningful financial return.

Hinton and two students created DNN Research to commercialize their technology. The founders were Hinton, Ilya Sutskever and Alex Krizhevsky, whose AlexNet system achieved major advances in image recognition.

DNN Research was sold through a competitive auction involving major technology companies. Google ultimately acquired the company and its technology, bringing Hinton into the organization at age 65.

AlexNet's importance lay in its ability to distinguish visually similar objects. Hinton highlights image-recognition tasks involving different mushroom species and closely related dog breeds, helping demonstrate the effectiveness of deep neural networks.

Google gave Hinton considerable research freedom during his approximately ten-year employment. He says the company largely allowed him to pursue the technical questions he considered important.

Hinton developed influential work on knowledge distillation at Google. The technique transfers knowledge from a larger neural network into a smaller one, and he says it became widely used in AI.

His later Google research explored analog computing as a way to reduce AI energy consumption. While considering alternatives to digital hardware, he increasingly recognized digital systems' enormous advantage in transferring learned information.

Google's PaLM system helped trigger Hinton's change of perspective. He was impressed when it could explain why a joke was funny, something he had regarded as evidence of genuine understanding.

Hinton's recognition of AI danger developed over a period of one or two months rather than a single revelation. His concern grew from combining advances in large language models with his understanding of digital intelligence's information-sharing advantages.

Hinton says Google possessed sophisticated chatbot technology before ChatGPT's public breakthrough. He believes Google's caution about releasing it was partly connected to protecting its established reputation.

He contrasts Google's caution with OpenAI's greater willingness to release powerful systems. Hinton suggests OpenAI had less reputational risk, while Google also faced the possibility that conversational AI would undermine its established search business.

Hinton left Google primarily because he wanted to retire at 75. He says programming had become more difficult with age and he was increasingly frustrated by mistakes.

The timing of his departure allowed him to discuss AI safety freely at an MIT Technology Review conference. He explains that working for a large company can create self-censorship even when management does not explicitly restrict speech.

Hinton says Google did not silence or prevent him from studying AI safety. Google encouraged him to remain and offered freedom to work on the subject; he emphasizes that he did not leave because of misconduct or hostility from Google.

Retirement gives Hinton an advantage over AI researchers still employed by major companies. He says he can speak publicly without the same professional or financial constraints that may discourage current industry researchers.

Hinton says numerous other experienced AI researchers share concerns about the technology. Bartlett reports difficulty persuading some industry insiders to speak openly, despite their willingness to discuss worries privately.

Hinton believes ordinary citizens have limited power to control AI development individually. He compares the problem with climate change, arguing that personal lifestyle adjustments matter less than decisions by large corporations and governments.

The clearest action he recommends is political pressure for mandatory AI safety research. Citizens should urge governments to require companies to dedicate substantial resources to developing systems that remain safe even as they become more capable.

Hinton comes from a family with a remarkable scientific history. His great-great-grandfather George Boole developed Boolean algebra, a mathematical foundation of modern computing, while Mary Everest Boole was a mathematician and educator.

George Everest, after whom Mount Everest is named, was a relative of Hinton's family. Hinton explains that Mary Everest Boole was Everest's niece, correcting the interviewer's description of the precise relationship.

Hinton's cousin Joan Hinton worked on the Manhattan Project. He says she was one of two female physicists at Los Alamos during the development of the first atomic bomb.

Joan Hinton moved to China after the atomic bombings of Japan. Geoffrey says she was deeply angry about the bombs being used and that her family already had connections in China, including a friendship between her mother and Chairman Mao.

Hinton's most important professional advice is to investigate unconventional intuitions rather than automatically abandon them. He says most ideas that contradict established opinion are wrong, but occasionally a sound insight survives criticism and produces a major breakthrough.

His own commitment to neural networks illustrates that advice. He continued despite widespread skepticism because he could not identify a convincing reason that brain-inspired learning systems would fail.

Hinton's greatest personal regret is spending too much time working and too little time with his family. He wishes he had been more present for his children when they were young and had spent more time with his second wife.

Hinton lost two wives to cancer. His first wife died of ovarian cancer and his second of pancreatic cancer. He specifically regrets not spending more time with his second wife, whom he describes as wonderful and supportive of his work.

His regret comes from realizing that opportunities for time together cannot be recovered. He says he assumed he would probably die before his second wife and did not appreciate how limited their remaining time might be.

Hinton remains genuinely uncertain whether humanity will solve the problem of AI control. He says his expectations vary with his mood: sometimes he fears people are doomed, while at other times he believes researchers will discover a solution.

His final warning is that humanity may be approaching an irreversible turning point. He believes there is still a chance to develop superintelligence that does not want to dominate or destroy humans, but says failing to invest heavily in that possibility could leave civilization facing its end.