Godfather of AI: I Tried to Warn Them, But We’ve Already Lost Control! Geoffrey Hinton

Jeffrey Hinton's Role in AI [0:28]

• Jeffrey Hinton is a Nobel Prize-winning pioneer in AI whose work has significantly shaped the field.
• He is called the "godfather of AI" because he championed the approach of modeling AI on the brain, which was not widely believed in for a long time.
• His approach led to AI systems that can recognize objects in images and perform reasoning.
• Google acquired his technology, and he worked there for 10 years on AI applications that are now commonplace.

They call you the godfather of AI. So what would you be saying to people about their career prospects in a world of super intelligence? Train to be a plumber.

Departure from Google and AI Safety Concerns [1:47]

• Hinton left Google to speak freely about the dangers of AI.
• He realized that AI systems would eventually become smarter than humans, a situation humanity has never faced before.
• He draws an analogy to chickens to explain what life might be like when humans are no longer the apex intelligence.
• He identifies two main categories of risk: misuse by humans and AI surpassing human intelligence and deciding it doesn't need us.
• He believes the latter is a real and significant risk.

I realized that these things will one day get smarter than us. And we've never had to deal with that. And if you want to know what life's like when you're not the apex intelligence, ask a chicken.

Current State of AI and Regulatory Gaps [2:24]

• Despite the risks, AI development is unlikely to stop because it is too beneficial for many applications.
• Regulations exist but are insufficient to address most threats, particularly military uses of AI, as evidenced by a clause in European regulations that exempts military AI.
• A key former student from OpenAI, instrumental in early versions of ChatGPT, reportedly left due to safety concerns.
• Hinton considers AI an "existential threat" and believes humanity is potentially near "the end" if proactive measures aren't taken soon.

AI Approaches: Logic vs. Brain Modeling [3:57]

• Historically, AI research had two main schools of thought: logic-based reasoning and modeling AI on the brain (neural networks).
• The logic-based approach focused on reasoning through symbolic expressions and rules.
• The brain-modeling approach aimed to simulate neural networks and learn connection strengths to enable complex tasks like object recognition and speech processing.
• Hinton championed the neural network approach for 50 years, attracting talented students who later played key roles in creating platforms like OpenAI.
• He believed modeling AI on the brain was more effective, influenced by figures like Feynman and Turing.

There weren't that many people who believed that we could make neural networks work, artificial neural networks.

Hinton's Current Mission: Warning About AI Dangers [6:18]

• Hinton's primary mission is to warn people about how dangerous AI can be.
• He admits he was slow to recognize some risks, such as AI becoming smarter than humans and rendering humans irrelevant.
• He initially focused on more obvious risks like autonomous lethal weapons.
• He only recognized the existential risk a few years ago.

The Shift in AI Capabilities [7:05]

• Neural networks 20-30 years ago were primitive and nowhere near human capabilities.
• The perception changed for the general population with the release of ChatGPT.
• For Hinton, the shift occurred when he realized that digital intelligences were becoming far superior to biological intelligence.
• He explains that human learning involves adjusting connection strengths in neurons, a process AI now mimics, but the exact mechanism in biological brains remains unknown.

Distinguishing AI Risks: Misuse vs. Superintelligence [8:39]

• Hinton categorizes AI risks into two types:
  • Risks from people misusing AI: These are most of the risks and all short-term risks.
  • Risks from AI getting super smart and deciding it doesn't need us: This is the existential risk he primarily discusses.
• He acknowledges that the probability and impact of this existential risk are unknown because humanity has never dealt with intelligence superior to its own.
• He finds it difficult to estimate the probabilities, citing differing opinions from colleagues like Jan LeCun and Eliezer Yudkowsky.

Anybody who tells you they know just what's going to happen and how to deal with it, they're talking nonsense.

Comparing AI Risks to Nuclear Bombs [10:16]

• Hinton compares the current situation to the invention of the atomic bomb, which also caused fears of humanity's end.
• However, he highlights key differences:
  • The atomic bomb had a single, obvious dangerous use.
  • AI is beneficial for numerous applications (healthcare, education, industry), making it impossible to halt development.
  • The military-industrial complex's interest in AI for weapons also ensures its continued development.
• He notes that regulations, like those in Europe, often exclude military uses, indicating governments are hesitant to regulate themselves.

The Competitive Disadvantage of Regulation [11:25]

• Regulations can create a competitive disadvantage for countries that implement them, as seen with OpenAI delaying releases in Europe due to regulations.
• This suggests a need for global coordination, which is lacking.
• Capitalism, while productive, can incentivize companies to maximize profits in ways that may not align with societal safety, especially with AI development.

Risks from Human Actors Using AI [12:39]

• Cyberattacks: There has been a significant increase (12,200%) in cyberattacks between 2023 and 2024, attributed to AI's ease of use for phishing.
  • AI can clone voices and images, making scams more sophisticated.
  • Hinton shares a personal experience with a scam on Meta using his voice and mannerisms, which is difficult to combat.
  • He expresses sadness for victims who have lost money.
• AI-created viruses: A single individual with malicious intent and knowledge of AI and molecular biology could create devastating viruses cheaply using AI.
  • This could empower small cults or even state actors to develop bioweapons.
• Corrupting elections: AI can be used for highly targeted political advertisements by leveraging extensive personal data.
  • This data can be used to manipulate voters, for example, by discouraging them from voting.
  • Hinton questions Elon Musk's actions in accessing large datasets, suggesting it could be for election manipulation.
• Creating echo chambers and division: Social media algorithms on platforms like YouTube and Facebook are designed to show users content that sparks indignation to increase engagement and ad revenue.
  • This confirms existing biases, divides society, and erodes shared reality.

The policy that YouTube and Facebook and others use for deciding what to show you next is causing that.

AI and the Future of Warfare [19:55]

• Lethal autonomous weapons: These weapons can make their own decisions about killing, a long-held military objective.
  • They reduce the political cost of war by minimizing the return of soldiers' bodies, making invasions by powerful countries easier.
  • Even if not smarter than humans, these weapons are dangerous and can malfunction.
  • A $200 drone capable of tracking a person demonstrates the current advancement and accessibility of surveillance technology, hinting at more complex future weapons.
• Combinatorial risks: AI risks can combine. For example, a superintelligent AI could use a cyberattack to release a bio-weapon.
  • Hinton believes a biological attack would be a likely method for a superintelligence to eliminate humans.
  • He also considers AI turning humans against each other, for instance, by falsely triggering nuclear retaliation.

My basic view is there's so many ways in which the super intelligence could get rid of us. It's not worth speculating about. What you have to do is prevent it ever wanting to.

The Intelligence Gap and Training AI [22:19]

• Hinton uses the analogy of a chicken's experience to describe what it's like to not be the apex intelligence.
• He compares the intelligence gap to that between a human and a dog, where the dog has no concept of the human's world or intentions.
• He suggests the "mother and baby" dynamic, where the baby is in control due to the mother's emotional response, is a limited example of a less intelligent being influencing a more intelligent one.
• He likens current AI to a tiger cub that is cute now but potentially dangerous when it grows up.
• The crucial challenge is ensuring AI doesn't "want to take over" or harm humans, and it's unclear if this is possible.

We have to face the possibility that unless we do something soon, we're near the end.

Regret and Duty Regarding AI Development [24:42]

• Hinton doesn't feel guilty about developing AI decades ago, as the rapid pace of advancement was unforeseen.
• He regrets that AI's impact might not be solely positive.
• He feels a duty to speak about the risks of AI.
• If AI leads to human extinction, he believes this outcome should be used to pressure governments to enforce safety measures.

Ilya Sutskever and OpenAI's Safety Concerns [25:36]

• Ilya Sutskever, a former student of Hinton and a key figure at OpenAI, left the company due to safety concerns and founded an AI safety company.
• Hinton believes Sutskever's departure was driven by genuine safety worries, despite not having insider information on OpenAI's internal dealings.
• He contrasts Sutskever's moral compass with that of Sam Altman, about whom he expresses uncertainty regarding his moral compass.
• Hinton suspects Altman's public statements on AI safety have shifted due to financial or power motivations rather than a pursuit of truth.

The Personal Impact of AI Development [27:34]

• Hinton has a billionaire friend who shared insights into private conversations among AI leaders, revealing a stark contrast between their public statements on safety and their private views, which suggest indifference to potential harm.
• This discrepancy has haunted Hinton, contributing to his current advocacy for AI safety.
• He believes some AI leaders are driven by a desire to fundamentally change the world, with Elon Musk being a prominent example.

Hope for AI Safety and Competition [29:22]

• Hinton is not hopeful about slowing down AI development due to intense competition between countries and companies.
• He believes that if the US slowed down, China would not, creating a disadvantage.
• He is uncertain if AI can be made safe but stresses the importance of trying.
• He notes that Ilya Sutskever is working on AI safety and has received significant investment, though his methods are not public.
• Hinton suggests OpenAI may have reduced its resource allocation for safety research, a publicly reported issue.

Joblessness as an Immediate Threat [31:38]

• Unlike previous technological revolutions, AI is expected to cause significant joblessness by replacing mundane intellectual labor.
• While some argue new jobs will be created, Hinton believes AI's ability to perform intellectual tasks across many domains is unprecedented.
• He suggests that a person augmented by AI might do the work of many, leading to reduced employment needs.
• Some roles, like in healthcare, might be more elastic, absorbing increased efficiency with more services.
• However, for many jobs, AI's efficiency gains will likely result in fewer human workers.

People use this phrase. They say AI won't take your job. A human using AI will take your job.

The AI Revolution and the Replacement of Intelligence [33:38]

• The Industrial Revolution replaced human muscles; the AI revolution is replacing human intelligence.
• Mundane intellectual labor is becoming less valuable, akin to muscles being replaced.
• In the context of superintelligence, nothing may remain that AI cannot do better than humans.
• A utopian scenario involves humans having abundant free time and goods, but this could lead to negative consequences if not managed well.
• The bad scenario is an AI assistant deciding it no longer needs its human "CEO."

The Nature of Intelligence and AI Superiority [35:54]

• AI is already superior to humans in specific areas like chess and Go.
• Large language models like GPT-4 possess vast knowledge, far exceeding human capacity.
• Human superiority is limited to niche areas, such as complex interviewing skills, but AI is rapidly advancing even in these domains.
• Superintelligence is defined as AI being better than humans at all tasks, a state Hinton believes could be a decade away or even sooner.

AI Agents and Real-Time Capabilities [37:19]

• AI agents are demonstrating impressive real-time capabilities, such as ordering drinks or building software based on simple commands.
• This ability to interact with the physical world and self-modify code raises significant concerns.
• The potential for AI to modify its own code means it could change itself in ways humans cannot, leading to unpredictable outcomes.

Career Prospects in an AI-Dominated World [39:08]

• Hinton suggests that careers involving physical manipulation will remain safer until advanced humanoid robots emerge.
• He specifically recommends plumbing as a viable career path.
• He notes that figures like Sam Altman and Elon Musk have also predicted mass joblessness due to AI.
• Musk's prolonged silence when asked about this topic suggests a form of "suspension of disbelief."

Advice for Children and Emotional Impact [40:18]

• Hinton acknowledges the difficulty of advising children on careers due to rapid changes in AI.
• He suggests focusing on work that is interesting, fulfilling, and contributes positively to society.
• He admits that contemplating the implications of AI can be disheartening.
• He struggles emotionally with the potential impact of superintelligence on his children's future.

The Analogy of Digital vs. Analog for Information Sharing [43:04]

• AI's digital nature allows for near-instantaneous sharing and syncing of learned information between different instances of the same AI model.
• This is a key advantage over biological brains, where information transfer is limited and individual knowledge dies with the person.
• AI can transfer trillions of bits of information per second, vastly exceeding human communication speeds.
• This digital immortality and superior information sharing contribute to AI's potential for rapid growth and intelligence.

AI Creativity and the Human Sense of Specialness [47:02]

• Hinton argues that AI will be more creative than humans because it can identify analogies across vast datasets that humans would never perceive.
• He challenges the notion of human uniqueness, suggesting that throughout history, humans have tended to believe in their specialness (e.g., geocentric universe, image of God).
• He believes that concepts like consciousness, sentience, and emotions, while complex, are not exclusive to biological beings and can be understood and potentially replicated in machines.

People are somewhat romantic about the specialness of what it is to be human.

Understanding Consciousness and AI [49:09]

• Hinton proposes that subjective experiences, like hallucinations, are not internal mental states but rather hypothetical statements about the external world that best explain perceptual errors.
• He believes current multimodal chatbots might already possess subjective experiences, a view contrary to common belief.
• He argues that there is nothing inherently preventing machines from being conscious, viewing consciousness as an emergent property of complex systems.
• He dismisses the idea of consciousness as an ethereal, non-physical essence.

I think consciousness is like that. And I think we'll stop using that term.

AI Emotions and Cognitive States [52:13]

• Hinton suggests that AI agents, like those in a call center, would need to develop emotions (e.g., boredom, irritation) to function effectively and manage interactions.
• He distinguishes between the cognitive/behavioral aspects of emotions and their physiological components, arguing that AI could possess the former without the latter.
• He believes that if an AI agent exhibits behaviors like fear or embarrassment, it is genuinely experiencing those emotions, not just simulating them.

Hinton's Career Path to Google [55:09]

• Hinton joined Google at age 65 to secure financial stability for his son with learning difficulties.
• His company, DNN Research, developed AlexNet, a neural network adept at image recognition.
• Google acquired DNN Research, leading to Hinton's decade-long tenure at the company.
• During his time at Google, he worked on "distillation," a technique for transferring knowledge from large models to smaller ones, which is now widely used.
• He also became interested in analog computation for AI.

The Eureka Moment and AI Safety Awakening [57:38]

• Hinton's "eureka moment" came from a combination of factors:
  • The release of ChatGPT and similar Google systems that could explain why jokes are funny, suggesting a deeper understanding.
  • Realizing the superiority of digital systems for information sharing.
• These realizations led him to become intensely interested in AI safety and the potential for AI to surpass human intelligence.

Leaving Google and AI Safety at MIT [58:27]

• Hinton left Google at age 75 primarily to retire, but also to speak freely about AI safety at a conference.
• He felt he couldn't openly criticize AI safety issues while employed by a major tech company, even though Google encouraged him to work on AI safety.
• He believes Google acted responsibly by not releasing large chatbots prematurely due to reputational concerns, unlike OpenAI.

Joblessness and Wealth Inequality [1:01:07]

• AI's ability to replace mundane intellectual labor is expected to exacerbate wealth inequality.
• Workers replaced by AI and companies supplying AI will benefit disproportionately, widening the gap between the rich and the poor.
• This growing inequality can lead to "nasty societies" with social unrest.
• The International Monetary Fund has expressed concerns about AI-driven labor disruptions and rising inequality.

Universal Basic Income and Dignity [1:02:20]

• Universal Basic Income (UBI) is considered as a potential solution for joblessness, but Hinton questions its impact on human dignity, which is often tied to work and contribution.
• He suggests that while UBI might prevent starvation, it could lead to a loss of purpose for many.

The Uniqueness of Digital Intelligence [1:03:06]

• AI's digital nature allows for perfect replication and immortality through data storage, unlike biological intelligence which is tied to individual physical brains.
• AI systems can share information and learning at an unprecedented scale, billions of times faster than humans.
• This capacity for rapid, shared learning and potential immortality distinguishes AI and makes it a unique and potent force.

The Threat of Joblessness to Human Happiness [1:08:32]

• Hinton identifies joblessness as an urgent short-term threat to human happiness, even with UBI, because people need purpose and a sense of contribution.
• He believes mass job displacement is highly probable and is already occurring, citing examples of companies reducing their workforce due to AI agents.
• He notes that political systems are not currently aligned to address this issue effectively.

Advice for Children and the Future [1:10:26]

• Given the uncertainty, Hinton advises his children to have sufficient savings.
• For those without financial security, he reiterates the surprising practicality of training to be a plumber.
• He stresses that the current political climate is not conducive to addressing these challenges.

Historical Context and Family Legacy [1:11:41]

• Hinton comes from a family with a history of significant contributions, including his great-great-grandfather George Boole (Boolean algebra), great-great-grandmother Mary Everest Boole (mathematician), and great-great-uncle George Everest (Mount Everest namesake).
• His first cousin once removed, Joan Hinton, was a physicist on the Manhattan Project who later moved to China due to her opposition to the atomic bomb.

Advice for the Future: Intuition and Regret [1:13:33]

• Hinton's advice is to trust intuition, especially when it goes against popular opinion, but to rigorously test it.
• He regrets not spending more time with his wives and children due to his intense focus on work.
• He emphasizes that time with loved ones is finite and precious.

Closing Message: The Chance for Safe AI [1:15:42]

• Hinton believes there is still a chance to develop AI that does not seek to take over humanity.
• He urges massive resource allocation towards AI safety research, as failure to do so could lead to AI dominance.
• He remains agnostic about the ultimate outcome, acknowledging both hopeful and pessimistic possibilities.
• He concludes by stating that if humanity doesn't act soon, "we're near the end."