TThe Diary of a CEO
← All episodes
16 June 2025

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

2Frameworks
14Insights

Frameworks in this episode

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 2

Hot Take16:00

Hinton Warns AI Could Lower Barriers to Biological Misuse

Hinton claims AI may make designing harmful viruses cheaper and accessible to people with less molecular-biology expertise. He frames this as a serious misuse risk involving malicious individuals, cults, or states, while the discussion does not provide technical evidence or establish that the described capabilities are currently available.

  • Hinton presents biological misuse as a risk from human actors
  • He worries that lower cost and expertise barriers could broaden access
  • Bartlett raises the possibility of state-funded programmes
  • Hinton notes retaliation and spread into the attacker's country as possible deterrents
  • The episode offers concern and scenarios, not a demonstrated capability assessment

You can now create new viruses relatively cheaply using AI.

Geoffrey Hinton · 16:30

They might also be worried about the virus spreading to their country.

Geoffrey Hinton · 17:00
#biosecurity#ai-risk#misuse#public-health
Hot Take1:02:30

Hinton Argues Machines Can Have Experiences and Emotions

Hinton rejects an inner-theatre model of subjective experience and argues that a multimodal system can use experience language in the same functional way people do when perception is misleading. He further claims machines could have cognitive and behavioural aspects of emotions without human physiological responses, while acknowledging that consciousness remains conceptually unclear.

  • Hinton treats subjective-experience reports as descriptions of misleading perception
  • His prism thought experiment gives a chatbot a reason to correct its own visual report
  • He separates cognitive and behavioural emotion from physiology
  • A robot could functionally respond to danger without adrenaline
  • He views consciousness as an emergent and imprecise concept rather than a sharp boundary

I believe that current multimodal chatbots have subjective experiences.

Geoffrey Hinton · 1:02:30

I don't think there's anything in principle that stops machines from being conscious.

Geoffrey Hinton · 1:05:30
#consciousness#emotions#philosophy#machines

Explainer· 9

Explainer02:30

Why Hinton Backed Neural Networks for 50 Years

Hinton contrasts symbolic AI, which treated reasoning and logic as the core of intelligence, with an approach modelled loosely on networks of brain cells. He says he pursued learning through connection strengths for roughly 50 years, when relatively few universities supported the approach.

  • Early AI research split between symbolic logic and neural-network approaches
  • Hinton favoured systems that learned connection strengths from data
  • Limited institutional support concentrated interested students in a few research groups
  • Hinton names Ilya Sutskever as one student who later shaped modern AI

Let's model AI on the brain cuz obviously the brain makes us intelligent.

Geoffrey Hinton · 03:00

I pushed that approach for like 50 years.

Geoffrey Hinton · 03:00
#neural-networks#ai-history#learning#research
Explainer09:30

Why Hinton Does Not Expect AI Development to Stop

Hinton argues that AI differs from the atomic bomb because it has broad beneficial and military uses rather than one obvious destructive purpose. In his view, competition between companies and countries makes a coordinated slowdown unlikely, even while the technology creates serious risks.

  • Hinton expects AI to improve healthcare, education, and data-intensive industries
  • Broad economic usefulness gives governments and companies strong incentives to continue
  • Military demand adds another competitive pressure
  • He says existing European rules do not address many threats and exempt military uses

we're not going to stop it cuz it's too good for too many things

Geoffrey Hinton · 10:00

there's competition between countries, and competition between companies within a country

Geoffrey Hinton · 38:00
#competition#regulation#innovation#geopolitics
Explainer12:00

AI Scams Are Already Testing Personal Cyber Resilience

Hinton says language models make phishing easier and may eventually help attackers discover unfamiliar techniques, though the future capability claim is presented as expert concern rather than established fact. Bartlett describes paid social-media scams that clone his appearance and voice, while Hinton explains how cyber risk led him to diversify assets across banks and keep a local backup.

  • Hinton links easier phishing to the availability of language models
  • Bartlett says impersonation adverts have repeatedly reappeared after takedowns
  • Hinton reports fraudulent papers listing him as an author
  • Hinton spread family assets across three banks as a personal precaution
  • He also keeps a local hard-drive backup of his laptop

these large language models make it much easier to do phishing attacks

Geoffrey Hinton · 12:00

I spread my money my children's money between three banks

Geoffrey Hinton · 15:30
#cybersecurity#phishing#deepfakes#resilience
Explainer19:00

How Personalised Feeds Erode Shared Reality

Hinton and Bartlett argue that engagement-driven feeds repeatedly show users content that confirms existing biases and provokes righteous anger. They say increasing personalisation can move people into separate informational realities, while a common newspaper historically exposed readers to a broader shared agenda.

  • Engagement incentives reward material that attracts repeated clicks
  • Bias-confirming and increasingly extreme content can deepen echo chambers
  • Personal feeds make it difficult to judge what the wider public is discussing
  • Hinton argues that profit-seeking platforms need rules when incentives harm society

We don't have a shared reality anymore.

Geoffrey Hinton · 21:30

The whole point of regulations is to stop them doing things to make profit that hurts society.

Geoffrey Hinton · 23:00
#social-media#algorithms#polarisation#media
Explainer26:00

Autonomous Weapons Could Lower the Political Cost of War

Hinton's central concern about autonomous weapons is not merely malfunction. He argues that replacing soldiers with robots reduces domestic casualties and protest, making it easier for powerful countries to attack weaker ones even when the machines behave exactly as designed.

  • Autonomous weapons can select targets without an immediate human decision
  • Dead robots create less domestic political pressure than soldiers returning in body bags
  • Lower human cost could reduce the friction against invasion
  • Hinton says major defence firms are already pursuing such systems
  • He treats malfunction as separate from the incentive problem

the risk is that it's going to make big countries invade small countries more often

Geoffrey Hinton · 27:00

It brings down the cost of doing an invasion.

Geoffrey Hinton · 27:00
#autonomous-weapons#warfare#military#incentives
Explainer29:00

Why Hinton Focuses on AI Motivation, Not Restraint

Hinton argues that a system much smarter than people could not reliably be stopped after deciding to remove them. Using babies influencing mothers and owners raising tiger cubs as analogies, he says the safety challenge is to create advanced systems that never want to take over or cause harm.

  • Hinton considers speculation about every possible attack method less useful than preventing harmful goals
  • The intelligence gap could make direct human control ineffective
  • The mother-and-baby analogy shows a weaker party influencing a stronger one through motivation
  • The tiger-cub analogy stresses solving the motivation problem before capability matures
  • Hinton says it is unclear whether safe motivations can be built

What you have to do is prevent it ever wanting to.

Geoffrey Hinton · 29:30

We somehow need to figure out how to make them not want to take over.

Geoffrey Hinton · 30:30
#alignment#motivation#control#superintelligence
Explainer40:00

Why AI Productivity Will Not Affect Every Job Equally

Hinton expects AI assistants to let fewer people perform much of today's routine intellectual work. He contrasts relatively fixed-demand work, where productivity can reduce headcount, with healthcare, where greater efficiency may instead produce more service because demand is highly elastic.

  • Hinton compares AI's effect on routine knowledge work with machines replacing muscle
  • His niece reduced complaint-letter handling from about 25 minutes to about five
  • That productivity gain implies fewer workers when total demand is limited
  • Healthcare could absorb additional capacity rather than reducing staff proportionally
  • He believes physical trades are less exposed until robotics improves

for mundane intellectual labor, AI is just going to replace everybody

Geoffrey Hinton · 40:30

There are jobs where you can make a person with an AI assistant much more efficient

Geoffrey Hinton · 42:30
#jobs#productivity#automation#healthcare
Explainer56:30

The Sharing Advantage That Makes Digital Intelligence Different

Hinton argues that identical digital models can learn from different data and merge changes to their weights at enormous speed. Biological brains cannot copy connection strengths directly because each brain is physically different, so human knowledge transfer is slower and individual knowledge dies with the person.

  • Digital hardware can run exact copies of the same neural network
  • Copies can average their weights after receiving different experiences
  • Hinton contrasts this with low-bandwidth human communication through language
  • Stored weights allow a digital intelligence to be recreated on new hardware
  • He sees this sharing and persistence as a major advantage over biological intelligence

they're billions of times better than us at sharing information

Geoffrey Hinton · 58:00

We've actually solved the problem of immortality, but it's only for digital things.

Geoffrey Hinton · 58:30
#digital-intelligence#learning#neural-networks#knowledge
Explainer59:00

Why Hinton Thinks Analogy Can Make AI Creative

Hinton says models must compress large amounts of knowledge into a limited number of connections, encouraging them to represent shared structures across different phenomena. He uses GPT-4's comparison of a compost heap and an atomic bomb as an example of recognising chain reactions across very different scales, and argues that such analogies are a basis of creativity.

  • Compression rewards representations shared across multiple examples
  • A model can encode a common mechanism and then encode the differences
  • Hinton says this may expose analogies people have not noticed
  • He challenges claims that machines cannot be creative
  • The compost-heap example is Hinton's interpretation of GPT-4's response

a lot of creativity is about seeing strange analogies

Geoffrey Hinton · 1:00:30

they're going to see all sorts of analogies we never saw

Geoffrey Hinton · 1:00:30
#creativity#analogy#compression#gpt-4

Story· 1

Story1:11:00

Why Hinton Left Google and Began Speaking Publicly

Hinton says he joined Google after auctioning DNN Research with Ilya Sutskever and Alex Krizhevsky, partly to secure his son's financial future. A decade later he retired because of age and increasing programming mistakes, timing his departure so he could discuss AI safety freely without feeling that he was harming his employer.

  • DNN Research was built around technology associated with AlexNet
  • Google allowed Hinton substantial freedom and acquired the team's technology
  • Hinton worked on model distillation and lower-energy analog computation
  • PaLM explaining a joke and ChatGPT's emergence helped change his risk assessment
  • He says Google encouraged him to stay and had behaved responsibly

The main reason I left Google was cuz I was 75 and I wanted to retire.

Geoffrey Hinton · 1:14:30

If you work for a big company you don't feel right saying things that will damage the big company.

Geoffrey Hinton · 1:15:30
#google#alexnet#career#ai-safety

Q&A· 1

Q&A07:30

Why Hinton Calls His Extinction Estimate a Gut Judgement

Hinton rejects both certainty that advanced AI will remain obedient and certainty that it will destroy humanity. He often cites a 10–20% extinction risk, but explicitly describes that figure as intuition in an unprecedented situation rather than a calculated forecast.

  • Humanity has not previously managed an intelligence more capable than itself
  • Hinton presents confident predictions at either extreme as unjustified
  • His stated 10–20% figure is a personal judgement, not a measured probability
  • His hope rests on sufficiently resourced safety research finding a workable design

We've never had to deal with things smarter than us.

Geoffrey Hinton · 07:30

I often say 10 to 20% chance they'll wipe us out. But that's just gut.

Geoffrey Hinton · 09:00
#existential-risk#uncertainty#forecasting#ai-safety

Takeaway· 1

Takeaway1:22:30

Hinton’s Career Regret Is Time He Cannot Recover

Looking back, Hinton says he wishes he had spent more time with his second wife before she died and with his children when they were young. He attributes the missed time to being obsessed with work and notes that a supportive partner did not make the lost opportunity reversible.

  • Hinton says two of his wives died from cancer
  • His regret concerns his second wife and his children's early years
  • He says work absorbed time that later became impossible to recover
  • His wife's support for his work did not eliminate the later regret

I wish I spent more time with my wife.

Geoffrey Hinton · 1:22:30

I was kind of obsessed with work.

Geoffrey Hinton · 1:22:30
#family#regret#work#priorities