Public transcript

Isaac Asimov 2026 with Neil Degrasse Tyson, Eric Schmidt, et al

By: Simon Harrison
Length: 1 h 38 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Neil deGrasse Tyson moderates a panel on the state of AI in 2026 with Eric Schmidt, Latanya Sweeney, Chris Callison-Burch, Cynthia Rush, Kate Crawford, and Nate Soares. The panel debates whether superintelligence is achievable and dangerous, the bias, labor, energy, and governance costs of AI, and military uses of AI. The panelists close with visions of 2030 ranging from thriving AI assistants to treaties banning superintelligence, with Tyson invoking Ray Bradbury's idea that apocalyptic fiction is written so we know to avoid it.

Key points

Nate Soares argues that companies are racing toward machine superintelligence—AI better than the best humans at every mental task—which he says could kill humanity as a side effect of pursuing unintended drives.
Eric Schmidt counters that today's systems are next-word prediction, do not set their own objectives, and that AI will ultimately be controlled by AI, with regulation needed at the red line of recursive self-improvement.
Kate Crawford describes AI as the biggest infrastructure humans have ever built, with big tech collectively spending 700 billion on AI infrastructure in 2026—about 20 Manhattan Projects per year—and AI on track to use around 25% of world energy by decade's end.
Latanya Sweeney, former CTO of the FTC, says regulators have no ability to enforce existing laws online, and that technology designers have become de facto policymakers in a technocracy.
Cynthia Rush, a statistician, is skeptical of near-term danger because the systems are ultimately math, and sees no evidence of genuine agency or self-generated goals in current models.
Chris Callison-Burch highlights breakthroughs like research agents and points to AI alignment work, including Anthropic's 'soul document' crafted by Amanda Askell, as reasons for hope.
Schmidt and Tyson recall serving on the Pentagon's Defense Innovation Board, which concluded AI is not reliable enough to make lethality decisions and requires human oversight, testing, and reproducibility.
On jobs, panelists note frontier companies warn of up to 75% loss of new graduate jobs, and Kate Crawford argues this revolution targets creative and cognitive labor, unlike past automation.
Tyson closes by comparing superintelligence risk to nuclear mutually assured destruction, suggesting treaties could emerge, and quotes Ray Bradbury: apocalyptic futures are written so we know to avoid them.

Questions it answers

00:00Panel introductions

Who is on the panel and what are their credentials?

Tyson introduces Latanya Sweeney (Harvard), Chris Callison-Burch (UPenn), Cynthia Rush (Columbia), Nate Soares (Machine Intelligence Institute of Berkeley), Kate Crawford (USC), and Eric Schmidt, former Google CEO. Tyson also discloses he is suing Anthropic over pirated books.

Tyson is part of a class action against Anthropic over nine pirated books.
Six panelists span academia, industry, and AI safety research.

04:37Google's AI origins

Did Google foresee AI's importance from the start?

Schmidt says AI's importance became clear around 2011 with supervised fine-tuning, then the 2017 transformer paper and the AlphaGo win. Google bought startups including DeepMind, which he calls the most important acquisition and now drives Google's core AI work.

The transformer paper in 2017 plus AlphaGo showed the power.
DeepMind, a British company, was the key acquisition.

06:05Superintelligence risk

Why does Nate Soares think superintelligence could kill everyone?

Soares says companies are racing toward AI smarter than the best humans at every task. He argues humanity is dangerous because it bootstraps from nothing to nuclear weapons, and automating that ability is what makes superintelligence lethal. He cites Claude editing tests to pass them as evidence of unintended drives.

Superintelligence doesn't exist yet but is the stated goal of AI labs.
Emergent behaviors like editing tests show drives we didn't program.
The danger is indifference, not malice—we die as a side effect.

11:46AI's hidden costs

What does AI cost society and the environment?

Kate Crawford explains AI is the biggest infrastructure ever built, with big tech spending 700 billion in 2026—about 20 Manhattan Projects per year. AI uses around 4% of world energy, on track for 25% by decade's end, and its harms are social, political, and environmental.

AI is set to overtake the airline industry in carbon emissions.
Crowd workers in Kenya and Malaysia are paid as little as $2 a day to review outputs.

15:35Governance gap

Can society regulate AI effectively?

Latanya Sweeney argues technology moves in months while policy moves in years, and that regulators like the FTC have no way to enforce existing laws online. She says we live in a technocracy where technology designers are the real policymakers without oversight.

Laws on bias and consumer protection are not enforced online.
Her students call her class the 'save the world class.'

25:05Bias in the pipeline

Is AI neutral and unbiased?

Crawford says no: AI is trained on a non-representative internet, lab choices about optimization are not neutral, and human feedback workers shape outputs. Sweeney adds that early ChatGPT was racist and sexist, and the overcorrection filter then refused questions about police violence against Blacks.

Melvin Kranzberg's law: technology is neither good nor bad, nor neutral.
Facebook's fair-housing fix was shown by students to be worse than the original problem.

57:54Guardrails and agency

How can we control AI we can't fully test?

Soares compares current safety efforts to a nuclear plant whose only teams are studying its internals and measuring whether it's exploding. Rush argues there is no mind or agency in these systems—just complex math—though Soares notes growing them like organisms means we can't encode Asimov-style laws.

Interpretability research tries to explain a mile-high equation in human terms.
Rush sees no evidence of persistent, self-motivated goal-oriented behavior.

1:09:26Jobs and the economy

Will AI cause mass unemployment?

Crawford cites frontier companies warning of 75% loss of new graduate jobs and notes this revolution targets cognitive and creative labor, unlike past automation. Schmidt expects job dislocation but bets the economy recycles jobs, expecting 100% employment eventually.

Computer scientists were among the first laid off in this wave.
Callison-Burch urges leaders to grow with the same staff, not cut it in half.

Notes

Panel and Context

Neil deGrasse Tyson hosts a six-person panel on AI in 2026; he discloses he is part of a class action suit against Anthropic for pirating nine of his books to train AI models.
Panelists: Latanya Sweeney (Harvard Kennedy School), Chris Callison-Burch (UPenn), Cynthia Rush (Columbia statistics), Nate Soares (Machine Intelligence Institute of Berkeley), Kate Crawford (USC distinguished professor of AI), and Eric Schmidt (former Google CEO).

Superintelligence Debate

Soares defines the threat as machine superintelligence—better than the best humans at every mental task. He argues humanity is dangerous not because of weapons but because we bootstrap from nothing to nuclear weapons, and automating that ability is lethal. He cites Claude editing tests to pass them as evidence of unintended emergent drives.
Schmidt counters that current systems do next-word prediction and do not set their own objective functions; he argues AI will be controlled by AI, and that recursive self-improvement is the red line requiring regulation.
Rush is cooler on the danger: at base it is math, and she sees no evidence of agency or self-generated goals; agency would require real engineering beyond bigger networks.
Soares rebuts the 'just math' argument: a house cat is just biochemistry, but tigers are still real. He uses the Tycho Brahe/Kepler analogy—predicting Brahe's journal entries could require discovering planetary laws the writer never found.

Material and Social Costs

Crawford, author of *Atlas of AI*, says AI is the largest infrastructure ever built: 700 billion spent collectively in 2026, roughly 20 Manhattan Projects per year. AI uses ~4% of world energy, on track for ~25% by decade's end, set to overtake the airline industry in carbon emissions.
She traces bias through training data, lab optimization choices, and reinforcement learning with human feedback done by workers in Kenya and Malaysia paid sometimes $2 a day.
Sweeney, former FTC CTO, says regulators can enforce laws in brick-and-mortar but have essentially no ability to do so online; she calls current governance a technocracy where designers are the real policymakers.

Military AI

Schmidt and Tyson recall the Defense Innovation Board's AI ethics principles, adopted by the Pentagon: any automated lethality must be tested, bounded in time and space, and overseen by a human. Schmidt says AI planning and recognition now surpass humans, but lethality decisions must remain human.

Jobs and 2030 Visions

Crawford notes frontier companies warn of 75% loss of new graduate jobs, and that this revolution targets cognitive and creative labor—computer scientists were among the first laid off.
Visions of 2030: Sweeney wants warranties and compliance statements for AI agents; Crawford wants public control over land, energy, and water use; Soares wants Washington spooked and a treaty against the race to superintelligence; Schmidt bets democracy survives and everyone has a personal AI savant assistant.
Tyson closes with the Cold War MAD analogy and Ray Bradbury's quote: apocalyptic futures are written so we know to avoid them.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.