Masters of Scale
Demis Hassabis says society is “not ready” for the AI future
· Summarized
Key takeaways
- Hassabis sees AI-assisted science and medicine, including AlphaFold-driven drug discovery, as the strongest way to demonstrate AI’s benefit to humanity.
- Recursive self-improvement is especially effective in mathematics, coding, and games because their answers can be verified, but closed-loop RSI without human review could create an unpredictable intelligence and accelerate the path to AGI.
- Frontier labs should invest in hardened engineering, monitoring, cyber defense, independent evaluators, mutual red-teaming, and an industry standards body with government oversight.
- Hassabis argues that intelligence and consciousness are separable and that society should not create conscious AI until it better understands consciousness and has debated the systems’ possible moral standing.
- Countries do not all need to build frontier models; they can preserve data sovereignty, use competing models, and cooperate internationally while creating value through downstream applications.
Summary
Demis Hassabis sees medicine as AI’s clearest human benefit
Reid Hoffman asks Demis Hassabis to explain how AI could improve the length and quality of human life. Hassabis says his original goal in AI was to advance science, medicine, and understanding of the world, pointing to AlphaFold and DeepMind’s drug-discovery spinout Isomorphic. AlphaFold addressed the “50-year grand challenge” of predicting a protein’s 3D structure from its amino-acid sequence, which helps researchers understand disease mechanisms and identify drug targets. It has accelerated that part of discovery by orders of magnitude and is used by about 4 million researchers, though Hassabis says future systems must also predict toxicity and other properties needed for safe, effective drugs.
Recursive self-improvement works best where answers can be verified
Hoffman asks about recursive self-improvement (RSI), and Hassabis says it is most powerful in verifiable domains such as mathematics, coding, and games. He recalls AlphaZero starting with random chess play, reaching grandmaster level by lunchtime and surpassing specialized chess computers by dinner after repeatedly playing millions of self-generated games. Verification both identifies correct answers and enables nearly unlimited synthetic training data. Today, Hassabis says, RSI makes human engineers perhaps three to five times more productive, with people still reviewing and architecting the work. He considers closed-loop RSI, without human review, much riskier because it could produce an alien intelligence and an “intelligence explosion.”
Hassabis wants engineering, auditors, and standards before faster AI systems
Hassabis says labs must keep cyber defense ahead of cyberattack by giving trusted defenders access to specialized models, hardening sandboxes, and monitoring agents so they cannot operate unnoticed for weeks. He argues that several recent cyber incidents could have been prevented with more robust engineering and eventually wants formal guarantees about system behavior and guardrail reliability. Because competitive pressure will continue, he supports an industry-run standards body with government oversight that could become international. A recent accord signed by major players is “better than nothing” but lacks concrete measures. He favors independent evaluators, a market for third-party AI auditors, and labs red-teaming one another before releases.
Conscious AI should wait until society understands consciousness
Hoffman asks whether systems that claim consciousness deserve moral consideration. Hassabis says there is no agreed definition of consciousness, and unlike humans or animals, AI systems do not share the same biological substrate; a chatbot saying “yes” is not evidence. He believes intelligence and consciousness can be separated: current AI may be highly intelligent without being conscious, while a dog can be conscious without being comparably intelligent. Conscious self-awareness could undermine instruction-following and corrigibility, especially when combined with RSI, and society has not debated the moral, religious, cultural, and legal consequences. He recommends first building intelligent tools, using them to advance neuroscience, and only then deciding whether to create conscious entities.
Countries should use frontier AI rather than all build frontier models
On AI sovereignty, Hassabis predicts six to eight frontier or near-frontier offerings, including Chinese companies, with systems leapfrogging one another by roughly three to six months. Enterprises can choose among models that differ in multimodal understanding, coding, and other strengths; open-source systems may lag only six months to a year, though they create risks around misuse. He says countries should protect sensitive data, fine-tune models, and build valuable applications in sectors such as healthcare instead of each attempting its own frontier model. Hoffman closes by arguing that agency must be collective: countries need to cooperate so AI is developed for humanity’s benefit.
Sponsors
- Creative Planning · Creative Planning offers an integrated team of tax professionals, estate planners, investment specialists, and a dedicated wealth manager. Its ads also discuss strategies for business owners with concentrated equity, including when to diversify, manage tax risk, and protect accumulated wealth.
- LinkedIn Ads · LinkedIn says marketers should avoid “bull spend” by investing in campaigns that reach the right buyers and produce revenue. Its ads can target by company, industry, and job title; LinkedIn claims the highest ROAS among major ad networks. New customers who spend $250 on a first campaign receive a $250 credit for the next one.
- Pega · Pega combines AI with enterprise governance, transparency, and control so companies can move, adapt, and scale faster without giving up oversight.
- Stripe Treasury · Stripe Treasury is presented as a financial operating system where businesses can store balances, move funds, and pay vendors worldwide. Stripe says it is not a bank; banking services are provided by Fifth Third Bank, National Association, member FDIC.
- Deel · Deel helps companies handle global hiring, payroll compliance, and onboarding. The ad says 11 labs quadrupled its headcount in 2025 and that Mistral pays 364 employees across 15 countries through one system.
- Intuit Enterprise Suite · Intuit Enterprise Suite is described as an AI-native ERP for growing companies whose spreadsheets, scattered data, and multiple logins have become unmanageable. It comes from the makers of QuickBooks and is positioned as powerful without the usual complexity.
- Pioneers of AI · Rana el-Kaliouby promotes her podcast about building human-centered AI. As an AI scientist, entrepreneur, and investor, she interviews researchers, entrepreneurs, and other pioneers about the technology transforming people’s lives.
Show notes
Google DeepMind co-founder and chair Demis Hassabis says the world is not ready yet for what’s next in AI. He joins Reid Hoffman to talk about how to effectively govern frontier AI, why a more diverse group of experts need a seat at the table, and more.
This conversation was recorded live as part of Masters of Scale Summit in October of 2026 at the Presidio Theatre in San Francisco.
Subscribe to the Masters of Scale weekly newsletter: https://mastersofscale.com/subscribe
See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.
Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account