Hear This Idea – Détails, épisodes et analyse

Jacob Trefethen oversees Open Philanthropy’s science and science policy programs. He was a Henry Fellow at Harvard University, and has a B.A. from the University of Cambridge.

You can find links and a transcript at www.hearthisidea.com/episodes/trefethen

In this episode we talked about open source the risks and benefits of open source AI models. We talk about:

• Life-saving health technologies which probably won't exist in 5 years (without a concerted effort) — like a widely available TB vaccine, and bugs which stop malaria spreading
• How R&D for neglected diseases works —
• How much does the world spend on it?
• How do drugs for neglected diseases go from design to distribution?
• No-brainer policy ideas for speeding up global health R&D
• Comparing health R&D to public health interventions (like bed nets)
• Comparing the social returns to frontier (‘Progress Studies’) to global health R&D
• Why is there no GiveWell-equivalent for global health R&D?
• Won't AI do all the R&D for us soon?

You can get in touch through our website or on Twitter. Consider leaving us an honest review wherever you're listening to this — it's the best free way to support the show. Thanks for listening!

Elizabeth Seger is the Director of Technology Policy at Demos, a cross-party UK think tank with a program on trustworthy AI.

You can find links and a transcript at www.hearthisidea.com/episodes/seger   In this episode we talked about open source the risks and benefits of open source AI models. We talk about:

• What ‘open source’ really means
• What is (and isn’t) open about ‘open source’ AI models
• How open source weights and code are useful for AI safety research
• How and when the costs of open sourcing frontier model weights might outweigh the benefits
• Analogies to ‘open sourcing nuclear designs’ and the open science movement

You can get in touch through our website or on Twitter. Consider leaving us an honest review wherever you're listening to this — it's the best free way to support the show. Thanks for listening!

Note that this episode was recorded before the release of Meta’s Llama 3.1 family of models. Note also that in the episode Elizabeth referenced an older version of the definition maintained by OSI (roughly version 0.0.3). The current OSI definition (0.0.8) now does a much better job of delineating between different model components.

Jon Y is the creator of the Asianometry YouTube channel and accompanying newsletter. He describes his channel as making "video essays on business, economics, and history. Sometimes about Asia, but not always."

You can see more links and a full transcript at hearthisidea.com/episodes/asianometry

In this episode we talk about:

• Compute trends driving recent progress in Artificial Intelligence;
• The semiconductor supply chain and its geopolitics;
• The buzz around LK-99 and superconductivity.

If you have any feedback, you can get a free book for filling out our new feedback form. You can also get in touch through our website or on Twitter. Consider leaving us a review wherever you're listening to this — it's the best free way to support the show. Thanks for listening!

Steven Teles s is a Professor of Political Science at Johns Hopkins University and a Senior Fellow at the Niskanen Center. His work focuses on American politics and he written several books on topics such as elite politics, the judiciary, and mass incarceration.

You can see more links and a full transcript at hearthisidea.com/teles

In this episode we talk about:

• The rise of the conservative legal movement;
• How ideas can come to be entrenched in American politics;
• Challenges in building a new academic field like "law and economics";
• The limitations of doing quantitative evaluations of advocacy groups.

If you have any feedback, you can get a free book for filling out our new feedback form. You can also get in touch through our website or on Twitter. Consider leaving us a review wherever you're listening to this — it's the best free way to support the show. Thanks for listening! Key links:

Guive Assadi is a Research Scholar at the Center for the Governance of AI. Guive’s research focuses on the conceptual clarification of, and prioritisation among, potential risks posed by emerging technologies. He holds a master’s in history from Cambridge University, and a bachelor’s from UC Berkeley.

In this episode, we discuss Guive's paper, Will Humanity Choose Its Future?.

• What is an 'evolutionary future', and would it count as an existential catastrophe?
• How did the agricultural revolution deliver a world which few people would have chosen?
• What does it mean to say that we are living in the dreamtime? Will it last?
• What competitive pressures in the future could drive the world to undesired outcomes?
  • Digital minds
  • Space settlement
• What measures could prevent an evolutionary future, and allow humanity to more deliberately choose its future?
  • World government
  • Strong global coordination
  • Defensive advantage
• Should this all make us more or less hopeful about humanity's future?
• Ideas for further research

Guive's recommended reading:

• Rationalist Explanations for War by James D. Fearon
• Meditations on Moloch by Scott Alexander
• The Age of Em by Robin Hanson
• What is a Singleton? By Nick Bostrom

Other key links:

• Will Humanity Choose Its Future? by Guive Assadi
• Colder Wars by Gwern
• The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter by Joseph Henrich (and a review by Scott Alexander)

Michael Cohen is is a DPhil student at the University of Oxford with Mike Osborne. He will be starting a postdoc with Professor Stuart Russell at UC Berkeley, with the Center for Human-Compatible AI. His research considers the expected behaviour of generally intelligent artificial agents, with a view to designing agents that we can expect to behave safely.

You can see more links and a full transcript at www.hearthisidea.com/episodes/cohen.

We discuss:

• What is reinforcement learning, and how is it different from supervised and unsupervised learning?
• Michael's recently co-authored paper titled 'Advanced artificial agents intervene in the provision of reward'
• Why might it be hard to convey what we really want to RL learners — even when we know exactly what we want?
• Why might advanced RL systems might tamper with their sources of input, and why could this be very bad?
• What assumptions need to hold for this "input tampering" outcome?
• Is reward really the optimisation target? Do models "get reward"?
• What's wrong with the analogy between RL systems and evolution?

Key links:

• Michael's personal website
• 'Advanced artificial agents intervene in the provision of reward' by Michael K. Cohen, Marcus Hutter, and Michael A. Osborne
• 'Pessimism About Unknown Unknowns Inspires Conservatism' by Michael Cohen and Marcus Hutter
• 'Intelligence and Unambitiousness Using Algorithmic Information Theory' by Michael Cohen, Badri Vallambi, and Marcus Hutter
• 'Quantilizers: A Safer Alternative to Maximizers for Limited Optimization' by Jessica Taylor
• 'RAMBO-RL: Robust Adversarial Model-Based Offline Reinforcement Learning' by Marc Rigter, Bruno Lacerda, and Nick Hawes
• 'Quantilizers: A Safer Alternative to Maximizers for Limited Optimization' by Jessica Taylor
• Season 40 of Survivor

Katja Grace is a researcher and writer. She runs AI Impacts, a research project trying to incrementally answer decision-relevant questions about the future of artificial intelligence (AI). Katja blogs primarily at worldspiritsockpuppet, and indirectly at Meteuphoric, Worldly Positions, LessWrong and the EA Forum.

We discuss:

• What is AI Impacts working on?
• Counterarguments to the basic AI x-risk case
• Reasons to doubt that superhuman AI systems will be strongly goal-directed
• Reasons to doubt that if goal-directed superhuman AI systems are built, their goals will be bad by human lights
• Aren't deep learning systems fairly good at understanding our 'true' intentions?
• Reasons to doubt that (misaligned) superhuman AI would overpower humanity
• The case for slowing down AI
• Is AI really an arms race?
• Are there examples from history of valuable technologies being limited or slowed down?
• What does Katja think about the recent open letter on pausing giant AI experiments?
• Why read George Saunders?

Key links:

• World Spirit Sock Puppet (Katja's main blog)
• Counterarguments to the basic AI x-risk case
• Let's think about slowing down AI
• We don't trade with ants
• Thank You, Esther Forbes (George Saunders)

You can see more links and a full transcript at hearthisidea.com/episodes/grace.

Michael Aird is a senior research manager at Rethink Priorities, where he co-leads the Artificial Intelligence Governance and Strategy team alongside Amanda El-Dakhakhni. Before that, he conducted nuclear risk research for Rethink Priorities and longtermist macrostrategy research for Convergence Analysis, the Center on Long-Term Risk, and the Future of Humanity Institute, which is where we know each other from. Before that, he was a teacher and a stand up comedian. He previously spoke to us about impact-driven research on Episode 52.

In this episode, we talk about:

• The basic case for working on existential risk from AI
• How to begin figuring out what to do to reduce the risks
• Threat models for the risks of advanced AI
• 'Theories of victory' for how the world mitigates the risks
• 'Intermediate goals' in AI governance
• What useful (and less useful) research looks like for reducing AI x-risk
• Practical advice for usefully contributing to efforts to reduce existential risk from AI
• Resources for getting started and finding job openings

Key links:

• Apply to be a Compute Governance Researcher or Research Assistant at Rethink Priorities (applications open until June 12, 2023)
• Rethink Priority's survey on intermediate goals in AI governance
• The Rethink Priorities newsletter
• The Rethink Priorities tab on the Effective Altruism Forum
• Some AI Governance Research Ideas compiled by Markus Anderljung & Alexis Carlier
• Strategic Perspectives on Long-term AI Governance by Matthijs Maas
• Michael's posts on the Effective Altruism Forum (under the username "MichaelA")
• The 80,000 Hours job board

Ben Garfinkel is a Research Fellow at the University of Oxford and Acting Director of the Centre for the Governance of AI.

In this episode we talk about:

• An overview of AI governance space, and disentangling concrete research questions that Ben would like to see more work on
• Seeing how existing arguments for the risks from transformative AI have held up and Ben’s personal motivations for working on global risks from AI
• GovAI’s own work and opportunities for listeners to get involved

Further reading and a transcript is available on our website: hearthisidea.com/episodes/garfinkel

If you have any feedback, you can get a free book for filling out our new feedback form. You can also get in touch through our website or on Twitter. Consider leaving us a review wherever you're listening to this — it's the best free way to support the show. Thanks for listening!

Anders Sandberg is a researcher, futurist, transhumanist and author. He holds a PhD in computational neuroscience from Stockholm University, and is currently a Senior Research Fellow at the Future of Humanity Institute at the University of Oxford. His research covers human enhancement, exploratory engineering, and 'grand futures' for humanity.

This episode is a recording of a live interview at EAGx Cambridge (2023). You can find upcoming effective altruism conferences here: www.effectivealtruism.org/ea-global

We talk about:

• What is exploratory engineering and what is it good for?
• Progress on whole brain emulation
• Are we near the end of humanity's tech tree?
• Is diversity intrinsically valuable in grand futures?
• How Anders does research
• Virtue ethics for civilisations
• Anders' takes on AI risk and whether LLMs are close to general intelligence
• And much more!

Further reading and a transcript is available on our website: hearthisidea.com/episodes/sandberg-live

If you have any feedback, you can get a free book for filling out our new feedback form. You can also get in touch through our website or on Twitter. Consider leaving us a review wherever you're listening to this — it's the best free way to support the show. Thanks for listening!