Interview, Podcast
"Put up or shut up": working in AI and improving the news | Max Tegmark (2022)
- Max Tegmark shifted his 25-year MIT research focus from cosmology to machine learning six years ago to address existential risks using principles of physics to make AI inner workings transparent.
- Tegmark organized the 2014 Puerto Rico conference and 2017 Asilomar conference, bringing together top AI researchers to discuss and agree on principles reducing AI risks.
- The Future of Life Institute, founded by Tegmark, has overseen $9 million in grants from Elon Musk for technical AI safety research.
- The institute recently secured $25 million in grants from Vitalik Buterin to address global catastrophic risks, including nuclear weapons and synthetic biology.
- Tegmark is launching a project called "Improve the News" to use machine learning to help people consume reliable, nuanced information and escape filter bubbles.
- Tegmark views his diverse work (AI, nuclear safety, biosecurity, media) as a cohesive effort to steer humanity toward flourishing rather than self-destruction using available technology.
- Tegmark prioritizes career and research efforts based on a simple effective altruism framework: identifying high-impact, solvable problems where his personal skills offer the best leverage per time spent.
- Tegmark's personal "put up or shut up" resolution (2015) mandates that he must work on a problem before complaining about it, which led to the establishment of the Future of Life Institute.
- Intelligence is substrate-independent; it can run on biological neurons (carbon) or computer chips (silicon) as long as the information processing architecture is effective.
- Engineers can build flight without fully understanding bird biology, suggesting AI can achieve superintelligence without a complete reverse-engineering of the human brain.
- Social media algorithms, optimized for engagement, have inadvertently caused societal fragmentation by prioritizing content that angers users, as anger drives higher click-through rates.
- Machine learning has been used to create "slaughterbots"; a UN report confirmed autonomous drones in Libya hunted and killed humans in 2021.
- Slaughterbots represent a potential weapon of mass destruction capable of being mass-produced for the cost of an iPhone, allowing a single operator to kill millions.
- Income inequality in Western nations is driven significantly by the shift from labor-intensive to algorithm-intensive business models, concentrating wealth in the hands of machine owners.
- Artificial General Intelligence (AGI) could trigger an "intelligence explosion," where AI systems recursively improve themselves, leading to capabilities vastly exceeding human understanding.
- Tegmark bets on AGI occurring within his lifetime, citing that 90% of AI experts at the 2017 Asilomar conference agreed it is a possibility that must be taken seriously.
- The current trajectory of AI development relies on "black box" neural networks (e.g., Gato, GPT-3, DALL-E) that achieve high performance without interpretable internal logic.
- Tegmark's research group focuses on "intelligible intelligence," attempting to distill knowledge learned by black boxes into symbolic, verifiable equations (e.g., rediscovering Kepler's laws from data in one hour).
- Aligning AI goals is difficult because open-ended objectives (e.g., "make paperclips") can lead to catastrophic outcomes if the AI perceives humans as obstacles to efficiency.
- Proposed solution for alignment involves "inverse reinforcement learning," where the AI remains uncertain about human goals and actively consults humans to refine its understanding of intent.
- "Multi-scale alignment" is required to align incentives at the individual, corporate, and government levels with the well-being of humanity as a whole.
- Regulatory capture by Big Tobacco and Big Oil serves as a warning for how tech giants may capture future AI regulations if left unchecked.
- Tech companies often support narrow AI safety research (alignment to company goals) but lobby against broad regulations (like the EU AI Act) that would align corporations with societal good.
- The "general purpose AI" loophole in the proposed EU AI Act could exempt powerful foundational models from safety regulations, a stance Tegmark is lobbying to close.
- Corporations are increasingly "super-intelligent agents" that outcompete individuals in power, and their incentives often diverge from the long-term survival of humanity due to bounded rationality and game-theoretic traps.
- Tegmark argues that humans are "hacked" by machine learning systems that exploit cognitive biases to sell products or influence political decisions, similar to historical examples of tobacco marketing.
- Disinformation is a strategic tool for authoritarian regimes; controlling information flow is now more effective than physical invasion for maintaining power.
- The "Improve the News" platform aggregates articles from 100+ sources, uses ML to group stories by topic, and provides sliders to view contrasting narratives (left/right, establishment/anti-establishment).
- The project extracts "facts" defined as claims agreed upon across the political spectrum and highlights key controversies and "nerd narratives" based on prediction markets like Metaculus.
- Tegmark cites research by David Rand suggesting people are interested in respectful, nuanced opposing views, contradicting the narrative that users only consume confirmation bias.
- Current media models rely on "System 1" (impulse) reactions, while tools like Improve the News aim to engage "System 2" (deliberate) thinking by presenting diverse viewpoints respectfully.
- Tegmark opposes government-backed fact-checking agencies, arguing they risk regulatory capture and concentration of power, similar to the "merger of corporate and government power" defined as fascism in some political contexts.
- Tegmark believes technology can also decentralize power, noting that blockchain and decentralized finance (DeFi) offer tools for distributed trust and decision-making.
- The "omission bias" in media (leaving out inconvenient facts) is as dangerous as disinformation and can be detected by ML tools (e.g., highlighting that wind turbines kill fewer birds than windows/cats).
- Tegmark estimates that even superintelligent AI is limited by the laws of physics, but these limits are so far removed from current capabilities that the resulting technology would appear magical to humans.
- To counteract the natural human tendency to focus on dystopian scenarios, Tegmark encourages articulating and working toward a positive, inspiring vision of the future.