AI's Scaling Plateau: The Urgent Return to Foundational Research
Related Episodes
AI's Maturing Market: Infrastructure, Devices, and Societal Risks
AI transforms computing with category-defining products, integrated devices, and powerful hardware, but rising costs and new business models introduce inflation and societal risks.
View Episode Notes →
AI's Generalization Gap Requires Research Over Scaling
AI models excel on benchmarks but fail in the real world due to reward hacking. True AGI requires shifting from scaling to research-driven innovation focused on human-like learning and value functions.
View Episode Notes →
AI: Amplifying Work, Not Replacing Jobs
AI transforms jobs by automating tasks, not eliminating roles, demanding continuous learning and critical evaluation to navigate its rapid, unpredictable evolution and unlock new opportunities.
View Episode Notes →
Mitigating Existential Risks: Aging, Cosmic Threats, Nuclear War, and AI
Aging is the most probable existential risk, yet "thick-skinned weirdos" champion mitigation of diverse threats like asteroids, nuclear war, and AI before they become obvious.
View Episode Notes →
Prioritizing Hard Tech "Good Quests" Over AI Hype
Focus on "good quests" in hard tech, not easy AI applications, to tackle essential, difficult problems and achieve true societal impact.
View Episode Notes →
Compounding AI Advantage: Reinvestment and Agentic AI Accelerate Leadership Gap
Leading organizations build compounding AI advantages by reinvesting gains into complex workflows, creating a widening gap that laggards struggle to match.
View Episode Notes →Resources
Resources & Recommendations
Research & Studies
- "RL scaling" paper (mentioned in podcast, specific title not provided) - This paper discusses the learning curve on RL looking like a sigmoid, which is different from the power law observed in pre-training.
Tools & Software
- Gemini 3 - An AI model used by the host to brainstorm ideas, find connections between research concepts, code experiments, and refine hypotheses, particularly in the context of RL scaling and its theoretical underpinnings.
- Google Colab notebook - Used by the host to run toy ML experiments coded by Gemini 3.
- Transcriber tool (developed with Labelbox) - A tool used for generating podcast transcripts, designed to make conversations read like standalone essays by addressing issues like fits and starts, and confusing phrasing.
People Mentioned
- Kasparov - Referenced in the context of early narrow AI in chess, where a chess AI could beat him but was limited to that specific task.
- Ilya Sutskever - The guest of the podcast, co-author on significant deep learning advancements like AlexNet and GPT-3, known for his research taste in AI.
- Alex Krizhevsky - Lead author of AlexNet.
- Geoffrey Hinton - Often credited with pioneering deep learning and being a key figure in the development of AlexNet.
Organizations & Institutions
- OpenAI - Mentioned as a frontier company in AI research, with its charter defining AGI, and as a participant in early collaborations on AI safety.
- Anthropic - Mentioned as a frontier company in AI research and as a participant in early collaborations on AI safety.
- SSI (Safe Superintelligence Inc.) - Ilya Sutskever's company, focused on a "straight shot" to superintelligence and emphasizing a different technical approach to AI safety and generalization.
- Meta - Mentioned as the company that attempted to acquire SSI and where Ilya Sutskever's former co-founder went.
- Google - Context of host using Gemini, a Google product.
- Stanford - Mentioned as a place where research in AI was conducted during the "age of research."
Websites & Online Resources
- Labelbox.com/dwarkesh - Website to learn more about Labelbox and their transcriber tool.
- Gemini.Google - Website to check out Gemini 3.
- Sardine.ai/dwarcash - Website to learn more about Sardine's AI fraud detection and download their guide.
Other Resources
- AlexNet - A convolutional neural network that significantly advanced the field of deep learning, mentioned as an example of research using relatively small compute.
- Transformer (architecture) - A neural network architecture that became a foundational model in natural language processing, mentioned as an example of research using moderate compute.
- ResNet - A deep learning architecture, mentioned as an example of research.