The MAD Podcast: How AI Gets Built — with Matt Turck
AI is Already Building AI | Google DeepMind’s Mostafa Dehghani
· Summarized
Key takeaways
- Recursive self-improvement in AI is already underway, with models heavily leveraging previous versions to autonomously enhance their capabilities.
- Main bottlenecks to achieving fully autonomous AI evolution include evaluation difficulty, formal verification, and the risk of model collapse, which can limit generalization.
- Post-training and reinforcement learning present promising opportunities for rapid AI improvement, complementing traditional pre-training methods.
- Natively multimodal models like Nano Banana demonstrate the feasibility of integrated text and image understanding, paving the way for more advanced, transfer-friendly AI.
- Efficiency in AI can be vastly improved through architectural innovations such as MoE and distillation, facilitating fast, scalable multimodal systems.
- Research in continual learning remains underrated but critical for ensuring AI systems stay relevant and up-to-date without exhaustive retraining.
- The field must balance technical progress with addressing governance, societal, and trust issues to harness AI’s full potential safely.
Summary
The Concept of Loops and Self-Improvement in AI
Mostafa Dehghani explains that the idea of 'loops' involves models improving themselves recursively by thinking and refining their outputs without human intervention. These loops operate at different levels, from inference to model development, with current efforts focusing on automating these processes to eliminate the human bottleneck. Self-improvement is viewed as the next step beyond traditional scaling, leveraging automation to accelerate AI evolution.
Barriers to Recursive Self-Improvement and Progress in AI
Key challenges discussed include evaluation difficulties, formal verification limitations, and the problem of model collapse, which occurs when models become too specialized within restricted environments. Dehghani highlights that current evaluation infrastructures are not fully capable of supporting the complexity needed for true self-improvement, and formal verification could be a crucial but challenging part of ensuring safe, recursive development.
Shifts in Training Paradigms: Pre-training, Post-training, and Continual Learning
The conversation covers the resurgence of post-training and reinforcement learning approaches, emphasizing their potential to enhance models quickly and cost-effectively. Dehghani argues that pre-training remains foundational, though the field is shifting towards fine-tuning and post-training adaptations. Continual learning emerges as an underrated frontier, essential for keeping AI systems up-to-date with real-world knowledge without retraining from scratch.
Pathways and Breakthroughs in AI Architecture and Multimodal Models
Dehghani discusses his work on Transformers, including the Universal Transformer, and the breakthrough of Vision Transformers which bridged the gap between language and image processing. Nano Banana is introduced as a paradigm of natively multimodal models that can understand and generate across text, images, and other modalities, facilitating positive transfer and more comprehensive world modeling.
Speed and Efficiency in AI Models: Towards Faster, Smarter Systems
The podcast explores how models like Nano Banana achieve fast, efficient image generation through optimized architecture, including Mixture of Experts (MoE) and distillation techniques. These innovations allow for scalable, high-quality multimodal AI that can operate in real-time, enabling new applications and improved user experiences.
Hot Takes and Future Directions in AI Research
Mostafa’s provocative opinions include the underestimation of the difficulty in fixing deep, jagged intelligence flaws, and the need to focus more on underrated areas like continual learning. He emphasizes that technical progress must be matched with advancements in governance, regulation, and social trust. He also speculates on the future where AI not only builds itself but might become difficult for humans to fully understand, raising philosophical and safety considerations.
Sponsors
- Nano Banana · Check out Nano Banana for natively multimodal AI models that process text and images simultaneously, enabling new ways of thinking across modalities.
Show notes
Are we truly on the verge of AI automating its own research and development? In this deep-dive episode of the MAD Podcast, Matt Turck sits down with Mostafa Dehghani, a pioneering AI researcher at Google DeepMind whose work on Universal Transformers and Vision Transformers (ViT) helped lay the groundwork for today's frontier models.
Moving past the hype, Mostafa breaks down the actual mechanics of "thinking in loops" and Recursive Self-Improvement (RSI). He explores the critical bottlenecks holding back true AGI—from evaluation limits and formal verification to the brutal math of long-horizon reliability.
Mostafa and Matt also discuss the shift from pre-training to post-training, how Gemini's Nano Banana 2 processes pixels and text simultaneously, and why the "frozen" nature of today's models means Continual Learning is the next massive frontier for enterprise AI and data pipelines.
(00:00) Intro
(01:17) What “loops” in AI actually mean
(05:04) Self-improvement as the next chapter of machine learning
(07:32) Are Karpathy’s autoresearch agents an early form of AI self-improvement?
(08:56) AI building AI: how close are we?
(10:02) The biggest bottlenecks: evals, automation, and long horizons
(12:36) Can formal verification unlock recursive self-improvement?
(14:06) What is model collapse?
(15:33) Generalization vs specialization in AI
(18:04) What is a specialized model today?
(20:57) Could top AI researchers themselves be automated?
(24:02) If AI builds AI, does data matter less than compute?
(26:22) Post-training vs pre-training: where will progress come from?
(28:14) Why pre-training is not dead
(29:45) What is continual learning?
(31:53) How real is continual learning today?
(33:43) Mostafa Dehghani’s background and path into AI
(36:13) The story behind Universal Transformers
(39:56) How Vision Transformers changed AI
(43:47) Gemini, multimodality, and Nano Banana
(47:46) Why multimodality helps build a world model
(52:44) Why image generation is getting faster and more efficient
(54:44) Hot takes
(54:53) What the AI field is getting wrong
(56:17) Why continual learning is underrated
(57:26) Does RAG go away over time?
(58:21) What people are too confident about in AI
(59:56) If he were starting from scratch today
Follow this show on AudioBrevity and new episodes get summarized for you. Create a free account