AI Is Already Building AI — Google DeepMind’s Mostafa Dehghani

Data Driven NYC
Data Driven NYCApr 2, 2026

Why It Matters

Recursive, automated AI development could compress years of research into weeks, reshaping competitive advantage while introducing urgent safety and evaluation challenges for businesses and regulators.

Key Takeaways

  • Labs now use prior models to train next generation AI.
  • Full automation of self‑improvement remains the primary technical hurdle.
  • Recursive self‑improvement could accelerate progress beyond human‑engineered loops.
  • Reliable evaluation and formal verification are critical to prevent model collapse.
  • Enterprise pipelines may be disrupted by continual‑learning, AI‑built‑AI systems.

Summary

In this episode of the Matt Podcast, DeepMind researcher Mostafa Dehghani explains how the AI field is moving from human‑driven model design to a regime where models iteratively build the next generation of models. He frames the shift as a series of "loops"—micro‑level loops that add compute at inference time and macro‑level loops that automate the entire development pipeline, effectively removing the human bottleneck that has historically limited progress. Dehghani highlights that today’s leading labs already train new architectures on the outputs of previous generations, but full‑automation and long‑horizon self‑improvement remain unfinished. He argues that once models can evaluate and update their own weights without external supervision, a dramatic acceleration—recursive self‑improvement—will follow, provided sufficient compute and robust evaluation metrics are in place. The conversation cites concrete examples such as the Kapathies auto‑research project, where models began contributing to research engineering tasks, and discusses formal verification as a promising, though incomplete, tool to ensure safe feedback loops. Dehghani also warns of model collapse when a closed loop lacks external grounding, defining it as loss of generalization after over‑optimizing on self‑generated data. If these challenges are solved, enterprises could see their data pipelines and retrieval‑augmented generation systems rebuilt around continuously learning AI, reshaping competitive dynamics and raising new governance questions about safety, evaluation, and control.

Original Description

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.
Mostafa Dehghani
Google DeepMind
Matt Turck (Managing Director)
FirstMark
Listen on:
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

Comments

Want to join the conversation?

Loading comments...