Apr 2, 2026 · 1h 4m · mad
AI is Already Building AI — Google DeepMind’s Mostafa Dehghani
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, Google DeepMind research scientist Mostafa Dehghani discusses the shift toward self-improving AI systems, examining test-time compute, autonomous research loops, architectural breakthroughs like Vision Transformers and Nano Banana, and the key bottlenecks remaining in AI evaluation and real-world grounding.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 14.1% of the talking time here. How this is scored →
Matt as informed peer 4.1 Guest teaching 3.5 Guest disagreement 0.3 Matt pushing back 0.8
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 58:24 Rejecting technical-only AI progress
Mostafa explicitly disagrees with the widespread optimism that purely technical advancements will automatically solve broader societal, governance, and institutional challenges.
Hardest push from Matt ▶ 28:14 Challenging 'pre-training is dead' narrativeMatt directly confronts the guest with the recent industry consensus that pre-training is dead, pushing Mostafa to explicitly defend his contrasting stance.
Biggest teaching moment ▶ 1:00:45 Compounding probability in multi-step agentsMostafa educates the host on agent reliability using concrete probability math, demonstrating that 95% accuracy over 100 steps results in less than 1% overall task success.
Matt holds his own ▶ 39:57 Citing exact Vision Transformer paper titleMatt demonstrates deep familiarity with Mostafa's research history by reciting the exact title and patch dimensions of the seminal Vision Transformer paper.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Decoupling AI Looping: Inference Compute versus Development Self-Improvement | 3 | 4 | 0 | 0 | Matt opens by asking Mostafa to clarify the concept of recursive AI loops. Mostafa provides a structured breakdown separating inference compute looping from development self-improvement. | |
| Recursive Self-Improvement and Autonomous AI Research via AutoResearch | 6 | 3 | 0 | 2 | Matt cites Karpathy's auto research project and ICLR papers, synthesizing the key distinction between AI assistance tools and direct weight updates. Mostafa validates this framing while noting missing long-horizon capabilities. | |
| Evaluation Bottlenecks and Safe Infrastructure for Self-Improving Agents | 4 | 4 | 0 | 1 | Matt identifies evaluation as a core roadblock in self-improving systems. Mostafa agrees and explains the philosophical and infrastructure hurdles of running safe evaluation environments. | |
| Formal Verification and Grounding Signals to Prevent Model Collapse | 6 | 4 | 1 | 2 | Matt references formal verification from a prior guest conversation with Karina Hong of Axiom Math and asks Mostafa to define model collapse. Mostafa notes formal verification works well for code/math but struggles in messier domains. | |
| Navigating the Trade-Off Between Model Specialization and Generalization | 5 | 5 | 0 | 1 | Matt asks about the tension between model specialization and generalization in self-improving loops. Mostafa explains how post-training overfitting and organizational strategy drive short-term specialization. | |
| Automating AI Researchers and Essential Future Human Skills | 3 | 3 | 1 | 1 | Matt asks whether top researchers like Karpathy are automating themselves out of a job. Mostafa reframes the question around high-level strategic decision-making rather than pure technical execution. | |
| Redefining Data: Physical Grounding and Multi-Sensory Environment Interaction | 4 | 4 | 0 | 1 | Matt questions whether data remains essential if AI creates itself, noting the rise of sensors-as-a-service. Mostafa expands the definition of data to include multi-sensory physical grounding. | |
| Balancing Pre-Training and Post-Training Gains in AI | 5 | 3 | 1 | 4 | Matt directly pushes back against the prevailing narrative that pre-training is dead. Mostafa agrees that pre-training continues to unlock major gains through exotic new recipes. | |
| Defining Continual Learning vs. Self-Improvement Loops | 4 | 4 | 0 | 1 | Matt prompts Mostafa to define continual learning versus self-improvement loops. Mostafa explains that continual learning focuses on staying current to prevent catastrophic forgetting. | |
| Mostafa Dehghani's Career Journey and the Universal Transformer | 5 | 3 | 0 | 0 | Matt demonstrates research background knowledge by citing Mostafa's 2018/2019 Universal Transformer paper. Mostafa reflects on working with Lukasz Kaiser and the origins of test-time compute. | |
| The Vision Transformer (ViT) and Unifying Multimodal Architectures | 6 | 3 | 0 | 0 | Matt recites the exact title of the Vision Transformer paper and synthesizes its role in unifying multimodal architectures. Mostafa details the simple patchify intuition behind ViT. | |
| Nano Banana: Multimodal World Models and Interleaved Generation | 5 | 4 | 0 | 0 | Matt details Google's Nano Banana release timeline and asks how native multimodality differs from standard text-to-image translation. Mostafa explains reporting bias and interleaved generation planning. | |
| Technical Efficiency and Serving Optimizations in Nano Banana 2 | 4 | 3 | 0 | 0 | Matt asks about the efficiency gains and architecture behind Nano Banana 2 Flash. Mostafa attributes speedups to model sizing, distillation recipes, and serving infrastructure optimizations. | |
| The Problem of Jagged Intelligence in AI Systems | 3 | 4 | 1 | 0 | Matt prompts Mostafa for a hot take on what the AI field is getting wrong. Mostafa highlights jagged intelligence as a fundamental structural issue rather than a patchable bug. | |
| Overconfidence in Purely Technical Solutions for AI | 3 | 5 | 1 | 0 | Mostafa rejects the assumption that technical progress alone is sufficient, warning against overconfidence and illustrating agent compounding error rates with concrete math. | |
| Concluding Remarks and Final Thanks | 0 | 0 | 0 | 0 | Brief closing pleasantries, sign-off, and audience housekeeping. |
Statements from this episode (27)
AI looping is a top R&D investment for almost every major lab
“Definitely one of the top is active areas for almost every lab to invest in looping”
Every major AI lab uses previous model generations to build new ones
“In almost every lab the new generation of the models are built heavily using the previous generation of the models. I think that's basically like the case again, everywhere.”
Fully automated AI self-improvement will eliminate human bottlenecks and trigger breakthroughs
“The moment that we had this full automation, I would say we can close the loop of self-improvement and then it becomes the Like, you know, the problems become like, you know, mostly providing compute for these models to actually do what they want to do. And as…”
Karpathy's AutoResearch is an early example of AI doing sensible research
“That is definitely. And I think that was one of the early examples of like seeing these models actually doing something super sensible on the research side.”
AI research lacks evaluations to measure progress toward true self-improvement loops
“And the fact that we don't have evals that like, or like even defining evals that, that can maybe measure, oh, not how close we are to the point that we can actually get, get a self-improvement loop. It's just like, we don't have that. And it's just making it …”
Safe execution environments are currently the bottleneck for autonomous AI development
“Like in a safe setup, you know, where they can like, because right now we definitely, we don't, we're not confident about, you know, them doing the right things all the time and measuring like how much they can push and how long they can push a task is very di…”
Formal verification is crucial for AI self-improvement, but not a silver bullet
“In my opinion, formal verification is one of the most powerful like keys to enable like self-improvement, but it's not beaky.”
Anchoring self-improvement loops with real external signals prevents AI model collapse
“Model collapse mainly happens when you have a loop that is Completely closed. Right. And if you don't have any outside signal and just the model, for example, talking to itself or operating in a very like a restricted environment there's a good chance that you…”
Building specialist models is currently the fastest path toward generalist AI
“Short term I would say like building a specialist model is like probably the fastest way to learn like what is actually possible. And in, in many cases, these like specialized model are becoming Stepping stone toward a generalist model, which is like super val…”
Becoming a deep expert in a narrow subject will soon lose value
“Becoming absolute expert about a very specific subject, most likely Is not going to be useful in, in, in like the near future.”
AI data work will shift toward physical world grounding and interactive environments
“At the end of the day, I think, like, the work that we're doing on the data side most likely is going to shift toward building environments or making sure that these models can interact with physical words, and then it becomes more of a problem of, okay, how c…”
AI model progress will alternate between pre-training and post-training breakthroughs
“We're going to be having a bit of a swing back and forth between pre-training and post-training.”
Post-training techniques cannot compensate for a weak base AI model
“Pre-training is still the foundation and like, you can never post-train your way out of a week-based model.”
New pre-training techniques will drastically boost base AI model capabilities
“The way that we used to do pre-training, maybe, like, you know, like two, a year ago or two years ago maybe, like, you know, diminishing return is, like, obvious, but I can see how new ideas are bringing, like, you know, fresh, fresh energy into the pre-traini…”
Self-improvement increases AI capability; continual learning keeps its knowledge fresh
“So, self-improvement is about a model getting smarter. Over time and improving its capability, like the model itself doing it. Continual learning is mostly about a model staying current, right?”
AI continual learning research still lacks a proven recipe for productionization
“Like one side is, I think like the research is not like yet to a very, like to a point that you think that, oh, you know, this is the recipe, you know, I just need to kind of like, you know, exploit it and push productionization.”
Dehghani initially thought Google's Transformer architecture was random and would die
“And I was like, I don't know if I want to go with this team. It's just like, they're doing something random. Like who, like everybody's doing LST. I'm like, why should I go and work with like a group of people who are working on this like random architecture, …”
Looping provides parameter-free FLOPS, whereas MoE architectures provide FLOPS-free parameters
“In mixture of experts, you have flops free parameters. So, so parameters that they're not actually bringing any flops. And in, in like looping, you have parameter free flops where you don't have extra parameters for the extra flops that you're throwing on this…”
Vision Transformers benefited heavily from software infrastructure built for language models
“There was also benefit of doing that simply because the rest of the, that the machine learning field, which was working on, on, on language, they were using this, like architecture. So they were building infra for it, making it faster. And, you know, like the,…”
Unified Transformer architectures simplified the training of natively multimodal AI models
“Even if this is not, like, the only architecture that would be, like, in a multi-model, but it made it really simple to train these models, like, natively, because you have, like, a single architecture and you can have all the modalities in, during training.”
Video data conveys physical world knowledge to AI more efficiently than text
“So because of that, like picking up a lot of knowledge about the word through language is just not really efficient. I don't want to say that it's impossible, but it's not efficient, you know, like to learn about gravity. If you kind of like, you know, have yo…”
Demonstrating that image training lowers text perplexity remains extremely difficult
“So it turned out to be a really, really good model, but it was like really hard to see that. Wow. You know, I train on images and then like Text perplexity goes down. That was hard to see. You know, like the fact that, you know, you train in native model and i…”
Interleaved multimodal generation overcomes single-shot limits via step-by-step planning
“But if you have incremental generation, so if you have text and then an image and text and image, you can get your model to generate these details one by one. So you never expect your model to generate an image, a perfect image in the first shot, right? So, so…”
Jagged intelligence in AI is a structural flaw, not a patchable bug
“Not easy to pinpoint like specific things, but again, like, you know, this is just like my personal opinion and maybe I have colleagues and like the other people like sharing this with me, but I think we're underestimating how hard like jagged intelligence is …”
RAG will shift from universal use to handling long-tail distribution cases
“Maybe it changes in a way that, you know, like it doesn't need to trigger RAG for like everything, but I'm pretty sure that they're going to be some tail of the distribution that we're going to do RAG still for it.”
Single AI failures damage user trust more than average performance builds it
“People don't experience average performance of these models. They experience, like, the failures. If you have your model doing a dumb mistake, the damage in the trust that it makes is, like, bigger than, like, you know, the benefit of getting hundred things ri…”
Real-world physical grounding will become the bottleneck for AI self-improvement
“As I said, you know, like soon, like the concept of data, like, you know, how to kind of like, you know, enable these models to kind of like, you know, be very good at like self-improvement becomes, how can I ground these models in, in real world? So this is d…”