Oct 22, 2019 · 39m · mad
Fireside Chat: Pedro Domingos, Head of Machine Learning, DE Shaw (FirstMark's Data Driven NYC)
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 Data Driven NYC fireside chat hosted by Matt Turck, computer scientist Pedro Domingos discusses the application of machine learning in financial markets, the theoretical quest for a unifying 'master algorithm,' and key debates surrounding AI explainability, societal risks, and biological intelligence.
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 9.4% of the talking time here. How this is scored →
Matt as informed peer 2.6 Guest teaching 3.9 Guest disagreement 2.1 Matt pushing back 0.6
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:30 Dismissing superintelligence threat in favor of dumb AI risks
Domingos forcefully rejects popular media narratives surrounding AI existential risk, asserting that stupid AI already making decisions is far more dangerous than superintelligent computers taking over.
Hardest push from Matt ▶ 2:47 Matt questions whether ML adoption in finance is still early daysMatt politely pushes back on standard industry claims, asking Domingos to verify whether machine learning adoption in capital markets is genuinely early stage compared to standard quant approaches.
Biggest teaching moment ▶ 19:40 Illustrating machine learning generalization inefficiencyDomingos educates the audience on the massive generalization gap by contrasting OpenAI's Rubik's cube robot requiring 10,000 simulated years with a human child mastering it in a single month.
Matt holds his own ▶ 17:03 Matt asking about capsule networks and two-way backpropMatt demonstrates deep familiarity with cutting-edge research topics by specifically naming Hinton's capsule networks and non-standard propagation methods.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Applying Machine Learning to Capital Markets | 3 | 3 | 1 | 1 | Matt Turck opens by asking why Domingos chose capital markets over other fields and notes that machine learning seems surprisingly early stage in finance compared to traditional quant strategies. Domingos clarifies that finance was actually an early killer app for neural networks in the late 1980s, though current computing power and data make modern ML applications vastly more capable. | |
| The Promise and Scope of Alternative Data | 2 | 2 | 0 | 0 | Matt inquires about alternative data and then transitions to Domingos's book, asking for its core premise. Domingos explains the necessity of writing a popular science book about machine learning for general decision-makers and introduces the concept of the master algorithm. | |
| Candidates for the Master Algorithm and Unifying Paradigms | 3 | 4 | 2 | 1 | Matt asks which paradigm is currently closest to becoming the master algorithm, suggesting reinforcement learning. Domingos argues that backpropagation alone will not reach general intelligence because it only solves credit assignment, emphasizing the need to unify connectionist, symbolic, and evolutionary approaches. | |
| Frontiers of AI Research and the Generalization Gap | 4 | 4 | 2 | 1 | Matt demonstrates technical awareness by asking about capsule networks and two-way propagation. Domingos explains the immense sample efficiency gap between human and machine learning, illustrating how current AI requires vast data or simulated time to accomplish what humans learn rapidly. | |
| Geopolitical Risks, Authoritarianism, and Incompetent AI | 3 | 4 | 3 | 1 | Matt asks about AI existential risks versus geopolitical competition. Domingos forcefully rejects Terminator scenarios as media-driven misunderstandings, arguing instead that state authoritarianism and dumb, incompetent AI algorithms making automated decisions pose the real dangers. | |
| Q&A: Human Brain Connections vs. Machine Scale | 1 | 5 | 4 | 0 | An audience member questions comparing biological neural networks to machine learning, arguing deep learning parameter counts are trivial compared to the human brain. Domingos corrects the premise, noting modern architectures like BERT possess parameter scale comparable to many biological nervous systems. | |
| Q&A: Explainability, Accuracy, and Objective Verification | 2 | 5 | 3 | 0 | Domingos addresses a question on explainability, pointing out an unavoidable trade-off between model accuracy and interpretability. He critiques blanket regulations like EU explainability mandates and offers NP-complete verification as a framework for trusting complex AI outputs. |
Statements from this episode (14)
Pedro Domingos: Very little of current machine learning capability has been deployed
“With the existing machine learning technology, of what we can do with that, how much have we done? Very little so far. So there's enormous amount for, enormous scope to do things there. Not just in finance, but in a lot of other fields, right?”
Domingos: AI is too early to predict 20 years into the future
“I would say we are so much in the beginning that we can't even really picture where we're going to be 20 years from now.”
Pedro Domingos: Every major ML algorithm can theoretically learn any function
“On a theoretical level, every one of these major machine learning algorithms has a theorem that says if you give it enough data, it can learn any function.”
Pedro Domingos predicts Backprop alone will not reach general AI
“I actually do not believe that that's the case. I think Backprop will not get us there.”
Domingos: Major deep learning wins come from combining multiple techniques
“A lot of the things that people think of as successes of deep learning are actually successes of combining deep learning with other things.”
Domingos: The hardest problem in AI is learning representation
“The hardest problem in AI is coming up with a representation. Once you've done that, the rest follows, and when you don't do that, you don't really get very far”
Domingos: Geoffrey Hinton hasn't made capsule networks work yet despite right intuition
“I think capsule networks are a very interesting, you know I like the intuition behind capsule networks. I think they point in the right direction. I think Jeff hasn't quite been able to make them work yet.”
Domingos: Machine learning's five major paradigms haven't changed since the 1950s
“There's been enormous progress. The five major paradigms are exactly the same as they were then.”
Domingos: Terminator-style AI scenarios are not happening anytime soon
“Terminator isn't happening anytime soon. First of all, because the technology isn't there, but second of all, because there is this, I mean, like a lot of these, I think, errors that people made come from anthropomorphizing AI.”
Domingos: AI is an amazing tool for authoritarianism
“AI can be a great tool for democracy. It is also unfortunately an amazing tool for authoritarianism.”
Domingos: Dumb AI is vastly more dangerous than smart AI
“The real danger of AI is not that computers will get too smart and take over the world. The real danger is that computers are too stupid, and they've already taken over the world, right? Computers are making decisions about us the whole time, and they're dumb,…”
Domingos: State-of-the-art AI models have more connections than many animals
“Having said that, if you look at the number of connections that the state-of-the-art machine learning systems for some of these problems have, they're more than many animals. So we're actually at the point where the, you know, they have hundreds of millions or…”
Domingos: Mandating total model explainability literally makes deep learning illegal
“And at the end of the day, you know, you can't, like the European Union, mandate that every model has to be explainable, right? Because then, if you read that law literally, it makes deep learning illegal, right?”
Domingos: Future human work will shift to setting goals and verifying AI output
“So we humans, right, I can see a future where our full-time occupation is to tell the, you know, algorithms what we want to do, set the objective function, set the boundaries, and then verify the solutions all the time, continuously.”