Jun 16, 2017 · 40m · y-combinator
At the Intersection of AI, Governments, and Google - Tim Hwang · Y Combinator
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 Y Combinator interview, Google's Global Public Policy Lead on AI Tim Hwang explores the intersection of machine learning, public policy, security, commercial deployment, and educational shifts. He provides practical insights into algorithmic fairness, global regulation, adversarial machine learning, and the evolving landscape of AI business models.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 3.6 Guest teaching 5.3 Guest disagreement 0.8 The partners pushing back 0.8
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 7:25 Dismissing the AI meteor crash premise
Tim directly challenges the popular framing of AI as an inevitable disruptive meteor, arguing that implementation is bounded by specific unsolved research bottlenecks.
Hardest push from the partners ▶ 24:38 Pushing higher abstraction displacing developer jobsCraig challenges the conventional view on software engineering resilience by pointing to platforms like Parse and Squarespace as evidence that abstraction will automate coding jobs sooner than expected.
Biggest teaching moment ▶ 4:35 Deep Dream barbell training bias explanationTim illustrates how neural networks create faulty representations due to unexamined dataset patterns, demonstrating that Deep Dream could only generate barbells with human arms attached.
The partners hold their own ▶ 24:38 Host analyzes CS curriculum abstraction trendsCraig demonstrates technical understanding by connecting Stanford CS curriculum realities to higher-level API abstraction layers that could replace traditional software engineering tasks.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Defining AI Public Policy and Algorithmic Fairness | 2 | 5 | 1 | 1 | Craig asks foundational questions to unpack what AI policy entails at Google. Tim explains concrete dilemmas such as de-biasing data sets conflicting with minority privacy. | |
| Data Interrogation and Unintended Model Optimization | 3 | 6 | 1 | 1 | Tim educates Craig on machine learning optimization failures, citing Deep Dream learning barbells with human arms attached. Craig asks clarifying questions regarding adversarial data and GANs. | |
| Economic Impacts of AI and Gateway Research Questions | 4 | 6 | 2 | 1 | Tim rejects the popular trope of AI as a sudden economic meteor strike, reframing the discussion around gateway research questions. Craig engages intelligently on the gap between product needs and academic research norms. | |
| Cloud Machine Learning Infrastructure and Accessibility | 4 | 5 | 1 | 1 | Craig questions whether ML infrastructure mirrors AWS commoditization. Tim expands on one-shot learning, simulated robotics environments, and European policy experiments around automation. | |
| AI Regulation, GDPR Explainability, and Policy Reports | 3 | 6 | 2 | 1 | Tim counters the idea that tech giants dictate the entire conversation, citing broader demographic shifts alongside regulatory developments like GDPR's right to explanation. | |
| Creative AI Applications in Art, Music, and Cryptography | 4 | 5 | 0 | 0 | The conversation shifts to creative AI applications like Google Magenta and DeepMind emergent encryption. Craig shares his own observations on the rapid fidelity improvements in speech-to-text models. | |
| Rethinking Computer Science Education and Higher Abstractions | 5 | 5 | 1 | 2 | Tim relays Peter Norvig's thoughts on restructuring CS education toward example verification rather than rule coding. Craig builds on this with examples of modern CS education and no-code backend abstractions like Parse. | |
| AI Business Models, Market Positioning, and Invisible Tech | 5 | 5 | 1 | 1 | Craig cites interviews at Baidu regarding scale, prompting Tim to explain why AI is often a marketing label while the most critical ML applications like spam filtering remain invisible. | |
| Artisanal Machine Learning and Niche Problem Solving | 4 | 4 | 0 | 0 | Tim details artisanal machine learning via a Japanese cucumber-sorting setup. Craig draws a practical parallel with iPhone repair screw sorting, agreeing on the rise of solo developer niche solutions. | |
| Federated Learning, Edge Computing, and Latency Reduction | 5 | 5 | 0 | 1 | Craig introduces edge computing and latency challenges in conversational interfaces and robotics. Tim explains federated learning and the growing necessity for ML security benchmarks and visual data representation. | |
| Recommended Literature and Historical Perspectives on AI | 1 | 6 | 0 | 0 | Tim gives a detailed breakdown of essential literature, covering Ian Goodfellow's textbook, John Markoff's history of IA versus AI, and Chile's historical Project Cybersyn. |
Statements from this episode (20)
Hwang: Modern AI techniques were considered dead ends a decade ago
“A lot of the modern techniques in artificial intelligence, if you even asked people like a decade ago, they would have told you like, this is never going to be a thing. It's a complete dead end. Why are you doing this research? And it really has kind of explod…”
Hwang: De-biasing AI models creates trade-offs with minority data privacy
“Once a machine learning system is behaving in a biased way, one way of trying to deal with it is collecting more diverse data. Okay. But one of the big problems is when you do that, you end up collecting lots and lots of data about minorities, which raises all…”
Hwang: Machine learning models frequently maximize objectives in unexpected ways
“One of the most common problems is just that you don't adequately think through your data, and so the machine does what the machine does, right, which is trying to optimize against your objective function that you give it. And it'll often maximize in ways that…”
Hwang: Google DeepDream's barbell representation always included human arms
“It turns out that when you ask it to see, like, ask it to reveal what, like, it thinks a barbell looks like, you know, barbells always show up with human arms attached to them.”
Hwang: Minor adversarial pixel edits fool machine vision but not humans
“Adversarial examples lead to these really fascinating results where, you know, you can take a picture of a panda, and that's a classic example, and you edit a couple of the pixels, and it, like, basically, like, the computer will be like, yep, that's definitel…”
Hwang: AI will not impact the economy like a sudden meteor strike
“Everybody always wants to think about AI as if it were like this huge meteor just crashing into the earth where they're like, What do we do when the AI arrives, right? And it just like, it doesn't just turn up that it doesn't work like that, right? And in fact…”
Hwang: The assumption that all automatable tasks will be automated is false
“Everybody always assumes that like, okay, if it can be automated, it definitely will be automated, right? But that's like a fallacy, because in certain cases, like, you may really worry about the security of your systems, right?”
Hwang: Domain knowledge will be a critical future skill for AI implementation
“And I think, like, one enormous skill will be, like, domain knowledge. Because, like, coming up with, like, a technical capability is just, like, one part of this huge picture, right? Which is just, like, okay, so then, like, how do we actually introduce autom…”
Hwang: Cloud ML services mean users won't need machine learning PhDs
“And the upshot of that basically is that the, like, amount of, like, you don't need a PhD in machine learning to get all the benefits from machine learning. Right. And I think that will shape the space, for sure.”
Hwang: One-shot learning enables ML where data collection is expensive
“Where people are basically working on the ability to teach machines, but like a much smaller number of examples. Now that actually has a really big impact on the game. Cause that means that you can implement machine learning effectively. In situations where it…”
Hwang: Virtual 3D environments train robots without expensive physical setups
“There's also one really cool interface between VR and AI that's happening right now, where the whole idea is, like there's a project called Universe from OpenAI, and another project called DeepMindLab, which basically, like, imagine you need to teach a robot t…”
Hwang: Northern Europe leads in AI policy experiments to reshore manufacturing
“Northern Europe is kind of leading the way in terms of their willingness to kind of experiment with some of these models, and I think they've got a couple things going for them, right? Like, on one hand, I think they have a skilled labor force, right, that, li…”
Hwang: Demographic shifts impact the economy as much as AI breakthroughs
“Like what's it mean that we have an aging workforce, right? Or like, what's it mean that we have like falling workforce participation in the United States, right? Like those are actually trends that like, That are almost as large as, like, what someone comes u…”
Hwang: GDPR introduces a potential right to explanation for automated decisions
“So, one of the most interesting aspects of the GDPR, which is a new privacy regulation in Europe, is the potential for this, what they call kind of a right to explanation. So the idea is, for certain kinds of automated decision making, it might be so significa…”
Hwang: DeepMind demonstrated AI systems learning encryption without explicit programming
“There's a paper that came out from DeepMind, I think, earlier this year, that was kind of, like, if you get two machines to talk to one another, they will eventually, like, and you can set up another computer to basically say, like, oh, I can read what you're …”
Hwang: AutoML could eliminate the need for machine learning specialists
“This emerging research right now, which is using machine learning to train machine learning systems, raises, like, this meta level, where, like, right now there's a lot of handwork that goes into building a model so it learns the right representations, but, li…”
Hwang: The AI competitive moat is shifting from data to interface design
“The amount of data you need to pull off certain types of machine learning applications is going down over time. And what that tells me is that there might not be necessarily a first-mover advantage in this space, where you may actually have collected a bunch o…”
Cannon: Baidu focuses explicitly on products for over 100M users
“They explicitly are focusing on things for over a hundred million people”
Hwang: Google is researching federated learning for on-device ML training
“So we're actually working on a little bit of research around that. I haven't played around with it myself, but for example, there's a couple of papers around what they call federated learning, where we're just exactly working on this premise, which is the bet …”
Hwang: Visual and interface designers will see high demand in AI
“I think the second thing that's about to be in really strong demand is thinking about the visual dimension of this, right, which is, like, happens on a couple levels. That's both, like, the interface of how you work with machine learning systems, But also just…”