Jan 2, 2019 · 40m · a16z
a16z Podcast | Revenge of the Algorithms (Over Data)... Go! No?
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 a16z Podcast, hosts Sonal Chokshi, Frank Chen, and Steven Sinofsky analyze DeepMind's AlphaGo Zero research paper to separate artificial intelligence media hype from practical engineering reality. They discuss the shift from data-heavy training to algorithmic reinforcement learning, the challenges of applying constrained AI models to complex real-world environments, and practical considerations for enterprise deployment and startup founders.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The host holds 21.5% of the talking time here. How this is scored →
The host as informed peer 5.2 Guest teaching 3.3 Guest disagreement 1.9 The host pushing back 3.4
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 17:44 Steven forcefully dismisses pure unsupervised hype
Steven forcefully critiques the hype surrounding modern AI trends, arguing that claims of replacing all previous technology are repeated historical fallacies and pointing out that even state-of-the-art NLP still relies on 1970s techniques.
Hardest push from the host ▶ 9:10 Sonal explicitly pushes back against guest skepticismSonal explicitly interrupts to challenge the guests' dismissive stance on generalizability, using the biological evolution analogy to argue that accelerated trial-and-error selection could lead to broader AI capabilities.
Biggest teaching moment ▶ 31:28 Frank clarifies the true origin of algorithmic biasFrank educates the host and listeners by reframing algorithmic bias from an inherent architectural flaw to an issue of human error in dataset selection and incomplete labeling.
The host holds their own ▶ 25:00 Sonal counters using Frank's own past self-driving car argumentSonal demonstrates strong knowledge and continuity by holding Frank accountable to his own past arguments regarding human black-box minds in autonomous driving to question strict machine transparency demands.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| AI Winter Fears and the Limits of Games | 4 | 4 | 2 | 4 | The guests caution against AI winter hype around claims of achieving AGI through board games like Go. Sonal interjects to note that the paper's authors themselves do not claim generalized intelligence, showing familiarity with the material. | |
| AlphaGo Zero's Breakthrough in Reinforcement Learning | 5 | 3 | 1 | 2 | Frank and Steven detail AlphaGo Zero's reinforcement learning mechanism. Sonal keeps pace by reciting technical facts from the paper, including hardware efficiency gains down to 4 TPUs and training time reduced to three days. | |
| The Evolution Analogy and Domain Constraints | 6 | 3 | 3 | 7 | Sonal explicitly pushes back against the guests' skepticism regarding generalized intelligence, introducing an analogy to evolutionary trial-and-error. Frank agrees with the evolution concept but reasserts domain constraint limits. | |
| Practical Applications for Constraint-Based AI | 4 | 4 | 1 | 2 | The conversation turns to practical applications like sales forecasting and cybersecurity code linting. Sonal helps guide the focus while the guests detail how fixed rules enable constraint-based AI. | |
| Hybrid AI Architecture and Revenge of the Algorithms | 6 | 3 | 2 | 3 | Sonal draws upon past podcast discussions and NLP history to discuss hybrid AI architectures. Steven emphasizes that practical breakthroughs combine old symbolic/rule techniques with new machine learning models. | |
| Tech Evolution Curves and Enterprise Software Debugging | 5 | 3 | 2 | 2 | Steven compares AI curves to past technology cycles like search engines and spell checkers, transitioning into enterprise debugging requirements. Sonal contributes historical context about Google being the 15th search engine. | |
| Black Box Transparency and Human Expectations | 7 | 3 | 2 | 6 | Sonal challenges Frank by citing his own past argument that human cognition is also an un-interrogatable black box, and later cites Stanford research by Clifford Nass and Byron Reeves to explain human trust in machines. | |
| Tabula Rasa, Algorithmic Bias, and Labeled Data | 5 | 5 | 3 | 3 | Sonal raises a philosophical query about tabula rasa algorithms, which Frank playfully deflects as above his pay grade before correcting the common misconception about algorithmic bias—explaining that bias stems from human dataset selection and labeling. | |
| Key Takeaways and Strategic Advice for Founders | 5 | 2 | 1 | 2 | Sonal summarizes the core takeaways around simple architectures and rule constraints. Frank and Steven offer strategic advice for founders on choosing appropriate technical paths rather than chasing trendiness. |
Statements from this episode (15)
Chen: Board game AI environments fail to mirror real-world fog of war
“All the rules are completely well known. The state of play is completely well known, right? In the real world, mostly we live in a fog of war situation where we know some things and we don't know other things. It would cost us something to go figure it out, an…”
Chen: AI startups spend $1M to $50M acquiring annotated training data
“And if you think about sort of the approach that most startups take to artificial intelligence today, they basically take the supervised learning approach. And step one, raise money so that you can go get a data set that's annotated, train your neural network,…”
Chen: AlphaGo Zero used 4 TPUs and 3 days, crushing prior versions
“It was 48 TPUs versus four. It was three days versus 40. It was thirty million trained games versus 4.9. So Order of magnitude improvement on all of those dimensions.”
Sinofsky: Nobody possesses codified rules for predicting protein folding
“Well, we know there are amino acids and we know that they have to be in three dimensions, but actually nobody else knows that there's no codified rules. Like nobody has the rules of protein folding.”
Chen: AlphaGo Zero discovered 1,000 years of human Go strategies in days
“Thousands of generations of games playing each other sort of arrive at places that maybe humans would have gotten to if we played another thousand years, but like, you know, it figured out in three days.”
Sinofsky: Code with nested parentheses is more likely to contain bugs
“Something with three sets of parentheses in it, it's likely to have a bug in it just because it has three sets of parentheses in it.”
Sinofsky: State-of-the-art machine translation still relies on 1970s NLP
“The best example for me of that is how everybody said machine learning was going to replace all of natural language processing. But if you dig into any of the work that's been going on, even the most state of the art translation, which, you know, goes any lang…”
Chokshi: Google was approximately the 15th search engine to launch
“Google was like the 15th search company to come around before it hit success, and that is kind of relevant to think about.”
Sinofsky: Algorithms are rarely a sustainable moat because competitors reverse-engineer them
“Very rarely is like an algorithm like this secret sauce for a company. Because you could look at what goes on from the outside and pretty much reverse engineer an algorithm.”
Sinofsky: Bing's biggest barrier to fighting Google was crawling the internet fast
“It turns out that was what the barrier to entering the search market, even for Microsoft and Bing, Was like, sucking in the entire internet fast enough.”
Chen: Regulators prohibit black-box AI models in lending without explainability
“The super active area of research right now, right, which is how do I make the deep learning models more transparent so that I can debug them, I can verify them, I can make sure there's no systematic bias in them, right? Because until that, you couldn't do imp…”
Sinofsky: Humans naturally ascribe excessive authority to automated computing outputs
“There's something about a mechanical device That produces answers that makes the human brain ascribe way more authority to it than there necessarily should be.”
Sinofsky: Microsoft could not formally prove to Boeing that Excel actually worked
“I couldn't go to people at Boeing or people at the Navy or wherever in Wall Street and prove that Excel worked.”
Chen: Algorithmic bias comes from incomplete datasets, not neural network architecture
“When people talk about bias in algorithms, they're mostly talking about this phenomenon, which is the human researcher or the human programmer selected an incomplete data set, and therefore you got biased results, as opposed to somehow the architecture of the …”
Chen: AlphaGo Zero proves reinforcement learning succeeds without labeled data
“Right, where there's so much momentum right now that says, basically, we're one labeled data set away from glory. Right, and this result basically shows you, wow, there's a lot of mileage that you can get out of reinforcement learning where there's no data, no…”