Jan 2, 2019 · 28m · a16z
a16z Podcast | Engineering Intent
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, engineering leaders from Pinterest and Airbnb discuss how machine learning and computer vision bridge digital platforms with real-world user behavior, alongside best practices for engineering leadership and culture.
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 30.1% of the talking time here. How this is scored →
The host as informed peer 4.0 Guest teaching 3.0 Guest disagreement 1.1 The host pushing back 2.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:08 Curtis re-frames the 10x engineer premise
Curtis gently pushes back against Fan's endorsement of the 10x engineer concept, reframing high individual output as a temporary alignment of project, skill, and personal energy.
Hardest push from the host ▶ 11:42 Sonal presses on extreme review sampling biasSonal challenges Curtis's assumption about review usefulness by bringing up the statistical problem of regression to the mean when sampling purely from extreme customer sentiments.
Biggest teaching moment ▶ 15:40 Fan clarifies modern deep learning classification limitsFan educates the host on modern computer vision capabilities by pointing out that classic dog versus bagel images are easily solved, explaining that actual technical limits stem from training data domain richness.
The host holds their own ▶ 24:48 Sonal introduces VC framework on technical debtSonal demonstrates domain expertise by introducing Martin Casado's arguments regarding how isolated engineering estimates create systemic technical debt without centralized VP of Engineering alignment.
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 |
|---|---|---|---|---|---|---|
| Using Offline Reviews as Predictive Matching Signals | 4 | 3 | 1 | 2 | Sonal frames the core challenge of inferring physical user intent from implicit digital signals and shares an insightful personal anecdote about Pinterest board clustering. Curtis and Fan cooperatively explain how review signals and visual lenses serve as feedback mechanisms into ranking algorithms. | |
| Transforming Unstructured Images into Structured Data | 3 | 4 | 0 | 2 | The host asks targeted questions about moving unstructured image data into structured data pipelines. Curtis and Fan walk through object detection and multi-attribute labeling in ML models. | |
| Distinguishing Aspirational Intent from Actual Outcomes | 5 | 3 | 1 | 3 | Sonal pushes the guests on review sampling bias from extreme opinions and fear of retribution. Curtis details Airbnb's double-blind review reveal mechanism to ensure data integrity. | |
| Modeling Visual Aesthetics and Personalization Signals | 4 | 4 | 1 | 2 | Sonal brings up classic computer vision edge cases like the dog versus bagel meme. Fan clarifies that basic classification is solved while deep learning subtle boundaries remain constrained by training data richness. | |
| Engineering Culture, Mission, and Technical Stacks | 3 | 2 | 1 | 1 | The host prompts a discussion on religious engineering stack debates and open source policies. Curtis and Fan articulate pragmatic engineering management principles for standardized tech stacks. | |
| Managing Technical Debt and Infrastructure Balance | 5 | 2 | 1 | 2 | Sonal displays strong domain insight by citing a thesis from Martin Casado on technical debt accumulation and leadership trade-offs. The guests detail fix-it weeks and engineering throughput metrics. | |
| Debating the 10x Engineer and Management Alignment | 4 | 3 | 3 | 2 | The guests exchange differing views on whether the 10x engineer is an intrinsic trait or a situational state of project alignment. Sonal draws a parallel to creative editing output. |
Statements from this episode (16)
Mike Curtis: Airbnb tests guest reviews as predictive signals in search matching
“And one of the areas that we're experimenting with now is being able to use the reviews that were given on a place as a predictive signal in the matching model all the way back.”
Lee Fan: Visual discovery on Pinterest differs from Google by lacking objective right answers
“In some way, this is different from Google. There's no right or wrong.”
Mike Curtis: Airbnb explores image embeddings to dynamically re-rank home listings
“Some of the areas that we're exploring now is like, how can we, you know, find other embeddings in those images that can take the unstructured data of the image, turn it into something that can actually be tagged and labeled and then used in that ranking algor…”
Lee Fan: Products should expand discovery before driving conversion based on intent
“What we could do as a product is expand your horizon and help you to discover new, interesting ideas. We don't have to push you into deep, say, purchase this or book this. But then later, we can tell users' behavior. They're going narrow and narrow. Now we, I …”
Mike Curtis: Airbnb reveals host and guest reviews simultaneously to stop retribution
“So, so one of the things that we did was we changed it so that we have a simultaneous reveal of the reviews. So both the host and the guest have two weeks to write a review, and then they're revealed at the same time.”
Lee Fan: Pinterest trains AI to identify visual cues that make photos inspirational
“We are trained computer to learn why this image of the same living room, you take a picture of this way and that way, that looks so different. One just looks so inspirational. The other, like maybe just boring and the computer will start to learn those cues an…”
Lee Fan: Pinterest tests skin tone detection for personalized visual recommendations
“We have some experiment in-house. If you take a picture of yourself, and we learn your skin tone. But by the way, we can also learn from the pins you like, and we know there's certain style, certain shape of a model you like to see.”
Lee Fan: Computer vision progress in subjective domains is limited by data
“Right now, I will say a lot of domain is limited by the data. If you only have a limited data to teach, let's say fashion, how can we know this is a fashion that are high end and more for the runway instead of a daily? It's a lot of data because it is a subtle…”
Airbnb open-sources all internal data pipeline and analysis infrastructure
“We regularly open source our own internal technologies, particularly around data, like everything we do from managing data pipelines to how we do data analysis, we pushed out to open source.”
Mike Curtis: Airbnb holds business leaders accountable for software bugs
“The way we've been thinking about it and the way that we try to do it now is that the business leaders who are responsible for furthering the business also have goals that are associated with bugs, performance, like all the things, like systems science.”
Mike Curtis: Engineering teams cannot hire their way out of technical debt
“And like you can't hire your way out of that. Every, every person that you hire just ends up becoming less and less productive.”
Lee Fan: Individual engineer throughput can differ by 10x to 100x
“Meaning that engineer versus engineer, the throughput can be 10 X or even hundred X difference.”
Lee Fan: Engineering compensation should reflect individual productivity differences
“I also believe that we should reward according to it.”
Mike Curtis: 10x engineering output stems from alignment, not static ability
“My thinking on that is slightly different in that I think that, you know, people have periods of time where they can be, like, producing X, what their peers might be, and it's that magical moment when they're aligned with the right project, with the right skil…”
Mike Curtis: Engineering managers' main duty is enabling peak alignment moments
“This is, like, one of the most core fundamental premises behind engineering management, is, like, understanding the person. What motivates them? What are their skills? What are they trying to develop? That magic moment that I talked about, like, being able to …”
Fan: Designing with non-technical founders reveals logical gaps in engineering
“When you talk to a designer, ask questions like, wow, I never thought about this. And the process of thinking about how to answer this and discover there's some hole in my logic or discover like maybe we were to, you know, go down this path and maybe we should…”