Nov 20, 2017 · 22m · mad
Where Should Machines Go to Learn? // Auren Hoffman, SafeGraph (FirstMark's Data Driven)
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 a keynote presentation and Q&A session at DataDrivenNYC, SafeGraph CEO Auren Hoffman argues that access to high-quality data—rather than algorithmic innovation—is the primary bottleneck for artificial intelligence, advocating for democratized world truth sets and secure data infrastructure to fuel global innovation.
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 8.4% of the talking time here. How this is scored →
Matt as informed peer 1.7 Guest teaching 3.7 Guest disagreement 1.4 Matt pushing back 1.1
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 13:30 Rejecting the definition of data companies
Auren dismisses conventional tech framing, asserting that most self-proclaimed data companies are merely application wrappers around hidden proprietary sets.
Hardest push from Matt ▶ 20:21 Challenging optimistic data access with China comparisonMatt presses Auren on whether authoritarian state data practices in China undermine his thesis regarding open democratized data driving global AI breakthroughs.
Biggest teaching moment ▶ 7:10 The reality of open-source algorithmsAuren educates the room on big tech dynamics, illustrating how Google can freely open-source TensorFlow because owning data confers the true monopoly power.
Matt holds his own ▶ 16:27 Host raising structural regulatory frictionMatt demonstrates domain expertise by highlighting specific legal constraints such as GDPR and index-level privacy rights that limit historical data aggregation.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Case Study: Data Beats Algorithms | 0 | 3 | 1 | 0 | This is a solo keynote segment where the guest details Microsoft Research data studies and oncology data acquisition challenges. The host does not speak during this segment. | |
| Startup Challenge #3: Data Munging and Cleaning | 0 | 3 | 1 | 0 | The guest continues his monologue on data cleaning, explaining how engineers spend 95% of their time on data munging rather than building models. Host scores remain zero for this uninterrupted presentation. | |
| Two Visions for the Future of AI | 0 | 4 | 2 | 0 | Auren outlines two potential futures for AI data monopolies and recounts a story about Google open-sourcing TensorFlow. The host does not participate in this monologue segment. | |
| Brute Force vs. Innovation & Keynote Conclusion | 0 | 3 | 1 | 0 | The guest concludes his keynote talk advocating for open data platforms over trade secrets to drive innovation. The host remains silent throughout the segment. | |
| Q&A: Internal Data Tools vs. World Truth Sets | 1 | 4 | 1 | 1 | Matt Turck opens the Q&A section with light banter, allowing Auren to explain the transition from internal data analytics to broad world truth sets. | |
| Q&A: Predicting Where AI Breakthroughs Will Occur | 5 | 4 | 2 | 3 | Matt Turck interjects to ask about data privacy frameworks and GDPR compliance when building historical indexes. Auren explains privacy-as-a-service concepts and containerized algorithms. | |
| Q&A: Strategic Investor Strategy Across Verticals | 6 | 5 | 2 | 4 | Matt challenges Auren's data access assumptions by bringing up China's lack of stringent privacy rules and potential global AI advantage. Auren agrees and expands on regulatory arbitrage. |
Statements from this episode (18)
Hoffman: Great data usually beats great algorithms in machine learning
“Great data usually beats great algorithms.”
Hoffman: Machine learning engineers spend up to 99% of time organizing data
“So now, you know, you have these great machine learning engineers, and they thought they were going to be spending You know, all their time predicting the future, but it turns out they're spending 95 to 99% of their time organizing the past.”
Hoffman: A data breach will put a data company out of business
“And of course, keeping that data secure is, is a huge issue, and you're going to build a security team, run penetration tests, you know, or of course you're going to be out of business if you have some sort of breach.”
Hoffman: Google ML engineers spend 95% of their time building models
“As a machine learning engineer, now you can spend 95% of your time predicting the future, which is what you want to do as a machine learning engineer.”
Hoffman: Data overhead is the core struggle for non-big-tech AI companies
“So this is the course, the core struggle for almost every single company, except for maybe Google, Facebook, Amazon, Tencent, that has to deal with all these types of issues.”
Hoffman: TensorFlow is one of the top 10 innovations of the past decade
“In my opinion, it's one of the top 10 innovations in, in the last 10 years.”
Hoffman: Advanced AI models cannot overcome exclusive access to core data
“In that world, even the most advanced models, deep learning, and machine learning frameworks can't beat them because they have access to this kind of core underlying data.”
Hoffman: Compute power access prices have dropped every single month
“And the price, because of things like containers, et cetera, the price of access and compute power has gone down every single month.”
Hoffman: Internal data represents under 0.01% of global data for most companies
“Most companies out there, their own data represents, like, point oh one percent of the world. It's, unless you're Google, Facebook, Amazon, Tencent, a couple of others, you have a very small sliver of what's happening in the world.”
Hoffman: Google has location data on 70% of US mobile phones
“Companies like Google have great access to this data because they have 70% of the phones in the U.S. That they have data on, and probably even a higher percentage of phones worldwide that they get to see all this great location data on.”
Hoffman: Most so-called data companies are actually application companies
“There are very few data companies out there. Most companies that are quote-unquote data companies are actually application companies.”
Hoffman: Datasets are significantly more valuable when they are temporal
“The other thing about data is that it's more valuable if it's temporal.”
Hoffman: Future technological innovation will occur where data exists today
“If you really wanna know, like, where is innovation going to happen in the future, I think you can really look at innovation as to where we have data today.”
Hoffman: Oncology will see huge innovation in the next 20 years
“So the probably oncology is a place where we'll probably will see some innovation because there's a defined data set. There's, there are definitely some vendors today that have access to really good oncology data. So I expect that we'll see a lot of really gre…”
Hoffman: Nutrition tech will see little progress over the next 20 years
“On the flip side, if you think about nutrition, likely, 20 years from now, we'll still have fad diets. Just be, it's just incredibly difficult to collect all the data of, you know, everything that goes into your body, you know, your eventual outcome, which cou…”
Hoffman: Privacy-preserving data querying solutions will emerge by 2022
“But I expect within the next five years, we'll probably have answers to some of those problems.”
Hoffman: China may lead in healthcare machine learning due to data regulations
“And so we could see, it's very possible we could see more machine learning innovations, or at least in certain areas, Like maybe in healthcare, for instance. We might see more machine learning innovations that happen in China than happen in, in other places.”