Dec 8, 2017 · 30m · a16z
AI: What's Working, What's Not
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 comprehensive presentation, Frank Chen of Andreessen Horowitz outlines the state of artificial intelligence, exploring real-world commercial applications, current technical limitations, and actionable implementation strategies for enterprise leaders.
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 host as informed peer 0.0 Guest teaching 0.0 Guest disagreement 0.1 The host pushing back 0.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 23:15 Refuting press narratives on automation job destruction
Frank directly challenges media alarmism that AI will destroy all jobs by pointing out that automation panics have recurred for 300 years while consumer demand constantly invents new occupations.
Hardest push from the host ▶ 15:45 Dismissing Super AI existential threatsFrank dismisses popular hype around existential Super AI threats by citing Andrew Ng's quote comparing Super AI anxiety to worrying about overpopulation on Mars.
Biggest teaching moment ▶ 19:40 Exposing the limits of backpropagation in deep learningFrank educates the audience on the fundamental limits of current AI by sharing deep learning pioneer Geoffrey Hinton's admission that backpropagation is a dead end requiring a complete paradigm reboot.
The host holds their own ▶ 24:50 Counterintuitive ATM vs bank teller dataFrank uses empirical historical data on ATM deployment expanding overall bank teller jobs to dismantle oversimplified claims about technology wiping out human labor.
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 |
|---|---|---|---|---|---|---|
| Analogy: AI Will Get Inside All Software Like Databases | 0 | 0 | 0 | 0 | Frank Chen opens his solo presentation framing AI as an foundational layer that will penetrate all software applications much like databases did. Because this is a monologue presentation without host interaction, host-side metrics are zero. | |
| What's Working: Natural Language Processing & Speech Recognition | 0 | 0 | 0 | 0 | Frank outlines real-world working examples in natural language processing and computer vision, such as Pinterest image detection and speech recognition surpassing human accuracy. Monologue presentation structure maintains zero host involvement. | |
| What's Working: Predictive Analytics & Machine Learning | 0 | 0 | 0 | 0 | Frank discusses machine learning applications in Airbnb pricing, Instacart routing, and smartwatch health predictions, pointing out practical lessons around human labeling. Host metrics remain zero for this presentation segment. | |
| What's Working: Autonomous Systems | 0 | 0 | 0 | 0 | Frank details autonomous technologies like Rwanda blood delivery drones and self-driving taxis alongside infrastructure enablers like Databricks. As a monologue, host scores are strictly zero. | |
| What's Not Working Yet: Stages of AI & Current AI Limitations | 0 | 0 | 0 | 0 | Frank contrasts narrow AI with general AI, illustrating limits through fourth-grade science tests and citing Geoffrey Hinton's view that current deep learning paradigms need a reboot. Monologue segment with zero host engagement. | |
| Hot Topics: US vs. China, Intelligence Augmentation, and Automation Jobs | 0 | 0 | 1 | 0 | Frank addresses popular press narratives regarding China's national policy, AI personal assistants, and automation taking jobs, offering mild pushback against media hyperbole. Host scores remain zero. | |
| Enterprise Action Plan: Implementing AI in Your Organization | 0 | 0 | 0 | 0 | Frank delivers an enterprise roadmap covering choices to buy, bespoke build, or call APIs for internal AI adoption. Monologue format. | |
| Practical Inspiration & Conclusion: The Cucumber Sorter Case Study | 0 | 0 | 0 | 0 | Frank concludes with a case study of a $2000 Raspberry Pi cucumber sorter to inspire practical organizational application. Monologue segment. |
Statements from this episode (18)
Chen predicts AI will be embedded in all software like databases
“So fundamentally, I believe that AI is going to get into every piece of software that we write in almost exactly the same way that databases got inside all the software.”
Over 300 corporate earnings calls mentioned AI as central strategy in 2017
“We now have 300 earning calls this year where the CEO said something about artificial intelligence being central to their strategy.”
Airware uses drone data and AI to automate mine safety compliance checks
“Before AI, literally people walked out with clipboards and yardsticks to figure out if they were in compliance. Airware instead gathers data from drones and then analyzes it.”
Chen: Speech recognition AI error rates dropped below human levels
“Human word error rate is about five or six percent. If you talk to a human, they will get five or six percent of the words wrong. The algorithms are now at four percent and folly.”
Chen: YouTube auto-subtitled over one billion videos using AI
“It's the same technology that allows YouTube to subtitle their videos. Which they've done for a billion YouTube videos with no human involvement.”
Instacart achieved dual 3-4% efficiency gains from traditional ML and deep learning
“Instacart uses prediction techniques to help get shoppers through the grocery store faster. They went through this in two waves. Wave one was they used traditional machine learning techniques. Things like gradient boosted decision trees. And they got a three o…”
Cardiogram predicts sleep apnea and hypertension using standard smartwatch heartbeat data
“Using that exact same data, they were able to predict whether you had sleep apnea. They were able to predict hypertension. And these things, traditionally, you'd have to have dedicated medical devices for. But just off the heartbeat data, we can predict whethe…”
Zipline performs 500 autonomous drone blood deliveries daily in Rwanda
“Our portfolio company Zipline is delivering blood where it's needed throughout Western Rwanda. There's one place they collect blood in Western Rwanda. There are tons of places they need it, but sometimes it's dangerous or time consuming to get to the place tha…”
Udacity Spinout Voyage Launched Autonomous Taxi Service in Nine Months
“They spun out of Udacity, created this company called Voyage. Nine months after they spun out, they turned on a self-driving taxi service in a retirement community in San Jose.”
Andrew Ng: Worrying about Super AI is like worrying about Mars overpopulation
“Andrew Ng, who is one of the foundation researchers in this area, said, I worry about super AI in exactly the same way that I worry about overpopulation on Mars.”
Chen: AI field lacks a consistent research agenda to reach AGI
“We don't really have a consistent research agenda that will get us to general AI, and as a result, I'm not too worried about the super AI, quite yet.”
Geoffrey Hinton: Backpropagation is a dead end for reaching general intelligence
“He says, I think we need a reboot. I think the path that I set everybody on, and fundamentally an algorithm called backpropagation, I think that's a dead end. I think if we want general intelligence, we're going to need to invent a new technique. We're going t…”
Frank Chen: China leads deep learning paper volume, but West leads citations
“But also if you look today, the number of deep learning papers in China has outnumbered The number of deep learning papers, but the number of citations in those papers still tilts towards the West. The most influential papers are Western papers.”
Frank Chen predicts hundreds of billion-dollar AI augmentation startups will emerge
“And I'm convinced over the next 2030 years, dozens if not hundreds of billion dollar companies will get created building Jiminy Cricket.”
Frank Chen: Automation won't cause joblessness due to infinite human wants
“I think so long as we have the capacity to want new things, we will always be creating jobs to satisfy those demands.”
Chen: Existing AI research enables decades of enterprise software development
“Even if we stopped the research pipeline today and said no more research is coming out of the universities, I think we have 20, 3040 years of software development to do to put all the AI inside the apps.”
Chen: Enterprise buyers should reject software vendors lacking AI roadmaps
“If you're evaluating, ah, software from a vendor, and they don't have a pretty sophisticated machine learning roadmap that says, Here's how my product gets smarter, makes better predictions, helps you save money. You should send them home. Because their compet…”