Dec 11, 2018 · 29m · mad
The State of AI & What's Next // Dileep George, Vicarious AI (FirstMark's Data Driven NYC)
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
At FirstMark's Data Driven NYC event, Vicarious AI co-founder Dileep George presents a neocortex-inspired vision for Artificial General Intelligence, contrasting probabilistic graphical models with brittle deep learning systems and demonstrating their application in real-world industrial robotics.
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 3.9% of the talking time here. How this is scored →
Matt as informed peer 1.4 Guest teaching 5.3 Guest disagreement 1.4 Matt pushing back 0.7
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 17:50 Dismissing game-playing AI achievements
Dileep forcefully dismisses hype around deep reinforcement learning game benchmarks, asserting that playing games does not solve real-world model-building problems.
Hardest push from Matt ▶ 16:46 Challenging AI market hype vs academic realityMatt Turck directly challenges the industry hype narrative by contrasting the excitement around deep learning since 2012 with academic sentiment that AI progress should calm down.
Biggest teaching moment ▶ 24:16 Lecture on No Free Lunch theoremDileep thoroughly educates the audience on machine learning theory, explaining why general intelligence requires a Goldilocks set of domain assumptions.
Matt holds his own ▶ 16:46 Framing post-2012 hype landscapeMatt demonstrates clear domain insight by accurately placing Dileep's presentation against the historical timeline of post-2012 deep learning hype and academic skepticism.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| The Limits of Current AI and the Old Brain Analogy | 0 | 5 | 1 | 0 | Dileep George gives an opening presentation illustrating how current AI acts like an old brain retina-reflex system. Because this is a monologue presentation, the host does not speak, resulting in host scores of zero. | |
| Towards General AI: Neocortex, Common Sense, and Generative Models | 0 | 6 | 2 | 0 | Dileep presents neocortex-inspired generative models, explaining common sense through mental simulation examples like hammering nails and Vicarious's captcha-solving capability. As a monologue segment, host scores remain zero. | |
| AI Applications in Modern Manufacturing and Industrial Automation | 2 | 6 | 1 | 1 | Dileep concludes his talk on industrial automation and Matt Turck prompts him to explain the underlying mechanics standardly without getting overly technical. Dileep educates the audience on function approximation versus probabilistic graphical models. | |
| Evaluating AI Hype and the Reality of Game Playing Benchmarks | 5 | 5 | 2 | 3 | Matt Turck cites the post-2012 deep learning hype cycle and academic calls to calm down, pushing Dileep on market reality. Dileep explains why game-playing benchmark enthusiasm was premature. | |
| The Frontier of AI Research and Machine Common Sense | 3 | 4 | 2 | 1 | Matt asks about interesting developments on the AI frontier, leading Dileep to reference DARPA's common sense initiative before transitioning to audience Q&A regarding corporate logic analogies. | |
| Audience QA: Generalizing Physics Models Across Games and Environments | 0 | 6 | 1 | 0 | Matt moderates audience questions while Dileep explains the No Free Lunch theorem and how Goldilocks assumptions enable generalization. Host scores are zero as Matt only acts as an audience moderator. | |
| Audience QA: Computational Costs and Hardware Realities | 0 | 5 | 1 | 0 | Dileep answers an audience question on computational complexity, clarifying how message passing avoids expensive sampling in probabilistic graphical models before Matt closes the session. |
Statements from this episode (15)
Cheap hardware exists in 2018, but general-purpose robotics does not
“Ok, so it's 2018, and we have cheap motors, and sensors, and processors, and hardware et cetera, but we don't have Rosie the Robot.”
The delay in creating multipurpose robots is software, not hardware
“And we don't have Rosie the robot not because we haven't, you know, we haven't solved the hardware problem. It's a software problem.”
Intelligence is the ability to model the world and act upon it
“Intelligence is the ability to model the world and to act purposefully on it.”
Deep neural networks are easily fooled by abstract adversarial visual patterns
“In fact even our sophisticated deep neural networks can be fooled very easily by showing creating these weird-looking patterns, and those patterns will get interpreted with very high confidence as, you know, things like starfish, freight car, remote control et…”
Vicarious AI broke text-based CAPTCHAs with a generative visual perception model
“So this is something that we published in science in like, you know, last year, almost a year back. It's a generative model for visual perception, so it's generative, so it can imagine things trains with very little data and you know, it can do inference in th…”
DeepMind's reinforcement learning fails on minor visual changes while Vicarious succeeds
“So if you use, ah, if you compare, ah, Vicarious, ah, system with, ah, DeepMind system, ah, for example, if you change the brightness of the screen DeepMind system will stop playing, ah, because it will, ah, get confused, ah, and if you offset the paddle we ou…”
Vicarious AI has begun pilot testing its robots in real factories
“In fact, we have just started pilot testing some of our robots in real factories.”
Generative world models enable AI to answer unexpected, untrained questions
“Building models allow you. You are able to answer questions that you did not think about during training the system.”
Deep reinforcement learning in games does not translate to real-world AI
“But those excitements Turned, turned out to be premature because it's, yeah, game playing is an easier problem compared to real world problems. Those techniques do not translate to real world problems.”
AI practitioners are shifting toward models inspired by neuroscience and cognitive science
“The mood is starting to shift in the field especially the practitioners that are close to the problem. They understand that, okay, the current wave of excitement has its limitations, and we need to look more deeply at the problems and import insights from neur…”
Corporations are currently more sophisticated in their decision-making than AI
“I think corporations are more sophisticated than today's AI.”
Embedding assumptions about object interactions enables AI models to generalize better
“Our, what we are doing differently compared to what was attempted previously was that we are putting more assumptions into our model in the sense of that there are objects in the world. There are interactions between objects, et cetera. Those assumptions are f…”
Deep learning systems outperform humans at classifying QR codes
“If I show you QR codes and ask you to classify QR codes, humans will completely fail, whereas deep learning systems will succeed.”