Jan 25, 2016 · 22m · mad
Can A.I. Become More Human? // Gary Marcus, Geometric Intelligence (Hosted by FirstMark Capital)
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 DataDrivenNYC keynote presentation, cognitive scientist Gary Marcus critiques the overhyped state of deep learning and demonstrates how integrating human cognitive development principles can overcome data scarcity, statistical brittle points, and narrow domain limits to achieve true Artificial General Intelligence.
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 1.5% of the talking time here. How this is scored →
Matt as informed peer 0.1 Guest teaching 2.6 Guest disagreement 2.1 Matt pushing back 0.1
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 2:45 Calling current AI systems 'idiot savants'
Marcus forcefully pushes back against industry hype by describing specialized driverless cars and translation systems as fundamentally limited 'idiot savants'.
Hardest push from Matt ▶ 19:06 Host mic interventionHost Matt Turck briefly halts the session flow to insist the audience member speak directly into the microphone.
Biggest teaching moment ▶ 19:38 Educating on symbolic AI versus statistical modelsMarcus provides a masterclass on historical AI paradigms, explaining why symbolic manipulation must be combined with modern statistical learning to achieve real common sense reasoning.
Matt holds his own ▶ 20:32 Host joke during interruptionHost Matt Turck playfully asserts administrative control over time when a ringtone interrupts the guest, stating 'It's not like the Oscars'.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| DataDrivenNYC Event Title and Speaker Introductions | 0 | 1 | 2 | 0 | Gary Marcus delivers an opening keynote monologue introducing the historical gap between narrow AI progress and general AI. Because this is a monologue presentation, host activity scores are zero. | |
| Historical AI Winters and Modern Narrow 'Idiot Savants' | 0 | 3 | 3 | 0 | Marcus critiques modern AI tools, labeling current chess engines and translation software as 'idiot savants' incapable of broad reasoning. The host does not speak in this presentation segment. | |
| Diagnosing AI Bottlenecks: Statistics, Big Data, and Hallucinations | 0 | 3 | 3 | 0 | Marcus diagnoses AI's reliance on big data correlations and statistics, demonstrating translation failures like 'word salad' and neural net 'hallucinations.' Host score is zero during the monologue. | |
| The Long Tail Problem and Contextual Computer Vision Failures | 0 | 3 | 2 | 0 | Marcus highlights computer vision failure modes in low-frequency long-tail scenarios, comparing current AI errors to cognitive hallucinations. Host involvement is absent in this keynote segment. | |
| Overview of Geometric Intelligence and Cognitive Approach | 0 | 2 | 1 | 0 | Marcus explains Geometric Intelligence's approach to learning efficiently from sparse data, contrasting human cognitive development with expensive annotated datasets. | |
| Physical Cost Bottlenecks of Reinforcement Learning in Robotics | 0 | 2 | 2 | 0 | Marcus explains why reinforcement learning fails in physical robotics due to catastrophic trial-and-error costs, comparing machine learning inefficiency to his toddler's rapid learning. | |
| Audience Question & Answer on Singularity and Hybrid AI Architecture | 1 | 4 | 2 | 1 | Host Matt Turck moderates an audience Q&A session where Marcus addresses questions on the singularity and symbolic versus statistical AI. Turck manages session logistics and timing without technical pushback. |
Statements from this episode (20)
Marcus: Siri's technology is not fundamentally different from ELIZA
“But the technology for Siri is not really that different from the technology for ELISA, and Siri is still very far from artificial general intelligence.”
Marcus: In AI, we wanted Rosie the Robot but got Roomba
“We wanted Rosie the robot, and instead we got Roomba.”
Marcus: Autonomous driverless cars can drive in Palo Alto but not Manhattan
“In fact, they can drive in Palo Alto, but they can't drive in Manhattan.”
Marcus: Current AI systems are fundamentally a collection of idiot savants
“AI is still basically a collection of idiot savants.”
Marcus: AI has made little progress toward strategic general game players
“There's still been very little progress in building a general game player that can play games that require strategy or insight, complex three-dimensional graphics.”
Marcus: AI is not close to achieving strong, human-level intelligence
“But we're not there yet. We're not even close.”
Google's Peter Norvig estimated strong AI is worth up to $2T annually
“I asked Peter Norvik, who at the time was director of research at Google I said, you know, how much money would it really be worth if you could have strong AI? And he wrote back about five minutes later, I'm suddenly a much louder, I think. And estimated somet…”
Marcus: AI is stuck due to over-reliance on statistics and big data
“And I have a three-part diagnosis, which is that it's fallen in love with statistics, it's fallen in love with big data, and it's forgotten its roots.”
Marcus: Deep learning models struggle with low-frequency edge cases
“People get very excited every time there's a new deep learning result. But it's always the case of doing better on the high frequency data than the low frequency data.”
Marcus: Deep learning vision errors are not human-like and resemble hallucinations
“So when you start to look at the errors that these things make, they're not human-like errors, and again, they seem almost like hallucinations.”
Marcus: AI researchers should look to child cognitive development for AGI clues
“So, where I think we should start looking for clues, in a nutshell, is children.”
Marcus: Information retrieval fails at state updates that human-level AI requires
“If you have human-level intelligence, you can make sense of that, but if you just memorize stuff off the web and do information retrieval, you could actually get stuck there.”
Gary Marcus: No great machine learning technique exists for natural language
“And so, as a result, there is no great machine learning technique in natural language.”
Gary Marcus: Financial markets are fundamentally a sparse data problem
“Finance is really fundamentally a sparse data problem because The financial markets aren't stationary.”
Marcus: DeepMind's deep RL required millions of data points per game
“This is what DeepMind used in their systems, and they used a version called deep reinforcement learning. And what they did is they collected billions or, you know, probably millions of data points for each game.”
Marcus: Reinforcement learning fails in physical robotics due to real-world damage costs
“But it's another thing if you import those same techniques into robots. If you have your robots doing reinforcement learning, which is basically trial and error learning, they start knocking over the furniture a 100,000 times, you're probably gonna send it bac…”
Marcus: Facebook M relies mostly on human operators rather than AI
“Facebook's new M service, which they have not rolled out at scale has humans on the back end. There's a little bit of AI in there, but it's mostly, ah, human beings, which is why they haven't rolled it out for a billion customers. They don't have enough human …”
Marcus: AI progress will be gradual, not a sudden explosion
“I don't think there's going to be like one special day with this big explosion. They're going to get better and better. It's going to take a while.”