May 7, 2025 · 52m · big-technology
Are We at the End of Ai Progress? — With Gary Marcus
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
Tech journalist Alex Kantrowitz interviews AI researcher and author Gary Marcus on the diminishing returns of deep learning scaling, emerging financial valuation bubbles, real-world safety hazards, and the necessity of transitioning toward neurosymbolic AI architectures.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 27.7% of the talking time here. How this is scored →
Alex as informed peer 5.6 Guest teaching 5.1 Guest disagreement 5.6 Alex pushing back 5.1
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:15 Marcus forcefully rejects the premise that he has been wrong
Marcus immediately counters Kantrowitz's challenge by demanding specific written proof of any failed predictions, recounting public arguments with Tyler Cowen and betting challenges to Elon Musk.
Hardest push from Alex ▶ 16:16 Kantrowitz challenges Marcus's stagnation claims with concrete use casesKantrowitz refuses the framing that models have not improved since GPT-4, laying out detailed personal test cases including visual rock-climbing coaching and end-to-end vibe coding in Claude.
Biggest teaching moment ▶ 9:55 Marcus distinguishes general reasoning from narrow synthetic verificationMarcus educates Kantrowitz on why test-time compute is not true abstract reasoning, clarifying that it is narrow pattern mimicry restricted to closed domains with synthetic verification like math and code.
Alex holds their own ▶ 28:50 Kantrowitz maps Marcus's critique to an Nvidia market crashKantrowitz demonstrates sharp market domain expertise by spelling out how an OpenAI compute plateau directly undermines the entire capex foundation driving Nvidia's multitrillion-dollar valuation.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| The Diminishing Returns of Large Language Model Scaling | 6 | 5 | 6 | 6 | Kantrowitz tests Marcus's thesis by citing statements from tech leaders and noting compute scaling continues to yield improvements. Marcus reframes the argument, differentiating between mathematical exponential scaling laws and modest linear gains with diminishing returns. | |
| Reasoning Models, Black Box Architecture, and Interpretability Challenges | 5 | 6 | 6 | 5 | Kantrowitz raises test-time compute as a viable continuation of progress. Marcus counters by arguing that test-time reasoning is an uninterpretable black-box trick that only functions in closed synthetic domains like code and math. | |
| Everyday AI Utility Versus Hallucinations and Cognitive Atrophy | 7 | 5 | 6 | 7 | Kantrowitz mounts a strong personal counter-argument, presenting his hands-on experience vibe-coding financial apps and using multimodal posture analysis on o3. Marcus responds by pointing out benchmark contamination, subtle hallucinated errors, and coding regurgitation limits. | |
| Valuation Bubbles, Lack of Moats, and Market Commoditization | 6 | 4 | 5 | 5 | Kantrowitz connects Marcus's scaling plateau thesis to downstream macroeconomic threats, particularly Nvidia's stock valuation and market price wars. Marcus agrees, highlighting commoditization and the absence of technical moats. | |
| Defining Artificial General Intelligence and Public Prediction Wagers | 6 | 6 | 8 | 7 | Kantrowitz puts Marcus directly on the spot by asking when he will admit he is wrong. Marcus vigorously rejects the insinuation, highlighting his public million-dollar wagers and explicit benchmark criteria. | |
| Biological Threats, Open Source Dangers, and Surveillance Monetization | 5 | 5 | 5 | 4 | The conversation moves into biosecurity risks, open-source hazards, and corporate surveillance models. Kantrowitz admits genuine concern over feeding sensitive personal data into chatbots that may be monetized via targeted advertising. | |
| System 1 Versus System 2 and the Neurosymbolic Future | 4 | 5 | 3 | 2 | Marcus breaks down the synthesis of Kahneman's System 1 neural networks with System 2 classical symbolic logic, proposing neurosymbolic AI as the genuine path to AGI. Kantrowitz facilitates the explanation and wraps up the interview. |
Statements from this episode (14)
Marcus: OpenAI's Project Orion failed and became GPT-4.5
“So OpenAI tried to build GPT-V and they had a thing called Project Orion and it actually failed. And eventually got released as GPT four and a half. So what they thought was going to be GPT five just didn't meet expectations.”
Marcus: Grok 3 delivers minimal gains despite 10x compute
“So he built Grok three and by his own testimony, it was 10 times the size of Grok two. It's a little better, but it's not night and day, right? Grok two was night and day better than the original Grok. GPT four was night and day better than GPT three. GPT thre…”
Marcus: Models like o1 are not systematically better than GPT-4
“Models like O-I are not systematically better than GPT-IV. There's, they're better in certain use cases. Especially ones where you can create data in advance.”
Marcus: 'Reasoning' AI models only mimic patterns without genuine abstractions
“Now, the reason I wouldn't call them reasoning models, though you're right that many people do, is what I think they're doing is basically copying patterns of human reasoning. They're getting data about how humans reason certain things, but the depth of reason…”
Marcus: OpenAI's o3 hallucinates more than preceding models
“I'll give you just one more example is O three apparently hallucinates more than the models that came before it.”
Marcus: AI scaling laws are no longer true
“So the scaling quote laws were empirical guesses about how these models work and they were true for a little while, which was amazing. And they're not true anymore, which is also amazing in a way.”
Marcus: Vals AI benchmark shows LLM accuracy under 10% on financial charts
“Where they looked at things like, can you pull out a chart based on a series of financial statements, SEC statements from a bunch of companies and these systems all claimed to do it, but accuracy was under 10%. And overall on this new benchmark, accuracy was a…”
Marcus: AI coding assistants regurgitate public code but cannot debug software
“That's their sweet spot is regurgitation. And so Yeah, they can build the stuff that's out there, but if you want to code things in the real world, you usually want to code something that's new, and these systems have a lot of problems with that. And another r…”
Marcus: Microsoft study suggests chatbot usage impairs human critical thinking
“Well, Microsoft did a study, in fact, suggesting that critical thinking was getting worse as a function of them.”
Marcus: OpenAI is not worth $300B and won't IPO at $3T
“I don't think LLMs will disappear. I think they're useful, but this, yeah, the valuations don't make sense. I mean, I don't see open AI being worth three hundred billion dollars. And you have to remember that venture capitalists have to like 10 X to be happy o…”
Marcus: No AI lab has a technical moat; OpenAI has users
“Nobody has a technical mode. OpenAI has a user mode.”
Marcus: Gen AI's core business model will be surveillance and ads
“The business model of gen AI will be surveillance and hyper-targeted ads, just like it has been for social media.”