Aug 28, 2024 · 50m · news
Arvind Narayanan: AI Scaling Myths, The Core Bottlenecks in AI Today & The Future of Models | E1195 · 20VC with Harry Stebbings
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
Princeton computer science professor Arvind Narayanan joins host Harry Stebbings on the 20VC podcast to critique the speculative hype surrounding generative AI, exploring the limits of scaling, the challenges of enterprise deployment, and the tangible societal impacts of the technology.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 24.5% of the talking time here. How this is scored →
Harry as informed peer 4.5 Guest teaching 4.8 Guest disagreement 2.3 Harry pushing back 3.7
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 30:12 Arvind firmly rejects host's deepfake panic
Arvind directly pushes back on Harry's assertion that AI image/audio creation is inherently dangerous, insisting it is not fundamentally an AI problem.
Hardest push from Harry ▶ 30:12 Harry refuses guest's dismissal of AI election misinformationHarry refuses to accept Arvind's calm dismissal, posing a direct hypothetical about producing a fake podcast of Donald Trump declaring war on China.
Biggest teaching moment ▶ 41:46 Arvind dismantles the AI nuclear weapon analogyArvind explicitly identifies a category error in Harry's comparison of AI to nuclear weapons, educating him on software security vs physical weaponry.
Harry holds his own ▶ 5:52 Harry counters data bottleneck claims with YouTube data statsHarry demonstrates deep preparation by citing 150 billion hours of YouTube video data and synthetic data capabilities to challenge Arvind's data shortage thesis.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Hook: The Model Scaling Wall and Data Bottleneck | 2 | 3 | 1 | 1 | Harry opens with warm fan praise and asks introductory questions comparing AI hype to Bitcoin hype. Arvind explains his past disillusionment with crypto and contrasts it with AI's net positive societal impact. | |
| The Generative AI Product Pitfalls | 4 | 4 | 2 | 3 | Harry cites Microsoft CTO Kevin Scott on compute scaling to question whether model scaling has hit diminishing returns. Arvind explains why parameter scaling is hitting data bottlenecks and why compute now goes toward making models smaller. | |
| Rebutting Pushbacks on YouTube and Synthetic Data | 6 | 6 | 3 | 5 | Harry pushes back on data bottlenecks by raising two specific counterarguments: 150 billion hours of YouTube video data and synthetic data. Arvind counters that text extracted from YouTube video is an order of magnitude smaller than current training sets and that synthetic data acts like a snake eating its own tail. | |
| Codifying Enterprise Knowledge and Deployment Speeds | 6 | 5 | 1 | 3 | Harry cites Scale AI CEO Alex Wang regarding uncodified enterprise knowledge ('showing your work') as an obstacle for AI agents. Arvind strongly agrees and compares enterprise AI deployment to the slow, iterative feedback loop of self-driving cars. | |
| Why Smaller Models and On-Device Processing are Winning | 4 | 6 | 2 | 3 | Harry questions why smaller models are winning and asks if Moore's Law will render cost concerns irrelevant in 3 to 5 years. Arvind educates Harry using Jevons Paradox, demonstrating that lower model costs actually increase total enterprise inference spend. | |
| Training vs. Inference Costs and the Hardware Obsolescence Cycle | 6 | 4 | 1 | 4 | Harry cites Sequoia partner David Cahn regarding data center hardware cycles and the rapid obsolescence of H100 GPUs. Arvind details the trade-offs between training and inference compute costs before noting hardware exponentials eventually follow sigmoid curves. | |
| The Minefield of AI Evaluations and Benchmarks | 5 | 5 | 2 | 4 | Harry quotes Arvind's essay on benchmark evaluation and contrasts conflicting CEO timelines for AGI. Harry then presses Arvind on whether chasing superintelligence and building practical products are mutually exclusive inside OpenAI. | |
| Creating Gods vs. Building Products | 5 | 5 | 2 | 4 | Harry challenges Arvind's thesis that startups can compete in AGI by pointing to Mark Zuckerberg's $50B compute commitment vs OpenAI's fundraising. Arvind argues that as foundation models commoditize, value moves to the application and agent layer where smaller teams can win. | |
| Foundation Model Oligopolies and the Reality of AI Regulation | 5 | 4 | 2 | 4 | Harry argues that foundation models will naturally consolidate around cloud cash cows like Amazon, Google, and Meta. Arvind agrees this is a valid antitrust risk and clarifies that AI regulation should target harmful acts rather than the underlying software technology. | |
| Debating the Misinformation Threat and the "Liar's Dividend" | 5 | 5 | 6 | 7 | Harry strongly challenges Arvind's dismissal of AI misinformation threats by arguing that deepfaked audio could declare war or incite real-world riots. Arvind repeatedly rejects Harry's framing, insisting that misinformation is a societal/distribution issue rather than an AI technology problem. | |
| Immediate Societal Costs: Deepfakes and the Education Shift | 4 | 4 | 2 | 6 | Harry forcefully rejects the hype around AI replacing general practitioners ('Are you high?... You're not gonna shove your smartphone up your nostril'). Arvind agrees with Harry's physical diagnostic skepticism and explains how AI is best integrated into back-office medical workflows. | |
| The Social Reality of Education vs. Self-Taught AI Developers | 4 | 6 | 2 | 3 | Harry challenges AI personalized tutoring claims by highlighting the essential human motivation in teacher-student relationships. Arvind validates Harry's point by sharing his background as a self-taught student in India and diagnosing how AI developers mistake themselves for typical learners. | |
| Why AI Job Replacement Fears are Overblown | 5 | 6 | 4 | 5 | Harry quotes Alex Wang's assertion that AI could be a bigger weapon than nuclear arms to question open-source models. Arvind firmly refutes the framing as a 'category error' and explains why defensive open-source security outperforms restricted closed models. | |
| Reflections on the Illusion of Rapid AI Progress | 3 | 5 | 3 | 2 | Harry asks Arvind what beliefs he has changed over the past two years. Arvind explains that the close release timing of GPT-3.5 and GPT-4 gave an illusion of rapid progress, and dismantles sci-fi fears of self-aware AI. | |
| Quick-Fire Round: Nvidia, Policy, and the Forgotten Children | 3 | 4 | 1 | 1 | Harry runs a rapid quick-fire round on Nvidia, policy, Yann LeCun vs Geoffrey Hinton, and neglected questions. Arvind gives concise answers, concluding that society needs to focus on AI's impact on children. |
Statements from this episode (36)
Narayanan: AI model scaling cycles are ending
“We're not gonna have too many more cycles, possibly zero more cycles, of a model that's almost an order of magnitude bigger in terms of the number of parameters than what came before, and thereby more powerful.”
Narayanan: AI models have exhausted accessible training data
“These models are already trained on essentially all of the data that companies can get their hands on.”
Narayanan: AI is a net positive for society, unlike Bitcoin
“While there are harms around AI, I think it has been a net positive for society. I can't say the same thing about Bitcoin.”
Narayanan: Generative AI companies deluded themselves into ignoring product-market fit
“They didn't think about actually building products, you know, making things that people want, finding product market fit, and all those things that are so basic in tech, but somehow AI companies diluted themselves into thinking that the normal rules don't appl…”
Narayanan: Skeptical GPT-5 will yield a leap comparable to GPT-4
“Are we going to see a GPT-V that's as big a leap over GPT-V as GPT-V was over GPT-V? I'm frankly skeptical.”
Narayanan: YouTube transcript text is an order of magnitude smaller than current AI training sets
“A hundred and fifty billion hours of video sounds, you know, really impressive. But when you put that video through a speech recognizer and actually extracts the text tokens out of it and deduplicated and so forth, it's actually not that much. It's an order of…”
Narayanan: Recursive AI scaling using synthetic data will not work
“And the other way to look at synthetic data is, okay, you take one trillion tokens, you train a model on it, and then you output 10 trillion tokens, so you get to the next bigger model, and then you use that to output a hundred trillion tokens. You know, I'll …”
Narayanan: AI training data quality matters far more than quantity
“What we've learned in the last two years is that the quality of data matters a lot more than the quantity of data.”
Narayanan: Enterprise AI deployment will be very slow
“It's got, you have to actually deploy AI to be able to get to certain types of learning, and I think that's gonna be very slow, and I think the a good analogy is self-driving cars, of which we had prototypes, you know, two or three decades ago, but for these t…”
Narayanan: Enterprise AI adoption is bottlenecked by cost, not capability
“My view is that in a lot of cases, the adoption of these models is not bottlenecked by capability. If these models were actually deployed today to do all the tasks that they're capable of, it would truly be a striking economic transformation. The bottlenecks a…”
Narayanan: Lower AI inference costs will increase total enterprise AI spend
“And I predict that we're going to see the same thing with models when models get cheaper. They're put into a lot more things, and so the total amount that companies are spending on inference is actually going to increase.”
Narayanan: Inference costs dominate training costs for popular AI models
“Over the lifetime of a model, when you have billions of people using it, the inference cost actually adds up, and for many of the popular models, that's the cost that dominates.”
Narayanan: Exponential AI growth will flatten and models will commoditize
“So I think that's going to happen both with models as well as with these hardware cycles. You know, I can't predict how long that's going to take, but we are, I think, going to get to a world where models do get commoditized.”
Narayanan: AI developers over-optimize models for benchmarks over real-world performance
“When there is so much pressure to do well on these benchmarks, developers are intentionally or unintentionally optimizing these models In ways that look good on the benchmarks, but don't look good in real world evaluation.”
Narayanan: GPT-4 passing bar and medical exams meant nothing for actual practice
“So when GPT-IV came out and OpenAI claimed that it passed the bar exam and the medical licensing exam people were very excited slash scared about what this means for doctors and lawyers, and the answer turned out to be approximately nothing, right? Because it'…”
Narayanan: Do not trust overconfident AGI predictions from tech CEOs
“I wouldn't put too much stock into these overconfident predictions from CEOs.”
Narayanan: OpenAI focuses on products as researchers depart to Anthropic
“Folks focused on superintelligence didn't feel very welcome at the company, and there has been an exodus of very prominent people, and Anthropic has picked up a lot of them. So it seems like we're seeing a split emerging where OpenAI is more focused on product…”
Narayanan: Oligopoly of tech giants dominating AI models is serious risk
“That might happen. I think that's a very serious possibility, and I think this is actually one area where regulators should be paying attention. You know, what does this mean for market concentration, antitrust, and so forth?”
Narayanan: AI bots influencing elections with misinformation is not a real danger
“So people have been worried, for instance, about bots creating misinformation with AI and influence in elections and that sort of thing. We're very, very skeptical that that's going to be a real danger.”
Narayanan: AI proliferation will force people onto trusted news sources
“That's right. So that, that actually is our prediction. People we predict are going to be forced to rely much more on getting their news from trusted sources.”
Narayanan: Misinformation affirms existing beliefs rather than changing them
“In a way, I think misinformation is more of a symptom than a cause. I think, you know, misinformation slots into and affirms people's existing beliefs as opposed to changing their beliefs. And I think the impact on AI here, again, has been tremendously exagger…”
Narayanan: Misinformation is a social media distribution problem, not an AI generation problem
“To the extent it's a technology problem, it's more of a social media problem, really, than an AI problem, because the hard part of misinformation is not generating it, it's distributing it to people and persuading them, and social media is often the medium for…”
Narayanan: No reliable method exists to detect AI-generated text
“There's no way really to catch AI-generated text or homework answers.”
Narayanan: AI companies should help cover education system adaptation costs
“It forces a lot of costs upon the education system, and ideally, AI companies should be bearing some of that cost.”
Narayanan: AI developers mistakenly assume average learners are self-taught
“You have a lot of AI developers who are thinking of themselves as the typical learner, and they're not.”
Narayanan: AI tools will not become the default way people learn anytime soon
“I'm very skeptical that these new kinds of learning are going to get to a point anytime soon where they're going to become the default way in which people learn.”
Narayanan: AI automates individual tasks, not entire jobs
“You know, the more abstract way of saying that is, as economists would put it, jobs are bundles of tasks, And AI automates tasks, not jobs. So if there are, you know, 20 different tasks that comprise a job the odds that AI is gonna be able to automate all 20 o…”
Narayanan: Comparing AI to nuclear weapons is a category error
“I think it's a bit of a category error there. I mean, a nuclear weapon is an actual weapon. AI is not a weapon.”
Narayanan: Near state-of-the-art AI on personal devices will accelerate
“We have you know, close to state of the art AI models that can already run on people's personal devices, and I think that trend is only going to accelerate.”
Narayanan: Restricting global access to AI models will fail
“Even if one country decides that models should be closed, the odds of getting every country to enact that kind of, ah, ah, rule are, you know, just vanishingly small. So if our approach to safety with AI is going to be premised on ensuring that quote unquote b…”
Narayanan: GPT-4's 18-month training period created an illusion of rapid progress
“So I think, like a lot of people, I was fooled by how quickly after GPT-III.V, GPT-IV came out. It was just, you know, three months or so, but it had been in training for 18 months.”
Narayanan: No AI model has meaningfully surpassed GPT-4 in 18 months
“And what we've seen in the nearly year and a half since GPT-IV came out is that we haven't really had models That have surpassed it in a meaningful way.”
Narayanan: Fears of spontaneous AI consciousness have no basis in reality
“When we look at the way that AI is architected today, that kind of fear has no basis in reality. Maybe one day in the future, you know, people are going to build AI systems where that becomes at least somewhat possible. And we should, you know, we should have …”
Narayanan: Public right to AI knowledge overrides private commercial interests
“I don't think I would be a good CEO, but if there were one thing I could change about OpenAI, I think I think the need for the public to know what is going on with AI development overrides the you know, commercial interests of any company, so I think there nee…”
Narayanan: Nvidia is trying aggressively to migrate from hardware to services
“I do find it interesting that NVIDIA itself Has been trying to migrate really, really hard out of hardware into becoming a services company.”
Narayanan: LLMs are an off-ramp to superintelligence requiring new breakthroughs
“I have to say, I really like Jan LeCun's perspectives on various things, including his view that LLMs Our quote unquote off-ramp to super intelligence that, you know, in other words, we need a lot more scientific breakthroughs, as well as tamping down the fear…”