Mar 20, 2024 · 1h 3m · big-technology
Predicting AI’s Next Advances — With Suhail Doshi
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
Playground founder and CEO Suhail Doshi joins Alex Kantrowitz to examine the rapid pace of AI breakthroughs, analyzing advancements in visual generative models, the realities of machine reasoning, and the shifting competitive landscape among major tech giants.
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 31.1% of the talking time here. How this is scored →
Alex as informed peer 4.9 Guest teaching 4.8 Guest disagreement 2.0 Alex pushing back 2.7
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 44:40 Rejecting search traffic stability
Suhail flatly rejects Alex's skepticism of a 25% drop in Google search traffic, asserting the decline will be far larger and arrive much faster.
Hardest push from Alex ▶ 38:27 Defending Sora's physics understandingAlex deliberately takes the counter-position to challenge Suhail, arguing that complex fluid simulations in Sora prove a real-world physics model.
Biggest teaching moment ▶ 38:42 Training data mimicry vs. physical lawsSuhail deconstructs Alex's belief that video generators understand physics by demonstrating that the model simply replicates video representations, including fictional CGI physics like The Matrix.
Alex holds their own ▶ 49:00 Citing GPU fleet metrics across Big TechAlex brings detailed external reporting on Meta's 650,000 GPU-equivalent capacity and compares it against enterprise benchmarks from ServiceNow.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Tracking the AI Innovation Curve: Beyond Chatbots to Multimodal Robotics | 4 | 5 | 1 | 2 | Alex asks open-ended questions about where AI sits on the innovation curve and the ultimate North Star for chatbots. Suhail educates on how robotics hit a ceiling that large language models and multimodal systems are now uniquely positioned to unlock. | |
| Deconstructing AI Reasoning and Out-of-Distribution Problem Solving | 5 | 6 | 3 | 4 | Alex pushes on OpenAI's Q* and whether solving novel math problems proves reasoning. Suhail pushes back on the premise, explaining philosophically and technically that true reasoning can only be tested out-of-distribution, as models may simply fit training data. | |
| The Evolution of Image Generation and Empowering Non-Designers | 5 | 4 | 1 | 2 | Alex shares his own experience using image generators for his newsletter and asks about the business case and threat to designers. Suhail explains how graphics will be democratized like 3D animation evolved from 2D without destroying creative jobs. | |
| Precise Image Editing and the Vision for Unified Pixel Models | 3 | 5 | 1 | 1 | Suhail outlines Playground's ambition to build a single unified pixel model that handles editing, generation, and vision tasks just as LLMs handle diverse text tasks. Alex mostly listens and facilitates with light humorous comments. | |
| The Future of AR Interfaces and Form-Factor Agnostic AI | 5 | 4 | 2 | 3 | Alex quotes Suhail's tweet defending smartphones to question the viability of AR headsets like Vision Pro. Suhail reframes his stance as form-factor agnostic, arguing the primary value is AI-driven pixel manipulation across any interface. | |
| Diagnosing the Homogenized Look of AI Art: Average Human Preference as Pop | 5 | 4 | 1 | 2 | Alex asks why current AI images have a recognizable, homogenized aesthetic compared to human designers. Suhail diagnoses this as models optimizing for average human preference, drawing a parallel to top-40 pop music. | |
| Prompting Artist Styles, Intellectual Property, and Fair Attribution | 5 | 4 | 2 | 4 | Alex presses on the ethical and copyright implications of prompting specific living artist names without compensation. Suhail acknowledges the issue, defends prompt modifiers as stylistic shortcuts, and explains Playground's attribution link-backs. | |
| Strategic Focus on Images and the Rapid Obsolescence of Video Demos | 4 | 4 | 1 | 2 | Alex asks if Suhail regrets focusing strictly on static images rather than jumping into video generation like Sora. Suhail explains the rapid obsolescence cycle of generative tech, predicting early video demos will look primitive within a year. | |
| The Physics Debate in Video Generation: Simulation vs. Training Data Mimicry | 6 | 6 | 4 | 4 | Alex contrasts Yann LeCun's V-JEPA approach with Sora's generative approach to argue Sora demonstrates physics understanding. Suhail challenges the claim, arguing Sora merely mimics visual correlations in training data rather than true underlying physics. | |
| The Vulnerability of Google and the Looming Collapse of Traditional Search | 6 | 5 | 4 | 5 | Alex expresses skepticism over a Gartner prediction that search volume will fall 25% by 2026. Suhail forcefully disagrees with Alex, predicting an even steeper drop sooner due to the rapid leap in conversational answer engines. | |
| Big Tech AI Strategies: Meta's Compute Arsenal and NVIDIA's Hardware Lock-In | 6 | 5 | 3 | 3 | Alex brings industry data on Meta's 600k GPU equivalent fleet and questions whether NVIDIA's CUDA software creates an unbreakable moat. Suhail nuances the moat argument, pointing out compute hardware costs drive decisions far more than software lock-in. | |
| Cloud Wars and Fast Followers: AWS Short-Sightedness and Apple's AI Positioning | 5 | 6 | 2 | 2 | Alex asks about Amazon and Apple's AI positioning. Suhail offers sharp insider critique on AWS's short-term price gouging on GPUs and praises Apple's patient fast-follower strategy. | |
| OpenAI's Singular AGI Mission and Satya Nadella's Ecosystem Mastery | 5 | 5 | 1 | 1 | Alex highlights Satya Nadella's corporate turnaround at Microsoft, while Suhail explains how Nadella is methodically neutralizing platform risk by hosting every foundation model on Azure. |
Statements from this episode (27)
Doshi: Image AI is probably a couple years behind a GPT-4 moment
“Images is interesting because it's probably a couple years behind a GPT-IV true moment, right?”
Doshi: Robotics research hit a ceiling around 2020 and hasn't recovered
“Broadly robotics kind of asymptoted and hit a ceiling. About like three or four years ago, and the research still isn't like kind of on a trajectory. That's amazing.”
Doshi: Powerful multimodal models will probably yield Westworld-style humanoid robots
“So I think that if the models get more powerful, We're probably going to see, we're probably going to see some kind of Westworld version of the world. We're going to go way beyond a chatbot.”
Doshi: Articulating explanations does not prove AI is genuinely reasoning
“Just because something is able to articulate its reason for doing, for getting to an answer, doesn't mean that it is necessarily reasoning.”
Doshi: True AI reasoning can only be tested out-of-distribution
“One way that you could maybe litmus test reasoning is if you gave the model something truly out of distribution, right? Like an example of something that's truly out of distribution that humans face was COVID, a pandemic that it had not yet seen before. And th…”
Doshi: The Predominant Use Case for Language Models Continues to Be Homework
“There are a lot of things posted on Twitter about how people are using language models, but the predominant use case continues to be homework.”
Doshi: Graphic Designers Won't Go Extinct From AI but Will Retool
“So did all the people that you know, that were people that drew the two D cartoons, they lose their jobs. And, you know, that was the end of the end of an era. Definitely not people retooled the stories that came from them were still really material to their c…”
Doshi: Computer vision lacks a single unified model equivalent to LLMs
“This is missing in vision, but definitely kind of exists in language and language. We can solve hundreds or thousands of different tasks. But in graphics, but in vision, it's all separated. It's kind of like where language was back three or four years ago wher…”
Doshi: Meta has one of the world's best graphics teams
“I think one thing that Meta is doing, regardless of where VR is headed or AR is headed, is they have one of the most world-class teams for graphics, and they have to because of all the stuff that they're doing in VR.”
Doshi: AI Image Models Homogenize Styles by Overfitting to Average Human Preference
“I think that the models are probably too curated and maybe overfit to be based on human preference. And human preference isn't your human preference. It's your preference. I mean, it's some kind of average of human preference.”
Doshi: AI Model Quality Is the Lowest It Will Ever Be
“Quality is the lowest it's ever going to be starting today.”
Doshi: Prompting AI image models by artist name will disappear this year
“I think that this thing is going to go away in the next year, this year.”
Doshi: In a year, OpenAI's Sora will look 'not even close'
“My feeling is my bet to people, you know, sort of listening is that in a year we will think, you know, we'll think something like Sora was not even close.”
Suhail Doshi: Video AI models mimic training data rather than understanding real physics
“What's really doing is it's representing
The physics it understands in the videos it's being trained on, which could be incorrect physics. It's really what it understands what it's being trained on is kind of my main, the main thrust of my point. And that to a…”
Doshi: Debate over whether AI truly understands physics is practically irrelevant
“I take the argument that it doesn't, I just like, I don't think it matters. Like at the end of the day, we are making these models could do it or could not, but either way, what matters is what utility it brings to humanity.”
Doshi: Google's bureaucracy is ill-suited for AI's rapid velocity
“And they have to contend with a significant complex bureaucracy that is not really well attuned for the velocity that AI is running at right now.”
Kantrowitz: Traditional search will remain for navigation, AI for curiosity
“I think that search is going to be, continue to be a way that we use web navigation and AI search like perplexity will be more to satisfy curiosity. And engage with different topics.”
Doshi: Search engine traffic will drop over 25% before 2026
“I mean, I think there's a very high probability it is greater than that number in a shorter span of time.”
Doshi: GPT-5 is probably slated for release in 2024
“GPT five is probably slated imminently this year.”
Doshi: AI advancement is outpacing Moore's Law by a lot
“We're still so at the beginning, the early phases of this thing that it is moving faster than Moore's law. By a lot.”
Doshi: AI companies would leave NVIDIA if cheaper compute alternatives existed
“It's not CUDA that's keeping, I think, keeping a lot of us. It's actually that there is nothing really dramatically better than NVIDIA's GPUs. And so if there's nothing dramatically better than, I mean, the reality is the cost for training and inference are so…”
Doshi: NVIDIA H100 is 1.9x faster but 2x costlier
“One of the problems with NVIDIA was that they released their H 100, but they didn't really reduce its cost. You know, it is, it's two is, you know, 1.9 X faster, but two X costlier.”
Doshi: AWS Quoted 5x Market Price for GPUs in 2023
“I mean, we look, we were looking for compute last year and AWS wanted to charge us five times more for the same GPU than 10 different providers all around them.”
Doshi: Apple Is Well Positioned to Copy Google's Best AI Experiments
“And I think Apple is super well positioned to just like let Google do all those experiments and then just pluck the ones that are the best ones and use its massive install base and distribution power to deliver an amazing experience, not a rushed rushed one.”
Doshi: Sam Altman only cares about AGI, treating video as stepping stones
“I think that Sam is genuinely focused on AGI and attaining that. And I think he does not care about Graphics and, you know, video necessarily. I think he, those are stepping stones and that helps research you know, get to the next point, but I think he is very…”
Doshi: Satya Nadella is brilliantly bringing all AI models onto Azure
“If you observe very carefully, he is actually partnering with everybody. He is bringing all the models into Azure, right? And he's doing it very methodically, and I think that he is really setting up Azure to be to leap and be a lot more competitive.”