Apr 6, 2025 · 33m · tbpn
Mike Knoop (Arc Prize) on Why Scaling AI Won’t Get Us to AGI
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
Zapier co-founder Mike Knoop explains why scaling pre-training and test-time compute is insufficient for reaching AGI, making the case for the ARC Prize to benchmark genuine fluid reasoning and stimulate decentralized algorithmic innovation.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.0 Guest teaching 4.8 Guest disagreement 1.9 The hosts pushing back 1.1
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 0:16 Mike dismantles economic definitions of AGI
Mike directly rejects the host's proposed GDP threshold metric, asserting that economic output conflates pattern memorization with novel problem-solving capability.
Hardest push from the hosts ▶ 0:26 Host presses on the tension between Marcus and SuttonThe host challenges the prevailing industry consensus by demanding a reconciliation between Gary Marcus's symbolic AI skepticism and the Bitter Lesson's scale-maximalist doctrine.
Biggest teaching moment ▶ 0:31 Mike clarifies ARC's underlying programmatic natureWhen the host proposes creating a coding-specific ARC benchmark, Mike corrects the premise by explaining that ARC tasks are already pure program synthesis problems evaluated as 2D numeric matrices.
The host holds their own ▶ 0:19 Host presents viral empirical data on visual style transferThe host commands the dialogue by citing his own controlled viral posting experiment to break down why Ghibli-style generation succeeds aesthetically where VFX photo-realism fails.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Mike Knoop on Zapier, LLM Reliability, and ARC Prize | 3 | 2 | 1 | 0 | The host warmly opens the episode and asks Mike Knoop to introduce his background at Zapier and the genesis of ARC Prize. Mike explains his realization that enterprise customers found LLMs too unreliable for unsupervised automation. | |
| ARC-AGI Evolution and the Limits of Test-Time Compute | 4 | 6 | 2 | 1 | Mike explains the mechanics of ARC-AGI 1 and ARC-AGI 2, detailing how recent models like o3 and test-time compute fail to generalize efficiently on novel benchmarks. Jordy and the host interject with questions about capital constraints versus algorithmic bottlenecks. | |
| Decentralizing Frontier AI Research and the Vesuvius Challenge | 6 | 4 | 1 | 1 | The host draws a strong parallel between ARC Prize and the Vesuvius Challenge to decentralized problem solving. The host then asks technical questions about Python execution and overfitting guardrails, which Mike clarifies by explaining Kaggle compute constraints and private test splits. | |
| Overcoming Reliability Bottlenecks to Build Consumer AI Agents | 4 | 5 | 1 | 1 | Jordy asks why consumer agent experiences like flight booking remain underwhelming across major tech companies. Mike lays out his framework of concentric rings of reliability risk and explains how fluid intelligence on ARC translates directly to robust agentic execution. | |
| Defining AGI Through Human-AI Task Gaps Versus Economics | 5 | 6 | 4 | 2 | The host proposes measuring AGI by macroeconomic contribution thresholds such as percentage of global GDP. Mike firmly pushes back against this definition, arguing that economic metrics obscure underlying capability mechanisms like memorization versus true adaptation. | |
| Viral Ghibli AI Models and Novel Architectural Exploration | 7 | 3 | 2 | 1 | The host demonstrates sharp domain expertise by sharing an A/B social media test comparing a real Ghibli still with an AI-generated scene, explaining the uncanny valley dynamics of artistic style transfer. Mike connects this to non-diffusion architectural exploration and warns startups against brute-force model pre-training. | |
| Reevaluating Gary Marcus, Scaling Limits, and the Bitter Lesson | 7 | 5 | 2 | 2 | The host articulates an informed synthesis between Gary Marcus's critiques of pure deep learning and Rich Sutton's Bitter Lesson. Mike agrees Marcus was largely vindicated on generalization limits and reinterprets Sutton's paper by highlighting that scalable architectures still originate from human ingenuity. | |
| Program Synthesis Pioneers and the Enduring Value of Coding | 4 | 7 | 2 | 1 | Mike highlights key researchers in program synthesis before the host asks if someone could create an ARC-like challenge tailored specifically for human programmers. Mike gently educates the host by revealing that ARC is fundamentally already a program synthesis challenge operating on discrete numeric matrices. |
Statements from this episode (14)
Knoop: AI Agent Failure Rates Make Unsupervised Automation Unviable
“Like, hey, I get the hype, but like, they just, they're not reliable enough yet. You know, they don't work two out of 10 times, and that just doesn't work for these unsupervised automation products.”
Knoop: ARC Saw No Progress Despite 50,000x Model Scaling
“Surprise that it basically hadn't, and not only hadn't been beaten, there'd basically been no progress in it which I thought was really fascinating given the fact that we've like scaled up these language model systems by almost like 50,000 times over the last,…”
Knoop: Pure LLMs Score 0% and o1 Scores 1% on ARC-AGI-2
“Pure LLM systems are scoring like zero percent now again on on arc B two single COT systems like R one and O one score like one percent.”
Knoop: Scaling Test-Time Compute Will Not Get Us to AGI
“And then there's a new story that's emerged over the last like five months, which is, oh, we're going to scale up this test time compute and that's going to get us to AGI. And I think what V two shows is that that's not quite either. We still need some structu…”
Knoop: ARC proves individuals can still advance frontier AI
“And I think arc shows that like, yeah, there actually are frontier problems. That are unsolved, that individual people and individual teams can actually make a difference on today.”
Knoop: AI intelligence benchmarks must measure compute efficiency, not brute force
“We do think that efficiency is actually a really, really important aspect of intelligence. You know, you can brute force your way up to intelligence, but we really do want to be shooting for like human targets and efficiency for this stuff.”
Knoop: OpenAI asked ARC Prize team to verify results on semi-private dataset
“Opening eye situation was a little different because they had reached out to us and said, Hey, we think we've got a really impressive result on the public eval set. And we'd like your help to verify. On the semi-private set, which is what we created that data …”
Knoop: AI agents will start working in 2025 due to ARC progress
“I actually think we're going to start to see agents start to work this year specifically because of progress on arc.”
Knoop: Language models operate by memorization rather than solving novel patterns
“Language models. Generally working like a memorization style regime where they're right. Learning lots of data. They're able to apply it to very similar types of patterns that they've seen before, but not novel patterns. That's what RKGI shows.”
Knoop: Startups just doing model training are lighting money on fire
“I think anyone who's like Just doing model training at this point is like lighting money on fire. If you really want to make a unique difference, especially if you're a small startup, like a founder, like you gotta go take an orthogonal approach. You gotta try…”
Knoop: Zapier never spent its $1M seed round from 2012
“It was about a million bucks back in 2012. We never spent the money. By the time we actually got the round closed, figured out who we wanted to hire, got the payroll started, like revenue had caught up. And so literally I think you could trace every dollar we …”
Knoop: Gary Marcus has been more right than wrong on deep learning
“I generally think he's been more right than wrong. I think if you like just take a limited five year view on this from 20, 20 up until 20, 20, end of 20, 24, you know, I think it was a generally right. Like he was making the right ideas.”
Knoop: Achieving AGI requires merging deep learning with program synthesis
“I actually don't think either is sufficient. I think some merger of the two is what's necessary to get to AGI.”
Knoop: People should still learn to code for technological leverage
“My hot take is, like, I guess, yes, you should still learn to code. Primarily because it's been, it gives you like leverage over technology today. And yeah, like I don't see that leverage over technology going away anytime soon, particularly if you want to wor…”