Jul 1, 2024, 8:00 PM ETInterview

Zapier’s Mike Knoop launches ARC Prize to Jumpstart New Ideas for AGI | Training Data

Zapier operates as a workflow automation platform supporting over 6,000 integrations, with a primary focus on non-technical users rather than engineers.
In January 2022, Mike Knoop transitioned from an executive product role to an individual AI researcher at Zapier after the "Chain of Thought" paper demonstrated that LLMs could perform reasoning tasks.
Over half of Zapier's internal staff now utilizes AI daily to build automations, primarily for content generation or data extraction from unstructured text.
A specific internal workflow for generating Zap templates achieved a 100x labor enhancement rate by shifting human labor from "doing" to "reviewing."
The production rate for Zap templates increased from approximately 10 per day to 1,000 per day using an automated AI system that generates and discards low-quality outputs.
Zapier currently executes approximately 10 million AI tasks per month, positioning it as a leading example of agentic AI systems operating with minimal human intervention.

The ArcPrize was established to evaluate "Artificial General Intelligence" (AGI) defined not by economic utility, but by the efficiency of acquiring new skills from limited core knowledge priors.
François Chollet (co-founder of ArcPrize) posits that true AGI requires systems to rapidly learn novel tasks without needing to retrain from scratch, contrasting with current LLMs that rely on memorization.
The ARC benchmark has remained unbeaten since 2019, showing a deceleration in progress rather than the acceleration toward human-level performance seen in other AI evaluations.
The current state-of-the-art score on the Arc benchmark is 39%, while the grand prize target is set at 85%.
A score of 85% would demonstrate a system's ability to synthesize core knowledge priors (e.g., symmetry, objectness) into programs for unseen tasks with exacting accuracy.

The competition enforces a strict "no internet" rule to prevent competitors from accessing frontier closed models (e.g., GPT-4, Claude) via APIs, which would bypass the efficiency constraints.
A compute limit of 1 GPU-hour (currently 12 hours) is imposed to force solutions to prioritize efficient reasoning over brute-forcing all possible program permutations.
The benchmark utilizes a private test set that remains unseen by the community to prevent overfitting and ensure that solutions generalize rather than memorize the puzzles.
The prize money is contingent on participants publicly releasing reproducible, open-source code to accelerate collective progress and prevent knowledge silos.

Knoop argues that the current dominance of "scale is all you need" narratives in frontier labs has diverted attention and investment away from the novel algorithmic ideas required to solve Arc.
The most successful recent attempts to improve Arc scores (e.g., Ryan Greenblatt) involve using LLMs as "outer loops" to generate and verify code traces, achieving scores in the low 40s on the public task set.
Historical winners of similar benchmarks (like Go or Chess) were often developed by big labs, whereas Arc solutions have largely come from "outsiders" or individuals from unrelated fields (physics, game dev).
Knoop predicts a 50% score is achievable by the end of 2024, but reaching the 85% grand prize threshold may require breakthroughs in deep learning-guided program synthesis that are not yet available.

Knoop advocates against premature regulation based on "mythical" scenarios of superintelligence, arguing that policy should be grounded in empirical evidence of what systems can and cannot currently do.
The introduction of AGI is expected to be an incremental, stair-step process rather than a singular event, allowing society to adjust to capabilities and risks as they emerge.
Solving Arc is viewed as a critical step toward fixing the "hallucination and trust" issues in current AI, which currently limit AI deployment to low-risk, high-cost environments.
Knoop believes that without solving efficiency-based generalization, human innovation will remain rate-limited, preventing AI from assisting in high-level scientific discovery and invention.
The interview concludes with the assertion that "new ideas" are necessary to break the current plateau, and that the ArcPrize aims to reinvigorate the search for these ideas through open competition.