# The Perils of the AI Exponential Source: https://www.youtube.com/watch?v=dztw1yctjI4 Recap page: https://rapidrecap.app/video/dztw1yctjI4 Generated: 2026-02-24T03:03:09.256+00:00 --- ## Quick Overview The video analyzes the recent deceleration in the improvement rate of AI agent capabilities, specifically citing METR's "Moore's Law for AI agents" chart which shows task length doubling every 7 months, and contrasting this with recent, seemingly slower progress seen in models like GPT-5.3-Codex and Claude Opus 4.6, which exhibit doubling times closer to 1.5 and 4.5 months respectively, suggesting that the rapid exponential gains seen previously are beginning to plateau or that new scaling laws are emerging, which is further supported by commentary from experts like Daniel Rein and Guy Berger questioning the sustainability and methodology of these metrics. **Key Points:** - METR's "Moore's Law for AI agents" research previously indicated that the length of tasks AIs could double roughly every 7 months (0:05). - Recent models like GPT-5.3-Codex (05:20) and Claude Opus 4.6 (07:07) showed faster improvements, with doubling times of 6.5 hours/17 hours and 14.5 hours, respectively, which translates to significantly shorter doubling times than the original 7-month trend (05:00, 05:39). - Expert commentary, such as from Daniel Rein (06:50), noted that the measurement is extremely noisy and that the recent apparent jump in coding capabilities might be an outlier due to changes in scaffolding/format (06:18). - Economist Guy Berger (10:18) questioned the sustainability of the trend and pointed out that if current growth continues, AI could consume the entire economy by 2028 (07:50), leading to massive unemployment. - The Citrini Research memo, "The 2028 Global Intelligence Crisis" (07:25), models this scenario of rapid AI advancement leading to economic collapse and mass joblessness. - The discussion concluded that while the overall trend is still positive, there is evidence of saturation in certain areas, and the consistency of the measurement itself is being questioned by experts (06:33, 08:27). ![Screenshot at 00:05: METR's chart illustrating "Moore's Law for AI agents," showing the doubling time for task length has been consistently around every 7 months over the past six years, which sets the benchmark for comparison against newer models.](https://ss.rapidrecap.app/screens/dztw1yctjI4/00-00-05.jpg) **Context:** This AI Daily Brief discusses recent developments and debates surrounding the progress rate of AI agents, focusing heavily on the findings and implications of research conducted by METR (Model Evaluation & Threat Research). The core of the discussion revolves around METR's chart tracking AI performance on long tasks over time, often referred to as "Moore's Law for AI agents," and whether the latest flagship models (like GPT-5.3-Codex and Claude Opus 4.6) confirm or challenge the established exponential growth rate. ## Detailed Analysis The video reviews recent updates to AI capability scaling, specifically revisiting METR's findings on the "Moore's Law for AI agents," which previously showed task length doubling every 7 months (0:05). However, newer models are challenging this benchmark: GPT-5.3-Codex achieved a time horizon of 6.5 hours (high setting) in February 2023 (05:19), and Claude Opus 4.6 achieved a 50% time horizon of around 14.5 hours in December 2023 (07:07), representing a doubling time of about 4.5 months, which is significantly faster than the 7-month historical trend. This acceleration was noted, but some experts cautioned against over-interpreting these results. Daniel Rein (06:50) pointed out that scaffolding/formatting issues might inflate performance, and METR itself noted that the GPT-5.3-Codex results were partially based on a different scaffold (06:18). Separately, economist Guy Berger (10:18) questioned the sustainability of the trend, raising concerns about who owns the agents and the potential for massive economic crisis if AI continues to consume all sectors, as modeled by the Citrini Research piece, "The 2028 Global Intelligence Crisis" (07:25). Fejau (08:48) noted that the excitement around the Citrini piece is widespread, but some subsets of people are overestimating the immediate impact. Ultimately, the video suggests that while progress is impressive, concerns about the methodology's consistency and the potential for saturation in current task suites temper the perceived exponential rate (06:33, 08:48). ### METR's Core Findings - "Moore's Law for AI agents" established a 7-month doubling time for AI task length - GPT-5.3-Codex showed a doubling time of ~1.5 months in coding tasks - Claude Opus 4.6 showed a doubling time of ~4.5 months in general software tasks (05:20, 07:07) ### Expert Skepticism - Daniel Rein suggests measurement noise and scaffolding issues impact results, noting GPT-5.3-Codex's coding performance might be an outlier (06:50, 08:27) - Dean W. Bal expresses decreasing confidence in the benchmark due to saturation (06:04) ### Economic Implications - Citrini Research models a 2028 crisis where AI consumes the entire economy, leading to mass unemployment (07:25, 08:00) ### Methodology Concerns - METR acknowledged initial scaffolding/format issues hurt performance, noting their own partial measurements used different scaffolds than the final results (06:18) ### Community Reaction - The Citrini piece generated significant buzz, with some calling the chart "absolutely ballistic" (05:47), while others questioned whether the rapid gains are sustainable or accurate (08:51) ![Screenshot at 00:05: METR's chart illustrating "Moore's Law for AI agents," showing the doubling time for task length has been consistently around every 7 months over the past six years, which sets the benchmark for comparison against newer models.](https://ss.rapidrecap.app/screens/dztw1yctjI4/00-00-05.jpg) ![Screenshot at 04:05: METR's tweet detailing Claude Opus 4.5's 50% time horizon of around 4 hours 49 minutes, noting this was the highest published time horizon to date, emphasizing the large jump in capability.](https://ss.rapidrecap.app/screens/dztw1yctjI4/00-04-05.jpg) ![Screenshot at 05:26: METR tweet showing Claude Opus 4.6 achieving a 50% time horizon of around 14.5 hours, noting this measurement is extremely noisy because the current task suite is nearly saturated.](https://ss.rapidrecap.app/screens/dztw1yctjI4/00-05-26.jpg) ![Screenshot at 06:18: METR tweet detailing how scaffolding/format issues can hurt performance evaluation, suggesting models might be more sensitive to the scaffolding being used.](https://ss.rapidrecap.app/screens/dztw1yctjI4/00-06-18.jpg) ![Screenshot at 06:50: David Rein's tweet questioning the gospel-like acceptance of the metrics, emphasizing that the measurement is extremely noisy and the task distribution used was slightly different.](https://ss.rapidrecap.app/screens/dztw1yctjI4/00-06-50.jpg)