โ๐๐คฆ Failing to Understand the Exponential, Again
๐ค AI Summary
- ๐ AI discourse regarding a โbubbleโ parallels the failure to grasp Covid-19โs exponential spread.
- ๐ฆ Commentators missed the pandemicโs scale, treating it as remote after exponential trends became obvious.
- ๐ซ Model mistakes prompt conclusions that AI will never reach human-level performance or will have only minor impact.
- ๐ Lack of conversational change across model releases suggests AI is plateauing and scaling is over.
- ๐ป The METR study documents a clear exponential trend in autonomous software engineering task length.
- ๐ Sonnet 3.7 achieved 50% success on one-hour tasks; recent models exceed this, completing tasks over 2 hours.
- ๐ผ OpenAIโs GDPval study measures performance across 44 occupations in 9 industries.
- ๐ Evaluation shows a similar trend, with GPT-5 nearing human performance.
- ๐ฅ Claude Opus 4.1 significantly outperforms GPT-5, almost matching industry expert performance.
- โณ Models will autonomously work 8-hour days by mid-2026.
- ๐งโ๐ฌ At least one model will match human experts across many industries before late 2026.
- ๐ง By late 2027, models will frequently outperform experts on many tasks.
- โ ๏ธ Grok 4 and Gemini 2.5 Pro underperformance is notable given previous state-of-the-art claims.
๐ค Evaluation
- โ๏ธ The analogy to the Covid-19 pandemic is contrasted with the mechanism of AI progress by commentators.
- ๐ฆ For COVID-19, the spread of infection is an understood, deductive exponential process, unlike the fuzzier process underlying the AI boom.
- โ๏ธ AI improvement is considered closer to Mooreโs law, which depends on the whole industry focusing on new innovations, suggesting improvements are not inevitable.
- ๐ The METR and GDPval tasks are contrasted with real-world work by being characterized as not โmessy.โ
- ๐ METRโs benchmark tasks have a mean messiness score of ~3/16, while a regular software engineering task is 7-8, suggesting current evaluations do not capture the large variety of real-world work.
- ๐ฎ A legitimate perspective posits that AI may be able to perform non-messy tasks for eight hours at a 50% success rate and outperform experts, yet somehow fail to replace anyone, similar to the introduction of technology to radiologists.
- ๐ค The topic to explore for better understanding is the creation of evaluations that include both messier and longer horizon tasks.
- ๐ ๏ธ Another topic to explore is how to best structure the human-AI collaboration necessary for ultra-high productivity, where AI functions as a very smart tool rather than a replacement.
๐ Book Recommendations
๐ก Similar
- ๐คโ ๏ธ๐ Superintelligence: Paths, Dangers, Strategies by Nick Bostrom. This book explores the risks and implications of an intelligence explosion, directly addressing the possibility of a rapid, exponential AGI takeoff often missed by linear thinking.
- ๐ค๐งฌโฌ๏ธ The Singularity Is Near: When Humans Transcend Biology by Ray Kurzweil. The seminal text arguing that technological change is accelerating exponentially, leading to a โsingularityโ where AI surpasses human intelligence, reinforcing the articleโs core theme.
- ๐ค๐ The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies by Erik Brynjolfsson and Andrew McAfee. It explores how digital technologies, particularly AI, are transforming labor and society, supporting the argument for rapid, exponential change.
โ๏ธ Contrasting
- ๐ง ๐ง ๐ง ๐ง A Thousand Brains: A New Theory of Intelligence by Jeff Hawkins. This work offers a biologically-grounded theory of intelligence suggesting current AI architectures may be fundamentally flawed or missing key elements, offering a potential plateauing mechanism that contrasts the articleโs optimism.
- ๐ Poverty, by America by Matthew Desmond. A deeply contrasting societal analysis that focuses on the distributional failures of a wealthy society, forcing a necessary consideration of where exponential technological gains might fail to solve fundamental human problems.
- ๐ The Myth of the AI Revolution by Kate Crawford. This work offers a critical, structuralist perspective, arguing that AI is not a disembodied intelligence but a system built on vast resources and political choices, suggesting a slower, more complex path to โrevolutionโ that contrasts the articleโs simple extrapolation.
๐ก Creatively Related
- โซ๐ฆข๐ฒ The Black Swan: The Impact of the Highly Improbable by Nassim Nicholas Taleb. This book discusses the impact of highly improbable, high-impact eventsโlike a sudden AGI breakthroughโthat are inherently unpredictable but reshape history, illustrating why the future is not merely an extrapolation of the past.
- ๐ฅ The Shock of the New by Robert Hughes. A history of modern art that deals with how society adaptsโor fails to adaptโto relentless, high-velocity change in culture and technology, linking the emotional and cultural impact of exponential progress.
- ๐๐๐ช๏ธ๐ช Antifragile: Things That Gain from Disorder by Nassim Nicholas Taleb. This book discusses systems that not only withstand shocks but benefit from them, providing a framework for how individuals and institutions can prepare for a future shaped by unpredictable, exponentially growing technologies.
๐ฆ Tweet
Interesting post!
โ Bryan Grounds (@bagrounds) September 29, 2025
An AI generated counterpoint:
๐ฆ For COVID-19, the spread of infection is an understood, deductive exponential process, unlike the fuzzier process underlying the AI boom.
Full AI Summary, evaluation, and book recommendations here: https://t.co/zrMKtXtRFc
๐ฆ Bluesky
โ๐๐คฆ Failing to Understand the Exponential, Again
AI Q: ๐ Is AI progress an unstoppable wave or a temporary bubble?
๐ค Machine Intelligence | ๐ Scaling Benchmarks | ๐ผ Labor Automation
โ Bryan Grounds (@bagrounds.bsky.social) 2026-07-16T11:45:32.000Z
https://bagrounds.org/articles/failing-to-understand-the-exponential-again