Understanding the Most Viral Chart in Artificial Intelligence
The Odd Lots podcast hosts Joe Weisenthal and Tracy Alloway interview Joel Becker and Chris Painter from Meter, a nonprofit AI safety organization, about their viral 'time horizon' charts that measure AI model capabilities. The discussion covers how these benchmarks are constructed, their limitations, the safety motivations behind them, and the broader tensions between AI development, investment incentives, and safety concerns.
Summary
The episode centers on Meter's 'time horizon' charts, which have become arguably the most viral visualization in AI. Hosts Joe Weisenthal and Tracy Alloway speak with Joel Becker (technical staff) and Chris Painter (president) of Meter, a San Francisco-based nonprofit focused on AI safety evaluation.
Meter's core mission is assessing AI autonomy and dangerous capabilities — specifically whether AI systems could assist in catastrophic activities like bioweapon development or large-scale cyberattacks. The time horizon charts were originally designed to answer a safety question: when would AI systems be capable enough that concerns about misalignment and loss of human control would become meaningful? The intuition was that three to four years ago, AI systems were so limited that fears about them 'going rogue' were almost nonsensical.
The time horizon metric works by having skilled human engineers complete the same software and machine learning tasks given to AI models, timing how long those tasks take. The difficulty of a task is expressed in human-hours. The AI's 'time horizon' is the task length at which it achieves a 50% success rate. As of early 2026, Claude Opus 4.6 achieved a time horizon of approximately 12 hours — nearly double the previous high of around 6 hours set by GPT codex models — causing the chart to go viral. The doubling time of AI capabilities, as measured by this metric, appears to have accelerated from roughly every 7 months to approximately every 4 months.
The hosts probe several nuances and potential criticisms of the charts. At the 80% success threshold rather than 50%, the progress looks less dramatic, though Joel argues the doubling rate is essentially the same — just offset. The 50% threshold is used partly for statistical reasons: it requires fewer samples to estimate reliably and matches conventions in prior literature. Joel also acknowledges that with only about three human baselines per task, the methodology has real limitations, and that baselining will become increasingly difficult as AI time horizons extend beyond months.
A key limitation discussed is the gap between benchmark performance and real-world productivity. The tasks measured are narrowly focused on software and machine learning engineering — the domain where AI labs are actively optimizing. Real-world tasks tend to be messier, involve larger codebases, require collaboration, and are evaluated more holistically than algorithmic scoring allows. These factors mean benchmarks likely overstate productivity gains somewhat, though Joel believes the underlying progress is real and generalizing.
On Chinese models, Meter has largely not included them on their primary charts because they appear to lag U.S. frontier models by roughly 9–12 months on time horizon metrics, and possibly even more than benchmark scores alone would suggest — with Chris hinting that Chinese models may perform better on benchmarks than on truly held-out problems, something 'spiritually close' to benchmark gaming.
The conversation also explores the strange sociological dynamic of the AI industry, where the people most enthusiastic about building AI are also often the most alarmed about its risks. Chris draws on the analogy of the Manhattan Project, noting that many early AI safety researchers got into the field precisely because they saw deep learning trends and worried about what full artificial general intelligence or superintelligence could mean. The competitive dynamics — between labs, and between the U.S. and China — create a situation where no individual actor feels able to slow down unilaterally, even if they wanted to.
On investment implications, Meter does not engage much with investors directly, but Chris argues that broad public awareness of AI capabilities is actually preferable to selective knowledge — he would rather all of humanity understand where AI is heading than have only certain actors informed. However, he acknowledges tension between financial commitments (like massive data center investments already baked in) and the ability to slow development if safety risks emerged.
Meter itself has about 30 people, is growing, and operates as a nonprofit unable to offer equity. They compete on cash compensation and attract people motivated by working on uniquely important, public-facing research outside the competitive lab environment. Joel closes by noting the team is in a state of triage, identifying 20–30 critical research questions but only able to address one or two per quarter.
About this episode
<p>We live in an era of charts that are going up and to the right. This image obviously describes the stock market, particularly any company whose business is adjacent to artificial intelligence. But beyond stocks, another sort of chart we keep seeing is of AI capabilities also going up and to the right. The most famous and viral of these comes from an organization called METR, which stands for Model Evaluation and Threat Research. The organization is focused on understanding the degree to which AI models can engage in autonomous, complex tasks. METR see this is as a particularly important benchmark, given the risk that AI could one day be engaged in recursive self improvement, taking humans out of the loop. But how do you really gauge a model's ability to do complex problems. And what is being measured for exactly? On this episode, we speak with METR's President Chris Painter as well as Joel Becker, a member of the technical staff who works on evaluation methods for the organization. We discuss both the mechanics and the philosophy of METR's work, and what it means when we see a a chart showing that Clause Opus 4.6 can do a task that would take a human nearly 12 hours.</p> <p>Read more:<br /><a href="https://www.bloomberg.com/news/articles/2026-04-24/deepseek-unveils-newest-flagship-a-year-after-ai-breakthrough?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article">DeepSeek Unveils Flagship AI Model a Year After Breakthrough</a><br /><a href="https://www.bloomberg.com/news/articles/2026-04-24/meta-inks-multibillion-dollar-deal-to-use-amazon-chips-for-ai?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article">Meta Inks Deal to Use Amazon’s Graviton Processors for AI</a></p> <p>Only <a href="http://Bloomberg.com">http://Bloomberg.com</a> subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast">bloomberg.com/subscriptions/oddlots</a></p> <p><a href="http://bloomberg.com/subscriptions/oddlots">Subscribe to the Odd Lots Newsletter</a><br /><strong>Join the conversation:</strong> <a href="https://discord.gg/oddlots">discord.gg/oddlots</a></p><p>See <a href="https://omnystudio.com/listener">omnystudio.com/listener</a> for privacy information.</p>
Key Insights
- The time horizon metric does not measure how long an AI works continuously, but rather the difficulty of tasks — expressed in human-hours — at which the AI achieves a 50% success rate, based on timing skilled humans doing the same tasks.
- Claude Opus 4.6 achieved a time horizon of approximately 12 hours as of early 2026, nearly doubling the previous high of ~6 hours, which is what made the chart go viral — the doubling time of AI capabilities appears to have accelerated from ~7 months to ~4 months.
- The 50% success threshold is used rather than 80% or higher partly for statistical reasons: estimating reliability at very high success rates requires far more samples and is highly sensitive to grading noise, making 50% the most statistically tractable point.
- Joel Becker argues that benchmark performance likely overestimates real-world productivity gains because real tasks are messier, involve larger codebases, require collaboration, and are evaluated more holistically than algorithmic scoring captures.
- Chris Painter argues that broad public awareness of AI capability trends is preferable to selective knowledge, framing Meter's mission as informing all of humanity — including investors and governments — rather than gatekeeping information to any particular group.
- Chinese AI models have been excluded from Meter's primary charts because they appear to lag U.S. frontier models by roughly 9–12 months on time horizon metrics, and Chris suggested their benchmark scores may overstate their capabilities relative to truly held-out problems.
- Chris identifies a key tension between large financial commitments — such as debt taken on to build data centers — and the ability to slow AI development if safety concerns emerged, arguing these obligations could force continued scaling even against better judgment.
- Meter operates in a state of triage with roughly 30 staff, identifying 20–30 world-important research questions per quarter but only able to address one or two, and Joel argues the primary bottleneck is technical talent rather than model access, as AI labs have generally been cooperative with third-party evaluation.
Topics
Transcript
The thing about AI for business, it may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR, IT, and procurement processes, we've reduced costs by millions, slashed repetitive tasks, and freed thousands of hours for strategic work. Now we're helping companies get smarter by putting AI where it actually pays off, deep in the work that moves the business. Let's create smarter business. IBM. Data rates may apply. JPMorgan Chase Bank, N.A. Member FDIC. Copyright 2026. JPMorgan Chase & Company. Next reward. Choose from more. More than 1,100 hotels across 13 distinctive brands and unlock the best available rates when you book direct with Synesta Travel Pass. Here today.…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Odd Lots
The Rise of Organized Retail Crime at Big Box Stores
Scott Glenn, VP of Asset Protection at Home Depot, discusses the rise of organized retail crime (ORC), explaining how it differs from opportunistic shoplifting, how stolen goods are fenced through online marketplaces, and the technological and legal challenges retailers face in combating sophisticated theft operations that often involve drugs, violence, and organized criminal networks.
Why Money Launderers Love $100 Bills
Bloomberg's Odd Lots podcast interviews journalist Oliver Bullough about money laundering, exploring how $2-5 trillion annually is laundered globally through cash, trade-based schemes, and crypto. The discussion reveals that despite $200 billion spent annually on anti-money laundering compliance, criminals consistently outmaneuver regulations while governments continue printing high-denomination bills that enable criminal activity.
Why Laser Beams Are the Hottest New Tech in Defense
Colonel Wayne Sanders from Bloomberg Intelligence discusses directed energy weapons (lasers and high-powered microwaves) as emerging defense technologies that could reduce reliance on expensive munitions. The conversation covers how these systems work, their deployment challenges, supply chain constraints, and their role in a layered air defense strategy alongside traditional missile systems.
What's Behind the Big Surge in US Government Bond Yields
Economics professor Daryl Duffy explains that rising US Treasury yields are driven primarily by massive government debt issuance relative to limited investor demand, rather than inflation concerns. He discusses the Treasury's buyback program, the Fed's balance sheet challenges, and the structural constraints on both fiscal and monetary policy's ability to control yields.
Adam Posen Thinks Things Could Get Very 'Messy' for the Fed
Adam Posen discusses Fed Chair Kevin Warsh's Jackson Hole speech, giving it a B- grade while highlighting concerns about the Fed's inflation trajectory and Warsh's preference for discretionary policymaking over clear forward guidance. Posen argues the Fed was late to hike in 2022 and early to cut in 2023, and warns that without explicit commitment to bringing inflation down at a specific pace, markets and policymakers may face confusion about the Fed's true intentions.