# The Penny Still Looks Like Nothing By Dave Glaser There’s a question most of us heard in grade school: would you rather take one million dollars today, or take a penny that doubles every day for thirty days? The million feels like the grown\-up answer\. The penny feels like a trick, and for most of the month it keeps looking like one\. On Day 30 alone, the penny pays $5,368,709\.12\. Add up all thirty payments and you get $10,737,418\.23\. But that isn’t the interesting part\. What matters is when it happens\. The daily payment doesn’t pass one million dollars until Day 28\. The running total doesn’t pass it until Day 27\. On Day 20, three weeks into the experiment, the check is only $5,242\.88\. If you were watching this unfold one day at a time, you’d have three weeks of evidence that the penny was a bad bet\. You wouldn’t be foolish\. You’d be reasoning honestly from what you could see, and you’d still be wrong because nearly all the money arrives at the end\. Nothing changes about the doubling\. It was doubling on Day 2, when nobody cared\. The base finally gets large enough for ordinary people to feel what the exponent was doing all along\. That’s the part that feels familiar now\. The penny is an analogy, not evidence\. It has a known rate and a known ending\. The real world has neither, and real exponential curves hit limits\. But the psychological trap is real\. When something compounds, the period when it looks manageable and the period when it plainly isn’t can be separated by just a handful of doublings\. So the question I keep coming back to isn’t simply whether AI is improving\. It’s which day we’re standing on\. ![](/writing/which-day/the-penny-still-looks-like-nothing/assets/hero.png) What’s actually compounding The loudest version of this argument is often wrong\. Today’s AI agents generally don’t rewrite their own model weights while they work\. The model is frozen\. When an agent appears to learn, it usually wrote something to memory, retrieved an old result, built a reusable tool or got replaced by a newer model\. That distinction matters, but systems can compound without rewriting their own brains\. Memory sticks\. Code sticks\. A useful workflow can be copied across a team for almost nothing\. Large models generate training examples and help teach smaller, cheaper ones\. Automated evaluations shorten the distance between an idea and a verdict\. And engineers at frontier labs are already using today’s models to help build tomorrow’s models\. The more consequential version is a research loop: generate a hypothesis, search, draft, formalise, verify, let an expert judge what matters, then feed the result into the next round\. Formal proof systems do not decide whether a question was worth asking or whether a result changes the real world\. But where a claim can be checked mechanically, they can turn plausible model prose into an auditable object and let experts spend their time on the harder judgement\. That points the compounding directly at the source\. The system does not need to rewrite its own weights while it works\. It needs to make the interval between possibility and checked knowledge shorter, then use what it learns to improve the system running the next cycle\. We can measure part of this\. METR found that the length of software and reasoning tasks AI agents could complete with 50% reliability doubled roughly every seven months from 2019 through 2025\. Its January 2026 update put the post\-2023 doubling at 131 days, though the exact number depends heavily on the task set\. \[1\] The International AI Safety Report described the same shift more plainly: coding agents can now reliably finish some tasks that would take a human programmer about half an hour, up from less than ten minutes a year earlier\.\[2\] In July 2025, systems from OpenAI and Google DeepMind each solved five of six problems at the International Mathematical Olympiad, crossing the gold\-medal line for the first time\.\[3\] None of this proves a singularity is coming\. It establishes something smaller and harder to wave away: important inputs to AI progress, including AI’s contribution to that progress, are improving on measured doubling curves\. The doubling times are counted in months, not generations\. The forecasts keep moving Ask the people closest to this work which day we’re on and you’ll get answers separated by decades\. Sam Altman has said takeoff has begun\. Dario Amodei has described a possible “country of geniuses in a datacenter” within a few years\. Demis Hassabis has pointed toward the end of this decade\. Andrew Ng says the question is overhyped\. Yann LeCun argues that human\-level AI will require architectures nobody has built yet\. The disagreement is real\. We shouldn’t smooth it over to make the story cleaner\. But the forecasts have generally moved earlier after each capability jump\. A survey of 2,778 AI researchers put the 50% probability point for machines outperforming humans at every task in 2047\. That was thirteen years earlier than the answer from the same survey one year before\.\[4\] That’s how a sensible person watching the penny would behave\. Each new number is larger than expected, so the forecast gets revised upward\. But the revision still trails the curve because it begins with what already happened\. This doesn’t mean the shortest timeline is right\. Vernor Vinge predicted superhuman intelligence by 2023\. It didn’t happen\. Confident dates deserve humility\. What bothers me is that our forecasting process is reactive\. We wait for a capability to become visible, accept that it happened, then adjust\. That works in ordinary life\. It’s dangerous in a transition where arriving late might leave no time to recover\. So which day are we on? I don’t think we’re on Day 27\. Nothing deployed today can simply seize control from humanity, and the field’s most careful public assessment says that plainly\. But it also says that if progress is extremely rapid, it’s impossible to rule out AI developing capabilities sufficient for loss of control within several years\.\[5\] That’s where I land\. Not on a date\. On a rate\. I use these systems every day, and I know how much they still get wrong\. They lose context, misunderstand instructions and confidently take the long way around simple problems\. Some days they look less like superintelligence and more like a talented new employee who misplaced the keys\. I’ve also watched one useful workflow turn into five, watched an agent turn a lesson into code every later agent could use, and watched tasks that once required a person at a keyboard become background work\. The changes don’t feel historic\. They feel like small improvements and another Tuesday\. That may be why they’re easy to underestimate\. The person on Day 20 isn’t stupid\. Their observations are accurate\. Their mistake is assuming the next part of the curve will resemble the part they’ve already seen\. I don’t know which day we’re on\. I know the measured doubling times are months rather than decades\. I know the forecasts keep moving in one direction\. And I know that on Day 20, the check is for $5,242\.88 and everything still looks fine\. If we can’t tell which day we’re on, what should we be doing differently today? \-\-\- \[1\]: METR, “Measuring AI Ability to Complete Long Tasks,” March 2025 (https://metr\.org/blog/2025\-03\-19\-measuring\-ai\-ability\-to\-complete\-long\-tasks/) and “Time Horizon 1\.1,” January 2026 (https://metr\.org/blog/2026\-1\-29\-time\-horizon\-1\-1/)\. \[2\]: International AI Safety Report 2026, Extended Summary for Policymakers (https://internationalaisafetyreport\.org/publication/2026\-report\-extended\-summary\-policymakers)\. \[3\]: Google DeepMind, “Gemini with Deep Think achieves gold\-medal standard at the International Mathematical Olympiad,” July 2025 (https://deepmind\.google/blog/advanced\-version\-of\-gemini\-with\-deep\-think\-officially\-achieves\-gold\-medal\-standard\-at\-the\-international\-mathematical\-olympiad/)\. \[4\]: Grace et al\., “Thousands of AI Authors on the Future of AI,” January 2024 (https://arxiv\.org/abs/2401\.02843)\. \[5\]: International AI Safety Report 2025 (https://internationalaisafetyreport\.org/publication/international\-ai\-safety\-report\-2025)\.