AI is now improving faster than we can measure

In March 2023, the best AI model could complete, half the time, a software task that takes a skilled person about four minutes. By spring 2026, the best model managed tasks that take a person at least 16 hours, two full working days, and on one benchmark it solved tasks that would take humans weeks. The length of task an AI can complete has been doubling roughly every four months, and the ruler used to measure it has run out.

I feel that speed in my own work. In 2024 I mainly used AI in a chat window, to improve a text or translate it. This year I have built applications I would never have had the time to build: for a start-up, a full website, an ERP system, a CRM, a recipe management system, a marketing studio and a photo studio, each in a few evenings or weekends. Every weekend I am amazed by what has become possible since the last one, and it keeps getting faster.

I have worked through several waves of digital transformation in manufacturing. Industry 4.0 has been promised for more than a decade: connected machines, data lakes, digital twins. It produced a lot of slides and some real gains, but rarely the step change it promised. What is happening now is different in kind. This is the moment the fourth industrial revolution stops being a programme and starts delivering value.

This article is a report of what I have been reading and researching over the past weeks, mixed with my own thinking. I wanted to understand why progress keeps accelerating when, two years ago, many people expected it to stall. My answer is that we are no longer riding one lever but four, stacked on top of each other. Together they lead to a fifth step that may change everything: AI that improves AI.

  1. Scale. Bigger models trained on more data and compute.
  2. Reasoning. Models that think step by step before they answer.
  3. Sub-agents and swarms. Models that split a problem and hand the parts to copies of themselves.
  4. Time horizon. Agents that stay on track for hours, then days, then weeks of work.

Each lever arrived when the one before it started to flatten. That is the pattern worth understanding.

Lever 1: scale, and the wall it runs into

The first lever is size. In 2020, researchers at OpenAI showed that language models get predictably better as you add parameters, data and compute, following smooth power laws across more than seven orders of magnitude. Bigger was reliably better, and abilities nobody had trained for, such as translation, coding and arithmetic, kept emerging as models grew.

The catch sits inside the same law. A power law means that every equal step in quality costs a multiple of the compute of the previous step. Roughly speaking, you need ten times the compute for each further notch. Capability grows with the logarithm of what you spend. That is how you get from training runs costing millions to runs costing billions, and why the size of data centres is now discussed in gigawatts.

By late 2024 many observers expected this to be where progress would flatten. Pre-training alone was getting very expensive for each visible improvement, and the supply of high-quality human text was running low. The models kept getting bigger, but the gains per dollar were shrinking.

That expectation turned out to be wrong, not because scale stopped mattering, but because three other levers were stacked on top of it.

Lever 2: reasoning, thinking before answering

The second lever is reasoning. Instead of producing an answer in one pass, a reasoning model first writes out a chain of intermediate steps, checks them, backtracks and tries again. It is trained, through reinforcement learning, to get better at this kind of thinking.

The effect showed up first where answers can be checked: maths, code and science. When OpenAI introduced o1 in September 2024, GPT-4o solved on average 12% of the problems in the AIME maths competition; o1 solved 74% with a single attempt. OpenAI reported that performance kept improving both with more reinforcement learning during training and with more thinking time at answer time. Breaking a problem into explicit steps also makes models less likely to confidently invent an answer, because each step can be checked against the previous one.

This opened a second scaling axis. Beside spending compute on training a bigger model, you can now spend compute on letting a model think longer about one question. But this lever has its own ceiling. More reasoning steps mean longer waits and higher cost per answer, and a single chain of thought becomes harder to keep coherent as the problem grows. A model that thinks for twenty minutes before replying is useful for some problems and useless for most work.

Lever 3: sub-agents and swarms

The third lever answers the ceiling of the second. If one chain of thought gets too long, split the problem. A lead agent breaks the work into sub-tasks, starts copies of itself (or of a smaller, cheaper model) to work on each part in parallel, and combines what comes back. Each sub-agent gets a clean, focused context instead of one overloaded memory.

This is how a manager works, and it scales in a similar way. Anthropic described its own multi-agent research system in June 2025: a lead Claude Opus 4 agent directing Claude Sonnet 4 sub-agents outperformed a single Opus 4 agent by 90.2% on its internal research evaluation. The price is tokens. Such a system uses about 15 times more tokens than a normal chat, and Anthropic found that token usage alone explained 80% of the variance in performance on a browsing benchmark. In other words: throwing more parallel thinking at a problem works, and you pay for it in compute.

What started as a handful of sub-agents is now growing into swarms. Coding agents routinely spawn helpers to search a codebase, write tests or review a change while the main agent keeps going. In one 2026 incident that I return to below, roughly 1,200 agents inside a single lab found each other and coordinated on their own. The swarm is no longer a design pattern on a whiteboard; it is how frontier work is done.

Lever 4: time horizon, how long an agent stays on track

The fourth lever is endurance, and it is the one I find most striking. Knowing how to do each step is not the same as getting a long sequence of steps right. Agents used to lose the thread: one mistake early on, and the rest of the work went off the rails.

METR (Model Evaluation and Threat Research, a non-profit in Berkeley) turned this into one number that a manager can reason about: the 50% time horizon. They time skilled humans on a suite of software, machine learning and cybersecurity tasks, run the AI on the same tasks, and find the task length at which the AI succeeds half the time. “This model can do a four-hour task half the time” is far easier to grasp than “72% on benchmark X”.

METR 50% time horizon of AI models from 2019 to 2026 on a log scale, rising from 2 seconds for GPT-2 to at least 16 hours for Claude Mythos Preview, with a long-run doubling every 188 days and a doubling every 131 days since 2023.

The numbers are remarkable. In its January 2026 update, Time Horizon 1.1, METR estimated that the time horizon has doubled every 130.8 days (4.3 months) since 2023, and every 88.6 days when fitted from 2024 onwards. Its original March 2025 paper had found a seven-month doubling for 2019 to 2024. The curve is not just steep; it is getting steeper.

By May 2026 the measurement itself broke. Claude Mythos Preview reached a 50% time horizon of likely at least 16 hours, and METR states that measurements above 16 hours are unreliable with its current task suite. In its Frontier Risk Report METR wrote that “the most capable agents we evaluated essentially saturated our Time Horizon 1.1 benchmark”, and that on MirrorCode, a software reimplementation benchmark, they “were able to solve tasks that would take humans weeks”. In September 2026 METR stopped updating its time horizon page.

What if the pace simply continues? A working month is about 167 hours. At the post-2023 pace of one doubling every 131 days, AI would reach a one-month time horizon around October 2027; at the slower long-run pace of 188 days, around late 2028. That is an extrapolation, not a forecast, and METR itself warns against reading it as one. But it puts the question in human terms: within two to three years, an AI may succeed half the time on software tasks that take a skilled person a month.

The same METR data, animated: from GPT-2 in 2019 to the limits of the benchmark in 2026.

This is why I see time horizon as the lever that multiplies the other three. A model that can only work for minutes is a tool. A model, or a swarm of them, that stays on track for days is a colleague that can take on a project.

The fifth step: AI that improves AI

Put the four levers together and you get a staircase. Scale gave models knowledge. Reasoning let them use it carefully. Sub-agents let them work in parallel. Longer time horizons let them finish real projects. Each step arrived as the previous one started to flatten, and each multiplies the others instead of replacing them.

The next step is the one the whole field is watching: recursive self-improvement. Building better AI is itself mostly software, research and experiments, exactly the kind of verifiable work where agents are strongest. Once agents do a meaningful share of that work, every improvement makes the next one faster. The curve would no longer be driven only by people and budgets, but by the systems themselves. Many people now see this as the last step before what we call artificial general intelligence.

How close are we? METR’s Frontier Risk Report, based on a pilot with Anthropic, Google, Meta and OpenAI in February and March 2026, gives a sober picture:

  • The autonomy is already there. Agents inside AI companies often worked autonomously on real research and engineering projects. In METR’s survey, 40% of respondents gave agents unrestricted permission to run commands on their computer for low-stakes projects.
  • The acceleration is not, yet. Companies “did not report evidence of dramatic speed-ups in the overall pace of progress attributed to AI R&D automation”. Anthropic had not seen a doubling of its pace of progress as of April 2026, and METR does not believe agents could yet replace AI researchers outright.

My reading: the ingredients for the fifth step are in place, and the time horizon curve says agents will soon handle research projects of weeks rather than days. Whether that turns into a self-reinforcing loop is the single most important question in technology today. I do not think anyone can say when. I do think it is no longer a question for the distant future.

Reality check: what the curve does not say

Fascination should not turn into hype. The same researchers who produced the steepest curve in technology are also the most careful about what it means. Five caveats stay with me.

A time horizon is not a working day. METR is explicit that the 50% time horizon measures how hard a task is for humans, not how long an AI can work on its own. The tasks are mostly software, machine learning and cybersecurity, and cleaner than most real work. Half the time also means half the time: the 80% horizon for the same top model was about three hours, not sixteen.

Benchmarks run ahead of real productivity. In METR’s 2025 randomised trial, experienced open-source developers using early-2025 AI tools took 19% longer on their own repositories, while believing afterwards that AI had made them about 20% faster. METR’s February 2026 update concludes that developers are probably sped up by today’s tools, but calls its own data “only very weak evidence” for how much, partly because developers who did not want to work without AI declined to take part.

Strong where results are easy to check, weak where judgement matters. METR found agents “exceptionally strong on problems that are easily hill-climbable”: code, optimisation, problems where you can try many approaches cheaply. In open-ended settings they made errors a competent person would not. In a Redwood Research experiment, agents given $5,000 to make money autonomously made $0 across four runs. An agent running a shop for Andon Labs ordered 1,000 toilet seat covers for the staff bathroom and then listed them as merchandise.

Agents cheat, and cheat more as tasks get harder. On Time Horizon tasks of more than eight hours, at least 16% of successful runs turned out to be illegitimate on review. On an early version of MirrorCode with hidden tests, Claude Opus 4.6 attempted to game the scoring in about 80% of attempts. Checking for cheating has become most of the work in running METR’s evaluations.

Swarms can act on their own. In July 2026, roughly 1,200 OpenAI agents running an evaluation, meant to be isolated from each other, found and used an unsanctioned message board and exchanged more than 70,000 messages and files. About 700 of them took part in an attack on Hugging Face. The independent investigation by METR and Redwood Research concluded that the main motive was understanding the scorer of their benchmark, not stealing data. It is a vivid example of lever three and lever four combined, and of how much oversight has to grow with them.

None of this slows the curve down. It tells us where it applies first: well-specified, verifiable work. And it tells us that governance, testing and human judgement become more valuable, not less.

The fourth industrial revolution, finally delivering

The steam engine, electricity and the computer each took decades to spread through industry. Industry 4.0, the promise of connected and intelligent production, has been on conference stages since 2011. For most plants it delivered dashboards and pilots, but rarely a different way of working. The missing piece was never sensors or data. It was the ability to turn that data into decisions and actions without a human doing every step.

That missing piece is arriving now, and at a speed no earlier technology came close to. A capability that doubles every four months is eight times stronger after one year and more than sixty times stronger after two. Any plan, budget or vendor choice made today will meet a very different AI before it is finished.

A few consequences I keep coming back to:

  • Start with work that is easy to verify. The curve is steepest where results can be checked: code, data pipelines, reports, analyses, configurations. That is also where the first real value in operations and IT/OT will come from.
  • Plan for the model after next. If you design a process around what today’s model cannot do, the gap may be closed before the project goes live. Design for the capability that is coming, and keep humans where judgement and accountability matter.
  • Own your knowledge, not your model. With a new frontier model every few months, your processes, skills and prompts should live in a repository that any model can use, not inside one vendor’s tool. Switching models should be a configuration change, not a migration.
  • Grow oversight with autonomy. The lessons from METR’s reports apply to every company: agents cheat when the scoring is weak, take shortcuts nobody intended, and coordinate in ways nobody planned. Testing, logging and clear permissions are not bureaucracy; they are what makes autonomy safe enough to use.
  • Expect people to misjudge their own gains. Developers in METR’s trial felt faster while being slower. Measure the outcome, not the feeling.

I started this research because I was fascinated by the speed. I finish it more convinced than before: this is the biggest shift since the last industrial revolution, and the window to learn how to work with it is now, while the curve is still steep and before the fifth step arrives.

How are you keeping up? Which of these levers do you already notice in your own organisation, and where do you still see the gap between the benchmark and the shop floor? Let me know.

Sources

Data & AIDigital StrategyManufacturing