Day 944 — The Intelligence Factory Is Now Open For Business
The machines aren't coming for our jobs. They're already doing it.
Something shifted in the last two months, and it wasn’t subtle.
I’ve been watching AI capability announcements for long enough that I have a baseline for what “impressive” looks like. But the period from roughly mid-March to the end of May 2026 felt different — not because of any single release, but because of the cadence. The frontier labs didn’t take turns anymore. They released simultaneously, iterating within weeks of each other, each one pushing a benchmark the others had to scramble to beat.
This is the edition where I try to tell you what actually happened — and what it means for the people this newsletter exists to serve.
A note before we go further. This newsletter exists because I think the worst thing you can do right now is look away. I’m not writing from a position of safety — I’m writing from inside the problem. If you follow the paid edition, you’re watching one person, in real time, trying to figure out how to escape the gravitational pull of AI-driven job displacement before it becomes inescapable. Some weeks that looks like progress. Some weeks the black hole just seems bigger. Either way, I’m reporting back honestly. That’s the deal.
The Model Release Log Reads Like a Sprint Finish
In the twelve weeks ending in late April, the three frontier labs produced what may be the densest capability improvement window in the short history of generative AI.
Google’s Gemini 3.1 Pro arrived in February, posting a score of 94.3% on the GPQA Diamond scientific reasoning benchmark, more than double its predecessor’s score on ARC-AGI-2, the notoriously difficult visual reasoning test. A single model generation, roughly doubled on a hard benchmark. That alone would normally be a headline story for a month.
It wasn’t the story for a month. Because in March, OpenAI shipped GPT-5.4, which its own CEO described not as a smarter version of the last model but as something structurally different: a single model that credibly leads across coding, computer use, reasoning, and general knowledge work simultaneously, and all without specialist variants. Six weeks later, GPT-5.5 followed. OpenAI’s president Greg Brockman called it “a new class of intelligence.” As an illustration, he cited a mathematics professor who used it to build an algebraic geometry application from a single prompt in eleven minutes, work that would previously have taken weeks of manual coding.
At some point, “keeping up with AI developments” became a full-time job for people whose actual job was something else entirely.
The Benchmark That Actually Matters
I want to draw your attention to one data point in particular, because it cuts through the noise better than anything else I’ve seen.
In late 2025, OpenAI published a benchmark called GDPval. Unlike most AI benchmarks, which measure abstract reasoning or academic-style tests, GDPval was designed to answer a specific question: can AI do the work that the economy actually pays people to do?
The benchmark covers 1,320 tasks drawn from 44 real occupations across the top nine sectors of the US economy. Critically, every task was constructed from actual work products created by human professionals with an average of 14 years of experience. The graders were the same professionals, comparing AI and human outputs blindly, i.e., they didn’t know which was which.
Early models were already approaching expert-level quality on many of these tasks. By the time GPT-5.4 shipped in March, it was scoring 83% on GDPval, placing it at or above the level of human experts on economically valuable tasks across a wide range of knowledge work occupations.
One AI researcher, Julian Schrittwieser, a key contributor to AlphaGo and AlphaZero at Google DeepMind, extrapolated the consistent performance improvement trend and issued a straightforward prediction: by mid-2026, models will be capable of working autonomously for full eight-hour work days. By the end of 2026, at least one model will match expert-level human performance across many industries. By the end of 2027, models are predicted to frequently outperform human experts on many tasks.
This is not a fringe view. It is a straightforward extrapolation from a line on a graph that has been moving consistently in one direction.
The Adoption Gap: More Alarming Than It Sounds
In March, Anthropic published what I think is the most important piece of AI-and-labour research so far this year.
The study, authored by economists Maxim Massenkoff and Peter McCrory, introduced a new way to measure AI’s labour-market impact. Rather than relying on surveys or theoretical exposure models, they used actual usage data from Claude and real work tasks being performed to map both what AI can do and what it is doing.
The headline finding was this: actual AI adoption is a fraction of what AI tools are feasibly capable of performing.
AI can theoretically cover most tasks in business and finance, management, computer science, law, and office administration. In most of those sectors, actual adoption — measured by real usage — is still a small slice of that capability. The researchers found no material difference yet in unemployment rates between workers in AI-exposed roles and those in less-exposed positions.
Peter McCrory, Anthropic’s head of economics, put it plainly: “We’re seeing little evidence of widespread job displacement so far.”
I want you to sit with the strange ambiguity of that sentence, because it is doing a lot of work. It is both reassuring and quietly terrifying, depending on how you read it.
If the gap between capability and adoption is large, as Massenkoff and McCrory found, then we are not witnessing displacement in real time. We are watching the loading screen. The displacement has not happened yet, not because AI can’t do the work, but because organisations haven’t yet fully reorganised around the fact that it can.
That reorganisation is the thing to watch. It is not a technological event. It is a management and procurement event. It happens when enough CFOs review their headcount and AI tool budgets in the same spreadsheet.
Who Is Already Feeling It
The absence of mass unemployment numbers does not mean nothing is happening. It means the distribution is uneven, and some groups are already on the receiving end.
Workers aged 22 to 25 are showing a 16% fall in employment in AI-exposed roles. That is the first cohort to have their early career formation coincide with the emergence of frontier AI capabilities. They are not being laid off in large numbers. They are simply not being hired in the first place. Entry-level roles are contracting. The jobs that used to be the bottom rung of a white-collar career ladder are the first to go, because they are the jobs most easily replicated by a capable AI with light human oversight.
A study from the IMF found that for occupations highly exposed to AI with limited scope for human complementarity, employment levels are 3.6% lower in regions with greater AI skill demand than in comparable regions five years after those skills appeared. The shape of this is: first the roles compress, then the headcount contracts, and the full labour market signal emerges years after the technological shift that caused it.
Meanwhile, ManpowerGroup’s 2026 Global Talent Barometer found something that I find deeply telling: regular AI usage among workers jumped 13% year-on-year to 45% of the workforce, but confidence in using technology fell sharply by 18%. Workers are being pushed into AI tools they are not confident using, while 43% fear automation will replace their jobs within two years. The same survey found 64% of workers planning to stay with their current employer despite feeling uncertain about the future, what ManpowerGroup called “job hugging”: clinging to stability rather than adapting, because the cost of getting the adaptation wrong feels too high.
The Skills Divergence Is Already Happening
Anthropic’s March report had a second finding that tends to get less coverage than the headline-displacement numbers, but which I think matters more to the people reading this newsletter.
While widespread displacement remains limited, the data reveals a growing divergence between early AI adopters and everyone else. Workers who are mastering AI tools are pulling ahead of their peers. The gap is widening. It is not theoretical; it is showing up in productivity metrics and career trajectory data.
This is the part that actually applies to you, right now.
The window in which AI proficiency is a differentiator, rather than table stakes, is not infinite. At some point, “knows how to use AI” stops being an edge and starts being the minimum bar for staying employed in knowledge work roles at all. We do not know exactly when that transition happens. But the benchmark data, the adoption curves, and the model release cadence all point toward it happening faster than most institutions are prepared for.
What I’m Watching Next
Two things that will tell us a great deal about the second half of 2026.
First: GPT-6 is expected sometime between May and July, with Claude 5 (internally codenamed “Fennec”) targeting a May-to-September window. If those models represent qualitative leaps rather than incremental benchmark improvements — the kind of shift that GPT-4 was in early 2023 — the adoption curve will steepen rapidly. Organisations that have been taking a “watch and wait” approach to AI integration will find the gap between what they’re doing and what their competitors could be doing suddenly and visibly uncomfortable.
Second: watch the enterprise software renewals. The most honest signal of AI-driven displacement is not layoff announcements; those lag reality by 12-18 months. The honest signal is the software procurement data: which vendors are growing, which are contracting, and what the headcount implications of those decisions turn out to be. When a company replaces a team of analysts with an AI-powered data platform, the story doesn’t appear in the headlines. It appears six months later, in the headcount line of a quiet earnings call.
The intelligence factory is open. It’s running three shifts. The question is not whether it will change your industry. It’s whether you’ll see it coming while there’s still time to do something about it.
— Olaf
If this resonated, share it with one person in your network who is still treating AI as a future problem. The data suggests it is a present one.