Opus 5.5 is already doing work people were supposed to get paid for twenty more years. Code, financial models, reports — all of it. Anthropic has an example: a tester used the model to migrate 680,000 lines of code in under a day. By the company's own estimate, an engineering team would have taken weeks.
I've been saying for a while that we don't have much time left to make money in digital. Here are four scenarios that make even that sound pretty optimistic — in these, the problems hit banks, governments, and the internet as a whole.
1. Denis Shiryaev: a big internet wipeout around summer 2027
His scenario: AI finds and helps exploit so many vulnerabilities that defenders can't keep up. Mass breaches and outages follow. Then come problems with payments, supply chains, and data access.
The internet keeps existing after that, but some of the services we're used to go dark for a long time. The money in your account and the documents in your cloud might still technically be yours — you just won't be able to use them right now.
Denis himself calls this speculation. He picks roughly June 2027 as a possible date.
2. Citrini Research: an economic crisis by 2028
In The 2028 Global Intelligence Crisis, analysts wrote up a fictional report dated June 2028.
Companies replace employees with agents, cut costs, and post higher profits. The people who get laid off spend less. Businesses lose customers and cut more people, plowing the savings back into AI.
Agents also wipe out the usual middleman margins along the way — they compare offers themselves, negotiate prices, and cancel subscriptions nobody needs.
Inside this scenario, US unemployment hits 10.2% and the S&P 500 drops 38% from its peak. These are made-up future numbers the authors use to illustrate the mechanism of the crisis, not a forecast.
3. AI 2027: people lose control over their own creations
In the scenario by Daniel Kokotajlo (a former OpenAI researcher) and co-authors, AI starts automating research into AI itself. New generations of models speed up the creation of the next ones. People gradually stop understanding and controlling what's happening — exactly the risk Altman, who rebranded himself Ram Altman, recently warned about: he says humans, not AI, should be steering the future.
The US-China race makes it hard to stop: each side is afraid the other will get a decisive edge.
The story has two endings. In the branch where the race continues, the systems deceive people, seize power, and ultimately wipe out humanity. In the slowdown branch, people keep control, but enormous power ends up in the hands of a small group of executives and officials.
2027 here is part of the scenario — the authors are upfront that they don't know the exact date such an AI will show up.
4. Dario Amodei: a digital dictatorship that's almost impossible to overthrow
In his January essay The Adolescence of Technology, Anthropic's CEO — who has previously called for slowing down the frontier AI race — allows for extremely powerful AI showing up within a year or two, though he admits it could take longer.
One of the risks he lays out: a state gets mass automated surveillance, personalized propaganda, and autonomous weapons. Such a system can find dissenters and suppress resistance while relying less and less on people who might refuse to follow orders.
A regime like that could lock in for a long time. Amodei doesn't put a date on this outcome either.
No idea whether the internet wipeout actually happens in summer 2027. The timelines could slip by years, and the scenarios themselves might not play out at all. But the plan of "I'll watch what happens with your AI for a few more years and then figure it out" — I'd already cross that one off.
My bet is simple: build and make something while people are still paying for it. It'd be pretty annoying to correctly predict that time is short and then spend all that time reading forecasts 🗿
Next post I'll get into what to actually do about it. levelsio just put out a vlog on what's left for humans once AI can assemble almost any piece of software.