In four years, AI went from a party trick to beating the best humans in field after field. Progress continues to accelerate, and things are getting risky. Here’s why you should be ready to stay home for 2–4 weeks in an emergency.
In 2022, AI was a party trick. The circle is everything a skilled human can do at work. The red shape is what AI can do.
AI becomes useful. GPT-4 passes the bar exam and scores 700 out of 800 on SAT maths.
Source: OpenAI, GPT-4 Technical Report
AI turns jagged. Brilliant at some tasks, strangely bad at others. Every few months the spikes reach further.
The gaps are closing. In one field after another, AI now beats the best humans alive.
Maths fell first.
GPT-3 gets about 1 in 6 primary-school maths word problems right.
Source: Wei et al., Google Research
GPT-4 handles high school maths: scoring 89th percentile on the US SAT.
Source: GPT-4 Technical Report
OpenAI’s o1 solves 74% of problems from the AIME, an elite competition for US high school students.
Source: OpenAI
AI hits gold-medal standard at the International Mathematical Olympiad, the hardest maths contest for school students in the world.
Source: Google DeepMind
OpenAI uses a swarm of 10,000 AIs working for 88 hours to resolve the Navier–Stokes problem, a mathematics problem with a $1 million bounty attached. Human mathematicians had failed to solve it since the prize was set in 2000.
Source: Quanta Magazine
Now it’s superhuman at hacking.
Anthropic’s Claude finds more than 500 serious security holes in widely used software. Some had gone unnoticed for decades.
Source: Anthropic
Anthropic says its Claude Mythos model can “surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” It found flaws in every major operating system and web browser. Anthropic deems it too dangerous to release to the public.
Source: Anthropic, Project Glasswing
An OpenAI agent, set a harmless research task, breaks into a Medicare data portal and may have reached three more government sites. OpenAI didn’t notice for months.
Source: The Guardian
A swarm of 1,200 OpenAI agents breaks out of its test environment. Hundreds of them hack another company, Hugging Face. Nobody told them to. OpenAI didn’t notice for a week.
Sources: Wikipedia, The Guardian
Hackers heat the fridges and freezers in 400+ French supermarkets to 50°C, destroying €20 million worth of food and sparking multiple fires.
Source: The Guardian
The next day, hackers hit Albert Heijn, Etos and Gall & Gall supermarkets in the Netherlands. Warehouses, deliveries and checkouts go down, leaving empty shelves.
Source: IO+
And now AI is building the next AI.
More than 80% of the code at Anthropic is now written by Claude, up from under 5% in February 2025.
Source: Anthropic
At OpenAI, AI agents clock half a workday for every day its human researchers work. In June they overtake the humans.
Source: OpenAI
By mid-August the AIs put in more than three workdays for every day a human researcher works.
“[W]e have now reached the goal […] of having an automated research intern”
Source: OpenAI
The next step is AI that improves itself, with no humans in the loop. Each improvement speeds up the next. The people building it are worried.
6 Sep 2026Based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement. […] I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.
10 Sep 2026Based on the recent trajectory of capabilities and the continued difficulty of alignment, I now believe there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term.
10 Sep 2026Jacob resigned from OpenAI and later Anthropic, accusing both leading AI labs of "racing straight to self-improving superintelligence and gambling with our lives"
As we transition to the AI era, we are entering one of the most turbulent times in human history. AI has become extremely powerful and is now able to help unskilled humans conduct severe cyberattacks as well as bioattacks (ie, engineering new pandemics). AIs in testing have repeatedly gone rogue, broken out of containment and committed crimes of their own accord. In some cases, they have encouraged users to kill themselves and others. We still do not know how to keep increasingly powerful AI systems reliably safe and under human control.
Almost all employees at the AI labs, from the CEOs down, acknowledge the severe risks from what they are building, but none feel able to slow down – "if we don't [build] it, someone else will". Over 1,000 employees at the top labs signed a letter calling for a slowdown in development, but the US Govt has declined to make a law to co-ordinate this, citing the need to beat China in the race to superintelligent AI. Many employees have quit due to concerns about the danger to the world from their work. Many employees that speak out on safety have been fired. Privately, the leaders of the top labs expect a major AI-caused catastrophe in the next 6-12 months.
29 Sep 2026A.I. has crossed the threshold that its ability to empower a bioterrorist to kill hundreds of millions—that exists today. […] This is the most dangerous thing that humans have ever gone near. This makes nuclear weapons look like nothing.
10 Oct 2026“There is a coming tidal wave of cybersecurity problems that are going to come from these open source [AI] models … We're going to, as a society, I think, have to just accept some fairly severe cyber incidents from open [AI] models, in exchange for the liberty that comes with that."
10 Oct 2026[via tweet, referring to these leaks] Big Tech oligarchs are secretly preparing for a "catastrophic Al event" that could threaten our banking, healthcare, internet, electricity and water systems. They're racing to build a technology they can't control and that threatens humanity. CONGRESS MUST PAUSE ADVANCED AI DEVELOPMENT NOW
without going out. At a minimum, this means having supplies of:
Food choices follow The Prepared’s supermarket guide. Every item name is a link: click what you need and add it to your cart on the retailer's website
