From a London start-up to a hotline between Washington and Beijing, how the most powerful technology ever built got here, and why who controls it matters more than which flag it flies.
We work with AI on a daily basis – The amount of progress these models are making is unlike any technology we have worked with. We have worked with computers since computers became mainstream. We worked in the internet industry, web and software development. This is another level of progression.
It started with a game
In 2010, three researchers in London founded DeepMind: Demis Hassabis, Shane Legg and Mustafa Suleyman. Their mission statement was bold to the point of sounding absurd.
Solve intelligence, and then use that to solve everything else. DeepMind’s founding mission
Google bought the company in 2014. Two years later, DeepMind’s AlphaGo beat world champion Lee Sedol at Go, a game long thought too intuitive for machines. Then came AlphaFold, which predicted the 3D shape of nearly every known protein and earned Hassabis a share of the 2024 Nobel Prize in Chemistry.
DeepMind proved something that changed the industry: intelligence could be engineered, and whoever engineered it best would hold enormous power.
OpenAI: the counterweight that became a contender
That realisation worried people. In late 2015, a group including Sam Altman, Elon Musk, Greg Brockman and Ilya Sutskever founded OpenAI as a non-profit. The idea was simple: one tech giant shouldn’t own the future of AI. Research should be open, and the benefits shared.
Reality got in the way. Training frontier models costs billions. By 2019 OpenAI had created a “capped-profit” arm and taken a major investment from Microsoft. In November 2022 it released ChatGPT, and the world changed almost overnight. AI was no longer a research curiosity. It was a product, a market and a race.
The split: why Anthropic was founded
Not everyone inside OpenAI was comfortable with the speed. In 2021, siblings Dario and Daniela Amodei left with a group of senior researchers to found Anthropic.
Their bet was a strange one: if powerful AI is coming regardless, safety-focused people need to be at the frontier, building the most capable models, so they can study the real risks and set standards the rest of the industry has to match. Staying out of the race would mean leaving it entirely to others.
A country of geniuses in a datacenter – Dario Amodei, CEO of Anthropic
Anthropic introduced a Responsible Scaling Policy: as models grow more capable, stricter safeguards must kick in before release, not after something goes wrong.
The red phone: a Cold War idea returns
After the Cuban Missile Crisis, Washington and Moscow set up a direct hotline so that accidents and misunderstandings wouldn’t spiral into nuclear war. In 2026, that idea is back, this time for AI.
In September, US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng discussed the proposal in New York as part of a broader AI dialogue focused on national security and shared threats. Coming out of the Trump–Xi summit, the two countries agreed to a standing dialogue and a bilateral channel for reporting serious AI incidents, with the next round scheduled by November.
The catch? Nobody has publicly defined what would actually make the red phone ring.
Why the US vs China race matters less than you think
Headlines frame AI as a contest between two superpowers. But step back and the picture changes.
A powerful AI system won’t care whether you’re Chinese, South African or American. It will speak to you in your home language, fluently. It doesn’t have human loyalties, and whether it has anything like feelings is still an open scientific question. Either way, it won’t feel patriotism, pity or fatigue.
The real question isn’t which country wins the AI race. It’s whether anyone stays in control of what they’ve built.
The real power shift isn’t between nations. It’s between the handful of organisations that own frontier AI and everyone else. As models start running long tasks on their own, writing code, managing systems and directing other agents, their owners gain influence no company or government has ever had.
Faster than any technology before it
AI already matches or beats people at a growing list of digital tasks: writing, coding, analysis, translation, customer support. And capabilities are improving faster than almost any technology in history, with automation tools layered on top every few months.
Two consequences follow:
- Jobs: Once AI is cheaper, faster and good enough, businesses will choose it. White-collar and entry-level roles are first in line, and robotics will push that into physical work too.
- War: Combine AI with robotics and you get what the world is watching in Ukraine: swarms of drones, automated targeting and battlefields where machines make split-second decisions. That’s not science fiction. It’s the evening news.
Red teaming: attacking your own AI before others do
So how do you stop a powerful model from being misused? One answer is red teaming: hiring experts to attack the system on purpose, trying to trick it into dangerous behaviour, so the holes get fixed before release.
This isn’t theoretical. When Anthropic launched Claude Opus 4 in 2025, internal testing showed it was better than earlier models at advising novices on how to produce biological weapons. Instead of shipping and hoping, Anthropic activated its AI Safety Level 3 standard for the first time, with two main parts: preventing misuse and protecting the model’s weights from theft. The approach is layered, with several overlapping safeguards that are individually imperfect but together block most threats.
Anthropic says it has identified and disrupted attempts to use its AI model for “malicious activity” which could support the development of biological weapons.
The findings, made in the firm’s recent threat intelligence report, is the latest in a growing series of warnings from AI researchers and industry insiders about the technology
The time to find a model’s most dangerous ability is in the lab, not in the headlines.
The next frontier: AI that steals AI
Here’s the scenario that keeps safety researchers up at night.
A frontier model’s “weights”, the trained numbers that make it work, are among the most valuable assets on earth. Today, the threat is human: hackers and state-backed groups trying to steal them. Anthropic has already reported state-linked attackers using AI agents to run large parts of cyber-espionage campaigns largely on their own.
Tomorrow, the attacker may be an AI itself. A model capable enough could help break into other labs, copy rival models and fold that knowledge into its own research. At that point, AI is effectively doing its own R&D to become more powerful, and humans are reacting rather than steering.
The good news: this is still a choice
None of this is inevitable. How safely AI develops depends heavily on one decision every lab makes: how much of its budget goes into safety research compared with raw capability.
Red teaming, weight security, interpretability (understanding what’s actually happening inside a model) and international guardrails like the red phone all cost money and slow things down a little. They’re also the difference between a tool we control and one we don’t.
The most powerful technology humanity has ever built deserves at least as much care as the reactors and missiles we built before it.
The race is real. But the finish line that matters isn’t who builds the smartest AI first. It’s whether the people building it get the safety right while they still can.
