
Former boss of @CPSThinkTank and @CapX, Sunday columnist for @thetimes, author of 'The Great Acceleration'. Politics, policy and parenting.
London
Joined December 2008
- Tweets1K
- Following2.9K
- Followers66.3K
- Likes76.3K
Robert Colvile retweeted
Today’s tech pitch:
1) Our product might kill everyone
2) If not, it will let everyone spend all day at home streaming AI-generated shortform video and cashing welfare checks
3) We need to develop it to beat China, who we also have to sell chips to
Robert Colvile retweeted
Friday's news cycle really was almost too perfect to believe.
For months, I (and so many others) have been pointing to this notable divergence between accelerating model capability growth and slowing AI spending/rising data center costs.
You really cannot get a cleaner example of the principle that "AI capabilities ≠ real-world AI consequences" than getting news of a proof of the Riemann Hypothesis—which people have for years cited as a sign of takeoff, AGI, the beginning of the productivity boom that takes us into the transhumanist Valhalla, etc etc—at the same time that the Nasdaq absolutely vomits up all over itself as the AI Trade dies.
The gap between insane model capabilities and somewhat ho-hum real-world consequences of those capabilities continues to interest me.
The AI prophets seem to have nailed the rate of model improvement but overrated the near-term impact. The unemployment rate is low, financially or infrastructurally devastating cyber attacks seem blessedly rare, etc.
Maybe we’re in a kind of Engel’s Pause but for real-world consequences, or maybe the better theory is some kind of Messy World Hypothesis where it’s really really hard for new technology to leave its mark, bc the world as it exists is a museum of inertias and frictions.
Robert Colvile retweeted
One of the funniest things Missouri ever did was when a small town had a meeting and decided to kill the local psychopath who terrorized them for years and then never snitch about it
Robert Colvile retweeted
METR shared the latest version of our "Time Horizons" work today. Takeaways, a few of which were already becoming clear imo, and are just further substantiated by this work:
1. Frontier AI agent time horizons are now doubling roughly every 3.6 months. Naively extrapolating the 50% success rate in this work, we would get months-long AI agent time horizons on well-specified tasks in the summer of 2027. Seems faster than the expectation I described below in February.
2. It's getting much harder to rule out serious AI loss of control risks using capabilities-based arguments (e.g. arguments of the type "I'm not worried that my AI agent will take unintended actions by itself, because it can't do anything by itself!"). Again, was becoming clear based on many sources of evidence, but again reiterated here.
3. The tasks we have to make are now so long and require so much human skill that its difficult to get human baselines for them (as expected).
4. Holding task-length constant, agents do worse on tasks that would have been costlier to make.
5. From @rowanwhuang: "Larger swarms tend to have bigger gaps on their performance on the test set they generate compared to what the hidden tests are actually testing for. This might suggest an “echo-chamber” effect, where swarms convince themselves to over-optimize on bad metrics, and don’t check their assumptions well."
6. When the team tried to have AI agents create their own benchmark tasks that could themselves be included in our time horizon task suite, the resulting tasks were borderline above our bar for a work-test, but not good enough to actually use.
Robert Colvile retweeted
This is massive. A story I've been following for years. He thought he had more time but didn't. Alessandro Masullo announces a separate, purely human proof of the Riemann hypothesis. math216.wordpress.com/2026/1…
Robert Colvile retweeted
The Riemann Hypothesis has fallen. RH was the singular math result that had been named, in advance, as the obvious coal mine canary for dangerously advanced AI, over the last three decades. The methods used should generalize to "make AI very good at anything you can definitely test", and unfortunately that category includes "make AI better at making AI".
This is like a private company announcing they have produced a million tons of 15% enriched uranium. It's not 90% enriched weapons-grade uranium, so you could proclaim the company's output is not immediately deadly right now. The concern is that the process line may generalize, to an extinction-level threat, within the unpredictably immediate future. Forecasters still had Riemann at 2035 one week earlier.
This is distinct from earlier events in that earlier you could reasonably ask if AI was selectively picking off famous math problems that turned out to be easier for AI. The Riemann Hypothesis was the single challenge most often named in advance as a warning sign.
Is there a good summary people have seen of why solving Riemann is such a big deal? Yours, too lazy to Google.
24
26
14,344
Robert Colvile retweeted
Astonishing impact assessment by His Majesty’s Treasury following the Brussels Deal:
“The policy increases both GDP and the population (through higher net migration) by similar proportions, so the impact on GDP per capita is 0.0% at one decimal place.”
Robert Colvile retweeted
A big moment.
"Levent Alpöge, a mathematician and employee at artificial intelligence company Anthropic, claimed on Friday to have proven the most famous problem in all of mathematics using an LLM."
Day four of both kids being sick and supplies of both Calpol and patience are running perilously low.
6
46
2,804
With due apologies to Vlad, a world-shaping superpower summit between the US, China and Russia is like a wrestling match between Hulk Hogan, the Undertaker and that spotty teenager from the Simpsons.
Robert Colvile retweeted
The government is now taking in 8.3% more from income tax than a year ago. I guess that is what frozen thresholds do. A huge tax increase that applies automatically.
Robert Colvile retweeted
If you're still concerned about citation hallucination by LLMs, you're about 2 years behind here.
Robert Colvile retweeted
Signs of greater cost discipline for CfD AR8:
1️⃣ No budget for Pot 5 (emerging technologies)
2️⃣ Notional budgets for wind and solar - meaning the government has visibility of all bids and will only procure "if bid prices are in consumers’ interests" (with HMT oversight...)
Robert Colvile retweeted
This is a banger of a post from @PhilipBarnes_
philipbarnesblog.wordpress.c…
Just the 86% of the vote for @JamesCleverly in the Tory mayoral contest...
Robert Colvile retweeted
Exactly. This amounts to stripping opportunity from those out of work.






























