AI safety research
Oct – Nov 2026
Catching AI sabotage on a budget
BlueDot Impact · Technical AI Safety Project
More and more code is written by AI, and a model could hide a
backdoor in it. A common defence is a monitor, a second AI that
reads each program and flags the suspicious ones for a human to
check. Smaller labs and businesses have no good guide to which
monitor is worth paying for.
With Anusha Asim and Marnick Vandecauter, I tested 13 monitors
on 500 coding problems with backdoors hidden in them, and priced
every run. When a human could check 2% of the code, Sonnet 5.5
caught 88% of the backdoors and Haiku 5.5 caught 85% for a fifth
of the price. If a human can only check 1% or less, Sonnet's lead
grows to between 12 and 27 points.
- More reasoning raised catch rates for 6 of 8 models, by up to 28 points
- Sent a fix to ControlArena, the UK AI Security Institute's control framework
Writing
Since Sep 2026
Explaining AI to people who don't work in it
The Pivot's Archive · Substack
Most people will live with AI's consequences without ever
reading a research paper. I write for them: no hype, no doom,
just enough to form their own view.
Every essay also gets a hand-built interactive version you scroll through.
Business
Taking clients
Pivotr Studio
My web studio for small businesses
Building software keeps getting cheaper, so a local business
shouldn't need an agency budget to get a good website.
Clean, fast websites with no ongoing fees, and one free build
every month for a non-profit.
Software
Live
Tools I couldn't find, so I built them
The alternatives were paid, or didn't exist
All of them are free and private by default. Where it can, your
data stays in your browser, and there's no account to make.
Open the tools hub
Interlude
A deliberate pause
I stepped away from full-time work and spent five months
travelling to recalibrate: working out what the world needs,
what I'm good at, and what someone will pay me for. Everything
above came out of it.