Mostly physics.
Melbourne 🇦🇺
Joined December 2014
- Tweets678
- Following604
- Followers76
- Likes341
Preprint from my work at @MATSprogram is now up: arxiv.org/abs/2610.07476
Looking at technical governance / verification Qs, specifically if power draw can be used to constrain covert compute.
Tom Kimpson retweeted
Well worth reading: proofsandprompts.com/2026/10…
Tom Kimpson retweeted
Great work from Paul and team. This is exactly right, with one further step I would add. The *short* paper - a few pages, very carefully written- needs to be done also. The "long" one, with fully open data in a standard format + all robustness - only for those who care. 1/2
This is a v. readable memo on verification thefuturesociety.org/interna…
Short piece for @LowyInstitute on why Australia is well placed for AI verification lowyinstitute.org/the-interp…
Tom Kimpson retweeted
Breaking news: Australian Prime Minister Anthony Albanese has revealed that an OpenAI agent hacked an Australian government health service website in June, in the latest high-profile cyber incident involving AI ft.trib.al/gloVAyn
Some thoughts on goal-driven research in in off-chip AI compute verification tkimpson.substack.com/p/goal… #AIGovernance #AISafety
One thing Scott raises is what are the rules for taking credit for a research paper in the age of AI?
scottaaronson.blog/?p=10062
The other extreme is the "luddite" pipeline: no AI, just pure human brainpower, like Adrian Newey designing F1 cars with a pencil and paper
This summer I was in Berkeley as part of @MATSprogram working on technical governance.
At the end of the program I gave a short symposium talk. You can check it out here: youtube.com/watch?v=G8d2uDdq…
More broadly, I thought MATS was fantastic!
As a recent convert to AI tools and AI safety (I wasn't really across any of these ideas until first playing with CC in Jan!) it really opened my eyes!
The density and level of talent was also super high
Interesting: arxiv.org/abs/2110.06581 Are all your SBI methods overconfident?
Tom Kimpson retweeted
A simple tip to massively improve your research presentations and faculty interview talks:
- Keep content simple but add periodic 'dips'.
Your ‘dives’ give experts food for thought while your simple content keeps the rest of the audience engaged.
Example:
- 1 'deep slide’ after each 4-5 simple slides.
- 2 ‘deep slides’ after each 5-6 simple slides.
This is crucial in faculty interviews, where the whole department is often listening. It also helps in postdoc interviews if you’re invited to a lab with a different expertise.
Myself, I stick to this rule most of the time. And it works very well.
__
The picture below is from our video podcast "How to Make a Great Research Talk: Structure, Slides and Confidence":
youtu.be/MfCZiJ8ekfE
I really like this paper “The Dead Salmons of AI Interpretability” by Méloux et al. arxiv.org/pdf/2512.18792 .The core argument re identifiability is v.interesting.
Tom Kimpson retweeted
📢 @ESA adopts two ambitious space science missions: Venus voyager #EnVision and gravitational wave surfer #LISA.
“These trailblazing missions will take us to the next level in two extraordinarily exciting areas of space science and keep European researchers at the forefront of these domains,” says ESA Director of Science @CGMundell
The ‘adoption’ step recognises that the missions' concept and technology are sufficiently advanced, and gives the go-ahead to build the instruments and spacecraft. Both spacecraft are planned to launch with an #Ariane6
ALT Illustration showing two black holes merging and creating ripples in the fabric of spacetime. Some galaxies are visible in the backrgound as well as a red triangle, meant to represent the position of the three LISA spacecraft and the laser beams that will travel between them.
ALT Illustration showing the EnVision spacecraft on top of two partially overlappint images of Earth (left) and Venus (right). Credit: NASA / JAXA / ISAS / DARTS / Damia Bouic / VR2Planets
Tom Kimpson retweeted
1,000 Bits of Good News You May Have Missed in 2023 - Human Progress @HumanProgress humanprogress.org/1000-bits-…
