Professor of Planetary Computing @Cambridge_CL
Cambridge, England
- Tweets 12,956
- Following 1,254
- Followers 6,281
- Likes 19,270
Pinned Tweet
Made the cover of @CACMmag with "A Decade of Docker Containers", recapping much systems work! Docker grew so fast in those early days that we never got a chance to write an academic paper about it, so this writeup has been a long time coming: cacm.acm.org/research/a-deca….
A heady mix of Go, OCaml, Swift, Linux, nested kernels, ancient network protocols like SLIRP, all mixed up into a container cauldron...
The article also comes with cool art and a video interview that was a lot of fun to film around @pembroke1347 and my overgrown office. I left some musings on my blog too anil.recoil.org/notes/cacm-d… with @mugofsoup and @justincormack
1
14
54
6,064
Good news for the Scottish Highlands in our latest paper mapping natural woodland regeneration up there with lidar, measuring the effect of deer population control on sapling survival into mature tree growth anil.recoil.org/notes/cairng…
1
5
316
Can university teams really train large-scale foundation models? Yes! In the @geotessera project, we've had a tremendous adventure training (multiple!) state-of-the-art FMs from scratch at Cambridge. This has involved vast amounts of data shuffling, stuffing every SSD we can find, overloading network links, jury-rigging GPU farms under our desks, and applying for grants on national clusters.
We also realised that we might be one of the few UK academic teams to have used Intel (Dawn), AMD (Vultr, Zenith) and NVIDIA (Isambard) GPUs at this petabyte scale. Our experiences are in this short paper, in the hope we can compare notes with other academic teams doing similar things. cl.cam.ac.uk/~avsm2/tessera-…
2
7
57
2,524
Just added support for searching for info about threatened species near you to the Dash of Life! dashoflife.org/near-me
4
419
Anil Madhavapeddy retweeted
ICLR 2027 has received more submissions than all previous years (2013–2026) combined
38
175
1,173
273,961
Anil Madhavapeddy retweeted
Are you an exceptional academic leader? Apply to join Cambridge as the founding Dean of the Rokos School of Government.
This is a once-in-a-generation opportunity to shape a major new academic institution from its inception.
We are seeking an individual of outstanding academic distinction and international standing. We welcome applications from a wide range of disciplinary backgrounds.
Explore the role and apply now: bit.ly/4cSFn47
7
19
55
12,335
Anil Madhavapeddy retweeted
Fully agree with Zhihao. As the cost of implementation and verification asymptotically approaches 0, there is little argument for broad generalization.
This is very much in the spirit of Unikernels (one of my all-time favorite projects, shoutout to @avsm). The hope here is that one doesn't need to write a declarative spec themselves as agents become more capable.
I think this comes down to a classic systems tradeoff: generality vs. specialization. vLLM/SGLang cover a huge space of models × hardware × workloads. It’s hard for a single system to be best at every combination, and focusing on a narrower set gives you more room to optimize.
Historically, the challenge was engineering cost: optimizations for one configuration often need substantial rework for another. Coding agents are lowering that cost, which I suspect is one reason we’re seeing more specialized inference engines.
Fragmentation is worth discussing, but so is the value of specialization. Many of these engines can complement vLLM/SGLang, especially when users need more performance from a specific setup. My 2 cents.
2
3
30
4,076
If AI's moving too fast for you, then consider training your agent with a hugging skill and deliver optimal load-bearing intra-hug squeezes regularly dl.acm.org/doi/10.1145/35261…
4
305
Anil Madhavapeddy retweeted
Lots been happening at @fp_launchpad over the past few months. Here’s an update of our activities.
fplaunchpad.org/blog/2026-09…
1
17
52
2,485
Anil Madhavapeddy retweeted
The contagion of fear
bcantrill.dtrace.org/2026/09…
45
229
1,144
246,366
Random UNIX history fact is that the original filesystems were just graphs and not trees, and '.' '..' are just conventions we adopted web.archive.org/web/20201002…
Thompson: No, the first one was a DAG. In fact, it wasnt even a acyclic. If you understand the UNIX file system, it was.... there was the I-list, which is a definition of all the files on the system. And then some of those files, were directories which just contained name and I-number. There's nothing in there that constrains it to a tree. So it was not in fact, not hierarchical at all. [...] Thompson: And we did not restrain it to a tree. We were experimenting with various topologies. What we ended up doing is turning into concrete and forcing the topologies that in fact were the topologies that came by convention from that system. The... Every time we made a directory, by convention we put it in another directory called directory - directory, [..]
2
12
225
22,501
A new preprint on breaking the security in brain-computer interfaces. This is not where you want to find vulnerabilities, so we categorise them as "NERVE Attacks":
- Neuro-mimetic Forgery synthesises a physiologically plausible brain signal that the classifier accepts without any data from the user.
- Evasion via desynchronization nudges the timing of the signal by a fraction of a second so discriminative features fall outside the model's receptive window.
- Replay-based Hijacking records a real epoch and replays it over the wireless link to reissue commands without the user knowing.
- Vein Tapping the unencrypted radio links, finding missing authentication, exporting world-readable model files, and (of course) memory-unsafe SDKs.
- Embedded Backdoors via trojaned models that behave normally until they see a trigger disguised as an eye blink, a muscle twitch or even mains interference.
Preprint at arxiv.org/abs/2609.08971 and my notes at anil.recoil.org/notes/nerve-… and then listen to piped.video/watch?v=B1BdQcJ2…
2
10
640
Term time's almost here, and I've refreshed my Cambridge MPhil/Part II student project ideas. A heady mix of planetary computing and systems/PL to whet anyone's appetite for challenge! anil.recoil.org/notes/studen…
1
5
490
Anil Madhavapeddy retweeted
Yeah it's been known for a while, it'll get sorted for 7.4 and headed to stable kernels post that to ensure the behavior is consistent.
2
1
4
604
io_uring will keep systems hackers employed for years to come after the AGI sublimation takes everyone else gist.github.com/avsm/03023cb…
2
3
83
5,920
The minimal C11 compiler required to build OCaml, itself written in OCaml. A fun Fable experiment from Mark: github.com/mtelvers/occ
5
10
59
3,059
Astra's accuracy in translating between two fairly obscure languages (OxCaml and Julia) for some complex climate simulations is quite spectacular compared to Sol and even Fable. It seems much more token efficient than Fable after a few days of using it as well.
21
707
The BBC radical podcasts are the perfect length for a run and distracting in a good way bbc.co.uk/sounds/brand/p0gg4…
321
The cybersecurity guards in frontier models perfectly illustrate the weird incentives in software engineering. They equally block vulnerability construction _and_ attempts to improve the correctness of software because they are both the same task.
3
1
16
1,319
spammed low-quality package contributions to OCaml are coming in with a really lovely smile now github.com/ocaml/opam-reposi…
3
10
1,452