Anil Madhavapeddy, Professor of Planetary Computing

Do not rule out nature from climate action; an open letter / Sep 2025

A world without nature feels rather impermanent, doesn't it? It's difficult to imagine a healthy future without clean air, fresh water and diverse wildlife. Yet important policy is being decided at the moment that will sideline "nature-based solutions" for net-zero carbon targets. While it is true that anything involving nature is fundamentally less predictable than human edifice, it is not true that it can't be quantified through science-based methods! Advances in remote sensing mean we have better resolution views into nature than ever before in human history, and we can leverage those towards protecting what's left. The wrong economic incentives are pushing us into a dangerous crossroads where several policy paths effectively abandon nature.

Back in January 2024, I hosted a workshop on permanence and durability at Pembroke College attended by sixty experts in this topic, with their recommendation reflected in the continuous improvement report on permanence from the ICVCM in May this year. I now join 40 other colleagues today in signing an open letter to the UN Article 6.4 supervisory body strongly calling out the scientific imperative to incentivise natural climate solutions on the path to net zero.

Exploring the biodiversity impacts of what we choose to eat / Sep 2025

Choosing where we source the food that we eat makes a difference to the environment, but by how much? After churning through around 100 petabytes of data, beginning with our LIFE metric and moving onto food provenance maps and import/export data for the world, we now know the answer can vary by three orders of magnitude for species extinction risks.

Our paper in Nature Food came out today with all the tasty details and implications for food policies worldwide. In order to make the data easier to explore, I knocked up an interactive global explorer using the data that the team (lead by Thomas Ball) generated.

GeoTessera Python library released for geospatial embeddings / Aug 2025

We've been having great fun at the EEG recently releasing embeddings of our new TESSERA geospatial foundation model.

TESSERA is a foundation model for Earth observation that processes Sentinel-1 and Sentinel-2 satellite data to generate representation (embedding) maps. It compresses a full year of Sentinel-1 and Sentinel-2 data and learns useful temporal-spectral features. -- Temporal Embeddings of Surface Spectra for Earth Representation and Analysis

A foundation model is designed to be used for downstream tasks without having to retrain a full model for every individual task. Our preprint paper describes what sorts of geospatial tasks you can solve more quickly, ranging from crop type classification, forest canopy height estimation, above-ground biomass calculations, wildfire detection, forest stocks, and many more.

Presenting our Ecology of the Internet ideas at Aarhus 2025 / Aug 2025

That's a wrap for the next decade with Aarhus 2025, where I presented our paper on "Steps towards an Ecology for the Internet". I was a little unsure about how to approach the presentation, largely because the ideas seem a little crazy if they'd been proposed even a year ago! Luckily my co-authors strengthened my spine with encouragement and gin, and the event was tremendous fun packed with useful insights.

Our key observation is that the Internet is dangerously ossifying into monocultures at multiple levels. Inspired by wild ecosystems, we're proposing mixing in more natural selection into edge deployments by using AI code models to mutate end-hosts and tailor them to their environment. Generative AI is notoriously unpredictable, which turns out to be a useful property if you actually want more local software diversity! For example, this lets us cook up "antibotty" networks that fight back against global viruses via locally adapted vigilantes (antibodies).

mlgpx is the first Tangled-hosted package available on opam / Aug 2025

Since I wrote about the new ATProto-powered Tangled Git forge a few months ago, it's come along by leaps and bounds!

First, and most excitingly, they've added continuous integration via Spindles which are built in a nice ATProto style:

When you push code or open a pull request, the knot hosting your repository emits a pipeline event (sh.tangled.pipeline). Running as a dedicated service, spindle subscribes to these events via websocket connections to your knot.

The pipelines are Nix-only right now, so I braved using it^[1] for a new GPS Exchange Format library in OCaml that I wrote. The pipelines should look familiar, and the description format very straightforward.

Secondly, the service has added support for JJ stacked pull requests, which are the closest I've seen to the Jane Street Iron diff workflow which I've been wanting to try in open source for ages. You can see the interdiff review process on a recent PR by Winter who add support for engine-agnostic Spindle workflows, which should pave the path for a Docker or BuildKit engine alongside the existing Nixery-based one.

And thirdly, the general quality-of-life of the web frontend has improved dramatically, with a nice timeline, repo list, and profile pages. I'm running two knots right now (one on Recoil, and one in the Cambridge Computer Lab), and both have been very painfree. I wrote one of the earliest Dockerfiles for it, but there's now a community-maintained Knot Docker setup which I've switched to. Doesn't take very long at all; give it a try!

Because I've been using Tangled so much, I added support for Tangled metadata to Dune to make OCaml package maintainence easier. This will appear in Dune 3.21 in a few months, but in the meanwhile enjoy the first Tangled.sh package on opam. It's a simple GPX library I used in my recent trip to Botswana. All you need in your dune-project will be:

(lang dune 3.21)
(name mlgpx)
(generate_opam_files true)
(source (tangled @anil.recoil.org/ocaml-gpx))

The only major thing I'm missing from Tangled is support for private repositories now, but I'm very content using it for public content today. Beware as usual that it's still in alpha, so don't trust super-ultra-mega-important stuff to it unless you've git mirrored elsewhere.

Tracking locations with OwnTracks, Life Cycle and Home Assistant / Aug 2025

I'm emerging reenergised from an epic trip to the Okavango Delta in Botswana, where we spent weeks in the wilderness gathering ground truth for TESSERA (and enjoying the wildlife!). Piecing together our locations was quite important, and so I took a cue from Ryan Gibb and deployed OwnTracks and HomeAssistant Device Tracker before I headed out there. There were four interesting tech pieces that resulted from this: a local iPhone app to determine my GPS accuracy while entirely remote; then merging my Home Assistant location database hile in the field, then reverse engineering 8 years worth of location data out of an old iOS app to backfill data, and finally deploying my own self-hosted OwnTracks on Recoil for the longer term.

Cresting the OCaml AI humps / Jul 2025

I've been hacking with Sadiq Jaffer (^), Jon Ludlam (^) and Ryan Gibb (^) on various approaches to improving the agentic coding experience for OCaml.

We jotted down our notes in a draft paper to keep track of everything going on, including summarising previous experiments with Qwen3 for FoCS. Since then, there's been a flurry of extra activity from others which we need to integrate!

• Marcello Seri started pushing to my vibe coded OCaml MCP library, making him user number 2 of that!
• Then Thibaut Mattio announced a bunch of software, starting with a collection of libraries and tools for numerical computing and machine learning and also another MCP server. I haven't had a chance to try the MCP server yet, but I hope I can retire mine...
• Thomas Gazagnaire started hacking on an agent-friendly merlint tool that spots common problems in style and choices and gives CLI feedback in a style easily consumed by claude. I've started using it despite its pre-alpha status.
• Jon Ludlam's been getting the opam embeddings into shape to be suitable as an MCP server that can search the entire opam ecosystem. odoc v3 has also gone live after lots of work, and David Sancho's support for Markdown odoc output on which makes this process easier.

This is all fairly straightforward MCP work that improves the short-term experience. We'll get to the RL-VR ideas later... If anyone else is hacking on something agent related do post on OCaml Discuss and let us know! I'm hoping to update the paper later in August to roundup the efforts above.

Using Kyutai's low latency audio models on macOS in one command / Jul 2025

I've just taken Kyutai's speech-to-text model for a spin on my Mac laptop, and it's stunningly good. As background, this is what the prolific Laurent Mazare has been hacking on; he has made a ton of contributions to the OCaml community as well, such as ocaml-torch and starred in a very fun Signals to Threads episode on machine learning at Jane Street back in 2020.

You can get the microphone-to-speech running on your Mac in a few commands, assuming you have uv installed (which you should!).

git clone https://github.com/kyutai-labs/delayed-streams-modeling
cd delayed-streams-modeling
uvx --with moshi-mlx python scripts/stt_from_mic_mlx.py

It understands my accent near perfectly; if that isn't a machine learning miracle, I don't know what is! I'm looking forward to trying this out more with our Low power audio transcription with Whisper project over the summer with Josh Millar and Dan Kvit.

Royal Society's Future of Scientific Publishing meeting / Jul 2025

I was a bit sleepy getting into the Royal Society Future of Scientific Publishing conference early this morning, but was quickly woken up by the dramatic passion on show as publishers, librarians, academics and funders all got together for a "frank exchange of views" at a meeting that didn't pull any punches!

These are my hot-off-the-press livenotes and only lightly edited; a more cleaned up version will be available from the RS in due course.

Is AI poisoning the scientific literature? Our comment in Nature / Jul 2025

For the past few years, Sadiq Jaffer and I been working with our colleagues in Conservation Evidence to do analysis at scale on the academic literature. Getting local access to millions of fulltext papers has not been without drama, but made possible thanks to huge amounts of help from our University Library who helped us navigate our relationships with scientific publishers. We have just published a comment in Nature about the next phase of our research, where are looking into the impact of AI advances on evidence synthesis.

EEG internships for the summer of 2025 / Jun 2025

The exam marking is over, and a glorious Cambridge summer awaits! This year, we have a sizeable cohort of undergraduate and graduate interns joining us from next week.

This note serves as a point of coordination to keep track of what's going on, and I'll update it as we get ourselves organised. If you're an intern, then I highly recommend you take the time to carefully read through all of this, starting with who we are, some ground rules, where we will work, how we chat, how to get paid, and of course social activities to make sure we have some fun!

BIOMASS launches to measure forest carbon flux from space / Jun 2025

The BIOMASS forest mission satellite was successfully boosted into space a couple of days ago, after decades of development from just down the road in Stevenage. I'm excited by this because it's the first global-scale P-band SAR instrument that can penetrate forest canopys to look underneath. This, when combined with hyperspectral mapping will give us a lot more insight into global tree health.

Weirdly, the whole thing almost never happened because permission to use the P-band was blocked because it might interfere with US nuclear missile warning radars back in 2013.

Meeting in Graz, Austria, to select the the 7th Earth Explorer mission to be flown by the 20-nation European Space Agency (ESA), backers of the Biomass mission were pelted with questions about how badly the U.S. network of missile warning and space-tracking radars in North America, Greenland and Europe would undermine Biomass’ global carbon-monitoring objectives.

Europe's Earth observation satellite system may be the world's most dynamic, but as it pushes its operating envelope into new areas, it is learning a lesson long ago taught to satellite telecommunications operators: Radio frequency is scarce, and once users have a piece of it they hold fast. -- Spacenews (2013)

Luckily, all this got sorted by international frequency negotiators, and after being built by Airbus in Stevenage (and Germany and France, as it's a complex instrument!) it took off without a hitch. Looking forward to getting my hands on the first results later in the year over at the Centre for Earth Observation.

Check out this cool ESA video about the instrument to learn more, and congratulations to the team at ESA. Looking forward to the next BIOSPACE where there will no doubt be initial buzz about this.

Update 28th June 2025: See also this beautiful BBC article about the satellite, via David Coomes.

Under the hood with Apple's new Containerization framework / Jun 2025

Apple made a notable announcement in WWDC 2025 that they've got a new containerisation framework in the new Tahoe beta. This took me right back to the early Docker for Mac days in 2016 when we announced the first mainstream use of the hypervisor framework, so I couldn't resist taking a quick peek under the hood.

There were two separate things announced: a Containerization framework and also a container CLI tool that aims to be an OCI compliant tool to manipulate and execute container images. The former is a general-purpose framework that could be used by Docker, but it wasn't clear to me where the new CLI tool fits in among the existing layers of runc, containerd and of course Docker itself. The only way to find out is to take the new release for a spin, since Apple open-sourced everything (well done!).

Visiting National Geographic HQ and the Urban Exploration Project / Jun 2025

I stayed on for a few days extra in Washington DC after the biodiversity extravaganza to attend a workshop at legendary National Geographic Basecamp. While I've been to several NatGeo Explorers meetups in California, I've never had the chance to visit their HQ. The purpose of this was to attend a workshop organised by Christian Rutz from St Andrews about the "Urban Exploration Project":

[The UEP is a...] global-scale, community-driven initiative will collaboratively track animals across gradients of urbanization worldwide, to produce a holistic understanding of animal behaviour in human-modified landscapes that can, in turn, be used to develop evidence-based approaches to achieving sustainable human-wildlife coexistence. -- Christian Rutz's homepage

This immediately grabbed my interest, since it's a very different angle of biodiversity measurements to my usual. I've so far been mainly involved in efforts that use remote sensing or expert range maps, but the UEP program is more concerned with the dynamic movements of species. Wildlife movements are extremely relevant to conservation efforts since there is a large tension between human/wildlife coexistence in areas where both communities are under spatial pressure. Tom Ratsakatika for example did his AI4ER project on the tensions in the Romanian Carpathian mountains, and elephant/human conflicts and tiger/human conflicts are also well known.

What I learnt at the National Academy of Sciences US-UK Forum on Biodiversity / Jun 2025

I spent a couple of days at the National Academy of Sciences in the USA at the invitation of the Royal Society, who held a forum on "Measuring Biodiversity for Addressing the Global Crisis". It was a packed program for those working in evidence-driven conservation:

Assessing biodiversity is fundamental to understanding the distribution of biodiversity, the changes that are occurring and, crucially, the effectiveness of actions to address the ongoing biodiversity crisis. Such assessments face multiple challenges, not least the great complexity of natural systems, but also a lack of standardized approaches to measurement, a plethora of measurement technologies with their own strengths and weaknesses, and different data needs depending on the purpose for which the information is being gathered.

Other sectors have faced similar challenges, and the forum will look to learn from these precedents with a view to building momentum toward standardized methods for using environmental monitoring technologies, including new technologies, for particular purposes. -- NAS/Royal Society US-UK Scientific Forum on Measuring Biodiversity

I was honoured to talk about our work on using AI to "connect the dots" between disparate data like the academic literature and remote observations at scale. But before that, here's some of the bigger picture stuff I learnt...