Will AI speed up literature reviews or derail them entirely? / Jul 2025
Will AI speed up literature reviews or derail them entirely?
Sam Reynolds, Alec Christie, Lynn Dicks, Sadiq Jaffer, Anil Madhavapeddy, and Bill Sutherland.
Journal paper in Nature (vol 643 issue 8071).
papersSteps towards an ecology of the Internet / Jun 2025
Every ten years, the city of Aarhus throws a giant conference to discuss new agendas for critical action and theory in computing. Back in 2016, Hamed Haddadi, Jon Crowcroft and I posited the idea of personal data stores, a topic that is just now becoming hot due to agentic AI. Well, time flies, and I'm pleased to report that our second dicennial thought experiment on "Steps towards an Ecology for the Internet" will appear at the 2025 edition of Aarhus this August!
This time around, we projected our imaginations forward a decade to imagine an optimistic future for the Internet, when it has exceeded a trillion nodes. After deciding in the pub that this many nodes was too many for us to handle, we turned to our newfound buddies in conservation to get inspiration from nature. We asked Sam Reynolds, Alec Christie, David Coomes and Bill Sutherland first year undergraduate questions about how natural ecosystems operate across all levels of scale: from DNA through to cells through to whole populations. We spent hours discussing the strange correspondences between the seeming chaos in the low-level interactions between cells through to the extraordinary emergent discipline through which biological development typically takes place.
Then, going back to the computer scientists in our group and more widely (like Cyrus Omar who I ran into at Bellairs), it turns out that this fosters some really wild ideas for how the Internet itself could evolve into the future. We could adopti biological process models within the heart of the end-to-end principle that has driven the Internet architecture for decades!
[…623 words]
TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis / Jun 2025
TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis
Frank Feng, Sadiq Jaffer, Jovana Knezevic, Silja Sormunen, Robin Young, Madeline Lisaius, Markus Immitzer, James G. C. Ball, Clement Atzberger, David Coomes, Anil Madhavapeddy, Andrew Blake, and Srinivasan Keshav.
Working paper at arXiv.
Solving Package Management via Hypergraph Dependency Resolution / Jun 2025
New preprint survey on energy-aware deep learning on embedded hardware / May 2025
Josh Millar has just released the latest survey paper he lead on energy-aware approaches to optimise deep-learning training and inference on embedded devices, such as those benchmarked in "Benchmarking Ultra-Low-Power µNPUs" recently.
We present an overview of such approaches, outlining their methodologies, implications for energy consumption and system-level efficiency, and their limitations in terms of supported network types, hardware platforms, and application scenarios. We hope our review offers a clear synthesis of the evolving energy-aware DL landscape and serves as a foundation for future research in energy-constrained computing.
Any comments, please do let any of us know!
Energy-Aware Deep Learning on Resource-Constrained Hardware
Josh Millar, Hamed Haddadi, and Anil Madhavapeddy.
Working paper at arXiv.
Out-of-the-box LLMs are not ready for conservation decision making / May 2025
Our paper on how the careful design of LLMs is crucial for expert-level evidence retrieval has been published today in PLOS One and is available fully open access!
Our findings suggest that, with careful domain-specific design, LLMs could potentially be powerful tools for enabling expert-level use of evidence syntheses and databases. However, general LLMs used "out-of-the-box" are likely to perform poorly and misinform decision-makers. By establishing that LLMs exhibit comparable performance with human synthesis experts on providing restricted responses to queries of evidence syntheses and databases, future work can build on our approach to quantify LLM performance in providing open-ended responses.
In a nutshell, we tested 10 LLMs with six different retrieval strategies on their ability to answer questions related to conservation, benchmarked against the Conservation Evidence database that has been hand-assembled by experts over the last two decades. In some of the retrieval scenarios, models were only allowed to use their pretrained knowledge, whereas in others they had access to the relevant parts of the hand-curated database.
We found that language models had very varying results when relying only on their pretrained data, and were particularly bad at answering questions about reptile conservation. However, given some extra training with the CE database, their performance improved dramatically. When we put these models head to head with human experts (from the conservation evidence team), with a set of questions and with RAG access to the database, we found that the models were just as good as our experts, but answered the questions much much much faster (near instant).
Essentially, LLMs without extra training are likely to perform poorly and misinform decision-makers. This is crucial when considering how to build AI infrastructure for public policymaking.
[…377 words]Radhika Iyer, Alec Christie, Anil Madhavapeddy, Sam Reynolds, Bill Sutherland, and Sadiq Jaffer.
Journal paper in PLOS ONE (vol 20 issue 5).
Conservation changed but not divided / Apr 2025
Conservation changed but not divided
Sam Reynolds, Sara Beery, Neil Burgess, Mark Burgman, Stuart Butchart, Steven J. Cooke, David Coomes, Finn Danielsen, Enrico Di Minin, América Paz Durán, Amy Hinsley, Sadiq Jaffer, Julia P.G. Jones, Binbin V. Li, Anil Madhavapeddy, Lloyd Peck, Nathalie Pettorelli, Jon Paul Rodríguez, and Bill Sutherland.
Journal paper in Trends in Ecology & Evolution.
New preprint on benchmarking ultra-low power neural accelerators / Mar 2025
Josh Millar just released our latest preprint on how to make sense of the growing number of dedicated, ultra-low-power 'neural network accelerators' that are found in many modern embedded chipsets. My interest in this derives from wanting to decouple from the cloud when it comes to low-latency local environments, and this needs fast tensor operations in hardware. Josh found a huge number of interesting NPUs in modern low-cost chips, ranging from ESP32-based boards over to ARM ones. All of these have quite a variety of tradeoffs, from the operations supported (which affects which models can be run on them) to the amount of memory and CPU power.
Benchmarking Ultra-Low-Power µNPUs
Josh Millar, Yushan Huang, Sarab Sethi, Hamed Haddadi, and Anil Madhavapeddy.
Working paper at arXiv.
Updated preprint on quantifying biodiversity cost of food consumption / Feb 2025
We've uploaded a revised preprint on our ongoing work on quantifying the biodiversity cost of global food consumption, lead by Thomas Ball. This is based on the recently published LIFE metric, combined with supply chain data and provenance modeling.
[…196 words]Updated preprint on LLMs for evidence-based decision support / Jan 2025
We have just updated our preprint on using LLMs for evidence decision support with more evaluation results and corrections from peer review.
Our findings suggest that, with careful domain-specific design, LLMs could potentially be powerful tools for enabling expert-level use of evidence syntheses and databases. However, general LLMs used "out-of-the-box" are likely to perform poorly and misinform decision-makers. By establishing that LLMs exhibit comparable performance with human synthesis experts on providing restricted responses to queries of evidence syntheses and databases, future work can build on our approach to quantify LLM performance in providing open-ended responses.
See also the fantastic EEG seminar talk that the student group who worked on this over the summer gave towards the end of last year.
Radhika Iyer, Alec Christie, Anil Madhavapeddy, Sam Reynolds, Bill Sutherland, and Sadiq Jaffer.
Journal paper in PLOS ONE (vol 20 issue 5).
LIFE metric published in Royal Society Phil Trans B / Jan 2025
After some years of hard work, our Mapping LIFE on Earth biodiversity metric was published today in a special issue of the Royal Society Philosophical Transactions B! The idea behind LIFE is that although human-driven habitat loss is known to be the greatest cause of the biodiversity crisis, we do not yet have robust spatially explicit metrics that quantify the relative impacts of human actions on species extinctions. And that's what LIFE provides: a way to compare the relative impacts of some landuse anywhere in the world, in a manner that is globally applicable.
[…409 words]LIFE: A metric for mapping the impact of land-cover change on global extinctions
Alison Eyres, Thomas Ball, Michael Dales, Thomas Swinfield, Andy Arnell, Daniele Baisero, América Paz Durán, Jonathan Green, Rhys Green, Anil Madhavapeddy, and Andrew Balmford.
Journal paper in Philosophical Transactions of the Royal Society (vol 380 issue 1917).
Horizon scan on AI and conservation published / Dec 2024
Back in July 2024, a large group of conservation and computer scientists got together in the CCI to prioritise the storm of AI-related projects that have been kicking off around the world. Our key goal was to harness AI to accelerate the positive impact of conservation efforts, while minimising harm caused through either the direct or indirect use of AI technologies.
The first horizon scan resulting from this has just been published in Trends in Ecology and Evolution. If you're looking for a gentle introduction to some of the terms in AI from a non-experts perspective, the first section does a good job of defining a glossary as well.
[…118 words]The potential for AI to revolutionize conservation: a horizon scan
Sam Reynolds, Sara Beery, Neil Burgess, Mark Burgman, Stuart Butchart, Steven J. Cooke, David Coomes, Finn Danielsen, Enrico Di Minin, América Paz Durán, Francis Gassert, Amy Hinsley, Sadiq Jaffer, Julia P.G. Jones, Binbin V. Li, Oisin Mac Aodha, Anil Madhavapeddy, Stephanie O'Donnell, Bill Oxbury, Lloyd Peck, Nathalie Pettorelli, Jon Paul Rodríguez, Emily Shuckburgh, Bernardo Strassburg, Hiromi Yamashita, Zhongqi Miao, and Bill Sutherland.
Journal paper in Trends in Ecology & Evolution.
Towards verifiable, privacy-preserving carbon emissions claims / Dec 2024
Customers of online services may want to take carbon emissions into account when deciding which service to use, but it's currently difficult to do so due to the lack of reliable emissions data that is comparable across online services. There's a lot of muddled data out there, and calculating accurate carbon emissions across a computing pipeline involves a number of stakeholders, none of whom are incentivised to accurately report their emissions for competitive reasons!
In this LOCO paper, Jessica Man lead our exploration of mechanisms to support verifiable and privacy-preserving emissions reporting across a chain of energy suppliers, cloud data centres, virtual machine hosting services providers and cloud services providers. The idea is that all of this can ultimately be exposed to APIs that can be consumed by client devices in order to let consumers make direct choices about their decisions based on relative environmental impacts.
Towards a frugal userspace for Linux / Dec 2024
All the work we've been doing on biodiversity (such as LIFE) comes at a fairly large computation and storage cost due to the amount of data that we churn through. This gets worse when you consider the exploratory nature of science -- we sometimes just need to mess around with the large dataset to test hypotheses which are often shown to be wrong. So then, when the LOCO conference came around, we wrote up our thoughts on what a frugal Linux userspace might look like.
The key insight is that the Linux kernel already exposes a number of namespace mechanisms (that we use in Docker, for example), and so we explore a new OS architecture which defaults to deterministic, reusable computation with the careful recording of side-effects. This in turn allows Linux to guide complex computations towards previously acquired intermediate results, but still allowing for recomputation when required by the user. We're putting this together into a new shell known as "Shark", and this first abstract describes our early results.
Prototyping carbon-aware domain name resolution / Dec 2024
Ryan Gibb and I have been thinking about how the current Internet architecture fails to treat the carbon emissions associated with networked services as a first-class metric. So when the LOCO conference came up, we tried extending the DNS with load balancing techniques to consider the carbon cost of scheduling decisions. A next step was then to build a custom DNS server written in OCaml to actively wake machines running networked services as a side effect of the name resolution.
Extending DNS means that we maintain compatibility with existing Internet infrastructure, unlocking the ability for existing applications to be carbon-aware. This is very much a spiritual follow on to the Signposts project that I worked on back in 2013, and have always wanted to return to!
Ryan Gibb, Patrick Ferris, and Anil Madhavapeddy.
Abstract in the 1st International Workshop on Low Carbon Computing.
Displaying the 15 most recent news items out of 133 in total (see all the items)