.plan-26-30: Bananas ripen, models open, and OxCaml releases

OpenUK's AI Openness report, published in July 2026, features a fireside chat with Anil Madhavapeddy on open foundation models like TESSERA, emphasizing that non-inspectable, non-reproducible models are dangerous for social land-use policy. Madhavapeddy also reports prototyping the Evidence TAP website using an agent on exe.dev, cutting Eio, cohttp and mdx releases, and releasing a comprehensive OxCaml and Claude skills update.

I was in Ireland most of the week attending to family matters, so another quieter weeknote My picture of the week is that the Pembroke College Cambridge banana tree has, err, sprouted real bananas, so I'm trying to pluck up the courage to attempt to pickle them and not poison anyone. I appeared up in the OpenUK openness report https://anil.recoil.org/ the-uk-openness-report-from-openuk , got the Evidence TAP website prototyped https://anil.recoil.org/ prototyping-the-evidence-tap-website-with-an-agent using an agent on exe.dev https://anil.recoil.org/ trying-out-exedev , cut Eio, cohttp and mdx releases https://anil.recoil.org/ ocaml-and-oxcaml-releases along with a comprehensive OxCaml and Claude skills update https://anil.recoil.org/ improving-oxcaml-support-and-the-claude-skill , started taking Windows seriously now that Forester https://anil.recoil.org/ eio-in-forester has found me a real user, and on matters of nature https://anil.recoil.org/ on-matters-of-nature followed a discussion on conservation versus ecology https://anil.recoil.org/ conservation-versus-ecology and celebrated funding for optimising ranger patrols https://anil.recoil.org/ ranger-patrols-get-funding . The usual fun links https://anil.recoil.org/ fun-links are at the end. OpenUK https://openuk.uk published their excellent AI Openness report https://openuk.uk/wp-content/uploads/2026/07/AI-Openness-Report-July-2026.pdf this month. I'm in there with a fireside chat on open foundation models like TESSERA https://anil.recoil.org/projects/tessera , and a followup to conversations I started with Amanda Brock at the Indian AI summit https://anil.recoil.org/notes/india-ai-summit : Openness matters here because a foundation model of the planet that you cannot inspect, cannot reproduce, and cannot train yourself would be a dangerous thing to derive social land-use policy from. The people who need it most -- conservationists, smallholder farmers, government agencies across the Global South -- are the least likely to be able to pay big money for proprietary access. -- me in the OpenUK AI Openness: 3 Years on From Llama 2 , July 2026 'Openness' is a little odd these days in that we have to manage community more these days than adoption. For example, every TESSERA hackathon we've run https://anil.recoil.org/notes/first-tessera-hackathon has filled to capacity, and the geotessera https://github.com/ucam-eo/geotessera issue tracker is flowing with issues, so people are certainly using our output But then every single user I talked to at the CNG summit recently https://anil.recoil.org/notes/cng-london-2026 was coding at high speed with AI assistance and not really reading the docs directly, and so getting people to talk to each other mentoring, trading tips of the craft, understanding how it all works, etc takes much more effort than it used to before the advent of Claude. We're in a new world of open source... I started working on a website for Evidence TAP https://anil.recoil.org/notes/2026w29 last week, and put up a work-in-progress site at evidencetap.exe.xyz https://evidencetap.exe.xyz/ and I hope to push it live next week once the content takes shape. We've had a staggering number of requests from policymakers who are interested in our approach, and so the site gathers up the basic idea of living evidence databases that continuously ingest, screen, appraise and extract from the literature, with every claim traceable back to its source. This began with the Conservation Evidence https://anil.recoil.org/projects/ce team, and we're moving next into an education strand with Jenny Gibson https://www.educ.cam.ac.uk/people/staff/gibson/ and Mélanie Gréaux. As an experiment in collaborative agent prompting, I've been using exe.dev https://exe.dev , which gives you a pool of CPU, RAM and disk and lets you spin up as many persistent VMs as you like. The lack of ephemereality is useful when dealing with authentication and state, i.e. when doing agentic programming. Their inhouse web agent is called 'Shelley' which also has a mobile app that's just been released. My only gripe is that Shelley can't seem to interact directlly with Claude Code, and it runs on usage credits rather than being able to use a subscription account like the one I already pay for. I suspect this will get resolved as these agent platforms mature, and I'm on the Kimi K3 waitlist https://x.com/Kimi Moonshot/status/2078855608565207130 if it doesnt Outside of agentic stuff, exe.dev also has really useful protocol proxies like an inbound HTTPS tunnel that takes care of certificates, and also a very good CLI. It's a very pleasant experience to anyone familiar with Unix-style environments. A year's worth of changes went into Eio 1.4 https://github.com/ocaml-multicore/eio/releases/tag/v1.4 , which is out thanks to Thomas Leonard https://github.com/https://roscidus.com and his excellent changelogging https://notes.roscidus.com/2026/07/24/ . This website is now running on Linux 7.x and using my uring socket/bind/listen conversion https://github.com/ocaml-multicore/eio/pull/884 , and the Evidence TAP literature downloader more on that soon uses the ftruncate/fdatasync https://github.com/ocaml-multicore/eio/pull/850 and other filesystem optimisations too. I also cut a minor cohttp 6.2.2 release https://github.com/ocaml/opam-repository/pull/30339 to be compatible with the new Eio, which is winding its way through opam-repository. Eio 1.4 changed the getaddrinfo 3 into an exception and to include the query in the error message. I also cut a mdx 2.6.0 https://github.com/realworldocaml/mdx/releases/tag/2.6.0 release which I made compile with OxCaml https://anil.recoil.org/projects/oxcaml . Unfortunately you can't actually use it yet, because this new version needs an explicit dependency on unix and the fix for that in dune https://github.com/ocaml/dune/pull/15592 is merged but not yet released. We'll have to wait for dune 3.25 before its fully usable, but that's already under preparation https://github.com/ocaml/dune/issues/15593 Once that's done, getting Eio compiling with the latest OxCaml minus39 release https://github.com/oxcaml/opam-repository/pull/54 will be much easier thanks to mdx compiling out of the box. I've also published a comprehensive update https://github.com/avsm/ocaml-claude-marketplace/commit/90cd8e910f755527d2a6a63b6eae6dd748adbd4b of my ocaml-claude-skills https://github.com/avsm/ocaml-claude-marketplace marketplace which seems to be accruing users . I'm going to do a massive update of my monorepo over to minus39 in the next week or two, which should be a good test of this Claude Fable skill Feel free to send PRs if you've got improvements to my skills. I'm now also looking at Windows support rather glumly no matter how much I try, I can't get a good CRLF conversion environment . The reason I'm investigating it is that we have a REAL USER I KNOW in the form of Jon Sterling https://jonmsterling.com In Jon's recent weeknotes https://www.jonmsterling.com/2026-W28/ he mentioned progress towards Forester 6.0 which Kento Okura is pushing along with support from Tarides and the OCaml Software Foundation . Patrick Ferris https://patrick.sirref.org chimed in on our Zulip thread https://eeg.zulipchat.com/ narrow/channel/522690-Blogs/topic/JMS.20Weeknotes.202026-W28/with/611607207 with some tips about cutting out a stat per file. Patrick knows Forester https://tangled.org/@jonmsterling.com/ocaml-forester well, having become familiar with its directory scanning logic while writing his own graft https://github.com/patricoferris/graft fork/variant which his website runs on. I also integrated a few quality-of life utilities like the null sink https://github.com/ocaml-multicore/eio/pull/894 into Eio 1.4 which a future Forester can hopefully use. Temporary files and directories are next on my list, as I've discovered O TMPFILE on Linux that allows for the creation of anonymous fds. I asked Jon whether he ships Windows binaries and he said it has long been a goal, blocked mostly on how Forester handles paths. After I achieved the milestone of working out how to map the \ key onto my Windows machine, I started understanding modern Windows paths. Kento's pathname PR https://github.com/ocaml-multicore/eio/pull/738 is a start, but we'll likely need to have Windows and POSIX specific pathname logic that's separate. Aneesh Naik https://aneeshnaik.github.io/ started an interesting thread https://eeg.zulipchat.com/ narrow/channel/522690-Blogs/topic/APN.20weeknotes.202026.2F07.2F20/with/612517496 on our Zulip off the back of his weeknote https://www.aneeshnaik.com/blogposts/20260720 weeknotes 2026 29.html about whether nature should be monitored at all. His thought came from a colloquium in the CCI by Hugh Possingham https://en.wikipedia.org/wiki/Hugh Possingham , who argued that one should only pursue a nature monitoring project when there is a specific on-the-ground action that will follow from it. However, a great deal of what we now know about effective conservation comes from people who kept their monitoring schemes running long past the point where they had any planned interventions left The paper behind the talk https://doi.org/10.1098/rspb.2025.2527 lays out many reasons for monitoring as a checklist and a theory of change, and does include plain societal curiosity Much of the remote sensing https://anil.recoil.org/projects/rsn work going on here is really ecological monitoring rather than conservation monitoring: the former is aimed at understanding an ecosystem itself and the latter at establishing the counterfactual impact of some intervention. Ecology does have a close sister discipline of conservation with some extremely vital imperatives attached, and a temporal urgency https://anil.recoil.org/notes/nas-rs-biodiversity with respect to human actions. Huge congratulations to Charles Emogor https://charlesemogor.com , who has landed one of the highly contended AI4PG 2026 fast grants https://www.ai4pg.com/grants/ai4pg-2026 for our work on optimising ranger patrol deployment in sub-Saharan Africa. The project is with Rob Fletcher https://www.zoo.cam.ac.uk/directory/professor-rob-fletcher , Lingkai Kong at Harvard and Inaoyom Imong at the Wildlife Conservation Society in Nigeria, and eleven proposals were selected from 135 submissions across 38 countries. This funds the next stage of the work Charles has been driving on validating anti-poaching predictions with ranger insights https://anil.recoil.org/ideas/validating-anti-poaching-predictions and mapping hunting risks https://anil.recoil.org/ideas/mapping-hunting-risks-for-wild-meat . The core problem is that patrol-collected data is biased since rangers only see what they walk past, and so a map of "where poaching happens" is often really just a map of "where rangers went". Charles plans to combine ranger patrols from 2016 onwards with GPS data collected from 120 hunters across eight communities, model the two independently, and use the mismatch to build a bias-correction layer over three protected areas in southeast Nigeria. Given that billions have been spent on anti-poaching globally https://thedocs.worldbank.org/en/doc/fa250077ab87b35d0451b82653e0f4f3-0320072025/original/GWP-2025-Analysis-of-International-Donor-Funding-to-Combat-IWT-2010-2023.pdf since 2010, even modest improvements in where the patrols go are worth a great deal