Davide Eynard (+mala)
-
-
CW: AI, LLMs
https://blog.mozilla.ai/llamafile-returns/
#llamafile has been instrumental in my first projects relying on language models (I purposely left “large” aside, as the one I used for #BYOTA is tiny).
While looking for support for more recent models I tried other applications, but I never felt “at home” with them as I did with llamafile. This is why I am so happy of this move: breathing new life into it is, for me, giving ppl a new chance to better understand (what ppl now call) “#AI” while tinkering with it.
-
It’s been three weeks already since #EndSummerCamp and so much happened that I haven’t had the time to sit and write something about it yet. As a super-short summary, it was a great experience as usual and if you happen to be in north-east Italy at the… well, end of summer 😬 I strongly suggest you to check it out.
This year’s talks are here: https://youtube.com/playlist?list=PL-8hpaXWo9guJT1fkMotyu0FPdY4qKVc9&si=LUFQSbd_WVcnLGcO
The slides of my talk about #opensource #AI #agents are here: http://3564020356.org/zelif/agents_to_the_people.pdf
-
http://3564020356.org/ch35t/wasm/ is the endpoint. The riddle you’ll find is the one that gives you access to my old website in case you are curious to check it out 🙂
If you change the URL to another valid JSON (see the ch35t/examples dir) a new riddle will load. You can already create your own riddle in #ch35t format, provide its URL here, and play it. There are still good chances that something will break 😇 so please lmk if it does and I’ll fix it. Or you can, the code is all there! 😜
-
#ch35t is also a python library that parses its own format. Until yesterday, it came with a CLI client to read a json, validate it, show all information from a chest (including decoding different hint formats), and unlock it with a key.
For my son’s riddle, I just built a @marimo_io client that provides a nice browser UI. I made sure all python deps are supported by pyodide, so the client becomes a portable static HTML file + JavaScript. This way, my son just needs to connect to a URL to play.
-
What does a #ch35t look like? Check out http://3564020356.org/ch35t/examples/ for a few examples.
What riddles can you create with it? The simplest ones hold a plain text question and the md5-encrypted answer you need to find. But you can also use eg. a base64-encoded zip file for your hint, and AES to encrypt another file using the riddle’s answer as a key.
This is already enough to encode almost all 356* riddles as ch35ts, so that anyone will be able to play with them even when the site is long forgotten.
-
#Ch35t is called this way because “it’s a chest with a little bit of 3564020356.org in it”🙂
It is first and foremost a format, defined with a json schema, that one can use to formalize a riddle. There’s a “label” with name and author, a hint telling you how to find the chest’s key (the riddle answer), and then the chest’s payload, that you only get if you have the right key.
Both hint and payload are data which can take any format. The payload is locked with the key via encryption.
-
After an almost 2-years hiatus, a new version of #ch35t (0.1.9) is out. If you were not expecting this, just know I am more surprised than you!
I just couldn’t hear my son say he’s bored anymore, so I built a riddle for him. Well, the intention was a scavenger hunt, but I had to downsize this to something I could finish in one day (but still hard enough so that I could create one more before he completed it).
Preparing the riddle was easy, but it became an excuse to pick up ch35t again… 🧵
-
New blog post: “Building a Custom Feed on ATProto Part 2: Text-to-Feed with BYOTA and Embedding Similarity”
https://aittalam.github.io/posts/2025-07-20-text-to-feed-with-byota/
-
Uuuh the video of the #BYOTA speed demo at #fediforum2025 is online! :partyparrot:
https://spectra.video/w/xtZeTuwpYAqPszVBK1uULd
If you missed the event, check this page out for videos of keynotes and demos and session notes: https://fediforum.org/2025-06/ 🔥
-
Today I woke up then I used #BYOTA to build an algorithmic timeline on #Bluesky... A good 8 hours before #FediForum starts 😅 so there's something more I can bring there!
Nice facts:
- everything runs on a #RaspberryPi 5
- the diff with the https://github.com/MarshalX/bluesky-feed-generator example is ~10 lines of code
- I dedicated my experiment to PrideMonth and I got a fabulous timeline in return :BlobhajPrideHeart: -
@powersource @tommi Yes AND I think the main problem is understanding where these algos will run. With ATProto it is a seamless experience but it requires you to have your own server, domain, and some technical knowledge.
I am working on #BYOTA (https://github.com/mozilla-ai/byota) to bring algorithms in the hand of their users (they run 100% locally), but it is still a PoC and I would like to find a better way to give ppl as much control as possible without leaving their current clients. Happy to chat :-)
-
@mario same! Well, I already started playing with it but mostly because I was interested in understanding LoRa’s capabilities… I have been playing with offline-first, self-hosted, low power servers with #PicoGopher some time ago and I always wondered whether it’d be possible to sync small websites or gopherholes with it.
-
https://web.archive.org/web/20191201105759/http://search.lores.eu/
+Fravia passed away on a Sunday, the 3rd of May, 16yrs ago.
Even after all this time, I keep meeting people who tell me how influential he's been for them. People I immediately feel connected to, as if we were meeting through his website once again.
I wonder what he would say today, after witnessing the enshittified Web. What I am sure about, he'd have a lot of fun learning (and teaching) how to break out of all this.
-
Let’s say you have an HTTP server running locally that returns a custom Fediverse timeline… How would you make it available to fediverse clients?
First thing I have tried to build was a Firefox extension that redirects `/timelines` requests for a masto instance to the local server, but mastodon does not look happy about the different IP.
On the other side, taking an existing client (e.g. @elk) and adding an extra timeline looks way easier, but rather ad-hoc.
-
New blog post: Building a custom feed on ATProto
https://aittalam.github.io/posts/2025-04-13-building-a-custom-feed-on-atproto/
(no, I am not moving to Bluesky, I just want to understand how generalizable #BYOTA is and break as many walls as I can in the process 🙂)
-
One thing I missed when writing the Blueprint post about #BYOTA was a more personal take about it. Well, here it is: https://aittalam.github.io/posts/2025-04-05-byota/
(If you are wondering about the Blueprint post: https://fosstodon.org/@mala/114202120125293155)
-
I already talked about why we should own our timeline algorithms (https://fosstodon.org/@mala/113967540546211928 ) and promised a blog post soon… Here’s the long awaited post AND a whole new Mozilla.ai Blueprint 🔥
https://blog.mozilla.ai/build-your-own-timeline-algorithm-a-blueprint-2/
I chose the Blueprint format as a commitment to share my work better, with updated code, more documentation, Docker images, and a demo you can play with right now: https://mozilla-ai-byota.hf.space/
-
This was #BYOTA, thanks for following this long thread! What you see here are some of the next steps I planned: in addition to natural ML extensions, I’d like to see it grow as a tool for people to experiment with different algorithms and easily share them, and for less tech-savvy people to use as easily as possible. For this to be true, I will invest some time in understanding how to bring this to fruition at protocol level, rather than a single application. Stay tuned!
-
So, what can you do with #BYOTA? The first thing is embeddings visualization. In these pictures you can see a 2D plot of embeddings calculated on four different timelines: home (blue, only people I follow), local (orange, all posts from my instance, which is fosstodon.org), public (red, federated posts from people followed by users of my instance), and the timeline that I got searching for the #gopher hashtag (light blue).
-
What does having a WASM-powered notebook mean? Consider #BYOTA: you can download it from its repo, pip-install its deps, and run it locally as any python notebook. But you can also deploy it as HTML+Javascript files, host it somewhere super cheap (coz the server won’t run any of your code), and people will be able to run it in their browser with no need to install anything else! Plus, this will work both with “my” algorithm and whatever alternative you might develop starting from BYOTA’s code.
-
In #BYOTA, we use #llamafile to calculate *sentence embeddings*.
If you don’t know what embeddings are, just think about them as numerical descriptors of your Mastodon statuses, which are closer (as in two cities’ coordinates being close on a map) the more semantically similar their respective statuses are. We’ll get back later to this with a more visual description. If you are interested in embeddings and wanna delve deeper, see https://vickiboykis.com/what_are_embeddings/ by @vicki. -
#Llamafile is a Mozilla tool that packages a language model in a single executable file that will run on most platforms. It is 100% local and has been optimized to run on slower hardware, from Raspberry Pis to my 8yo laptop. It is based on llama.cpp which supports a plethora of models, not just LLMs: I chose all-minilm because it’s tiny (50MB) and has open code, research papers, and datasets. https://github.com/Mozilla-Ocho/llamafile https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
-
I got back home from the amazing #SocialWebFOSDEM and while I am drafting a “Build Your Own Timeline Algorithm” blog post to follow up on my talk I decided to keep the momentum going with a thread about it. You'll find my code at https://github.com/aittalam/byota and the talk slides here: https://fosdem.org/2025/schedule/event/fosdem-2025-5601-build-your-own-timeline-algorithm/ (a video with the recording of the talk should be available soon too). Now let’s dig into #BYOTA!
-
https://genesistemple.com/a-swashbuckling-tale-of-italian-software-piracy-1983-1993
This is the “gaming scene” I grew up with - I still remember going to the newspaper kiosk as an 8yo to buy new games with no clue they were pirated 😅
When I created my first videogame with C64 SEUCK (called “The Night Of The Living Dead”) I submitted it to one of those magazines hoping to be published. That never happened AFAIK, but I sometimes wonder whether a copy of that game is anywhere out there…
-
Experimenting with #marimo, mastodon API, and #llamafile for my "Build Your Own Timeline Algorithm" project. 100% local, except those 11 seconds it took me to download my home timeline.
I have been pitching this idea around for almost two years now, and while I felt bad for not pushing more for it, I am glad that in the meantime new cool tech became available allowing me to do something even better than I imagined.
-
Back on a plane for my flight back home from 3 great days at #EndSummerCamp. I am grateful for the time I spent there and for the reception of Think Smol (despite its abysmal length). I loved the people, the food, the vibes.
Talk (in Italian): https://m.twitch.tv/videos/2257027599
Slides: http://3564020356.org/zelif/think_smol.pdf
Thanks everyone 💙
-
It took a while, but my gemini server for #PicoGopher is out!
https://github.com/aittalam/PicoGopher/tree/mala/adding_gemini
This is heavily inspired by Solderpunk's GeGoBi (https://git.sr.ht/~solderpunk/gegobi) and allows you to share BOTH capsules AND gopheroles via #GeminiProtocol ✨
The code is still a mess, but it was too cool not to share it! The merit goes to Solderpunk of course, while the mess is 100% my fault 😉
-
First #oldcomputerchallenge year for me and I am fashionably late with a live Knoppix 3.7 running on my Travelmate 514TXV. The Mozilla browser version dates back to 2004, which made me confident it still supported #gopher 🙂
The Netgear 802.11b pcmcia card 😅 could not connect to my AP so I used a #PicoGopher as an AP (did anyone say DIY?) and connected to both its gopher and its HTTP server with lynx and Mozilla.
This was fun, and a great way to spend my latest merged-PR endorphins 😁
-
This weekend I played with PG's gemini plugin so I could serve gwit sites while offline. I relied on solderpunk's GeGoBi (https://tildegit.org/solderpunk/gegobi) for the gemini server implementation, as I liked the idea of it being able to seamlessly serve gopherholes too 💙
Due to pico restrictions + lack of libs I had to simplify it quite a bit, but gemini seems easier to serve than gopher (if we ignore SSL...🤦♂️ which works great now!)
#PicoGopher #gemini #GeminiProtocol #gwit #SmallWeb #SelfHosting #offline -
I have been playing with gwit for a bit now, next step is testing it in PicoGopher. And as the very first gwit sites are gemini capsules… 😛
-
https://mastodon.social/@MozillaAI/111987207612394959
In our quest to democratize LLMs, two crucial efforts stand out: making open, trustworthy models which run on affordable hardware and evaluating their performance, in particular with respect to large, closed models. That's why I consider this collaboration in assessing LLMs trained under resource constraints a significant step for us. I am glad to share my team's latest (and first 😛) blog post, and I hope it sparks many more insightful discussions!
-
Some days my mind feels like those Emule/Torrent sessions with so many simultaneous downloads: so much stuff going on, nothing seems like it will ever be complete.
But today I mindfully reduced my downloads and let hyperfocus do the rest. It’s 2am now, but knowing that a few things are done finally gives me some piece of mind.
What I enjoyed the most this weekend? I finished reading Cory Doctorow’s Homeland and added AES support and a few new features to #ch35t 💪😁
-
Weekend plan:
- add Gemini support to #picogopher
- build web sandbox for #ch35t
- build a #brachiograph
- finish reading #homeland
- do some indoor bouldering
- play #turingcomplete
- play #projectzomboidWhat I actually did:
- an as beautiful as unplanned trip to Cliveden House and Windsor
- a bouldering session
- read Homeland (far from finishing it though!)
- read quite a lot of #manga
- took two loooong baths
- ate one Magnum (required for the brachiograph) -
I ran Falcon-180b on a freshly installed system (just to make sure it was not me having a weird set of library versions). I still have the same auth issue with HuggingFace Hub. It’s ok (my patch works!) but I wonder if out of all those talking about the new Falcon model (including a few who put together tutorials) there’s anyone who has tried to download it from HF and run it (not its demo!) since it has been made public 🤔
-
watching @blaine's talk at #ipfsthing about the Name Name Service #nns and getting more and more excited. A simple, decentralized keybase that makes DIDs usable? Sign me up
https://fission.codes/blog/prelude-to-ipfs-thing-2023/
https://talk.fission.codes/t/nns-the-name-name-system/3684 -
Thinking about #dynamicland and #realtalk : imagining a work/productivity interface made, in part, from two piles of stones and a candle
-
As someone whose #IFPS node stip dates back to 500BC or something, I didn't know that you can now update the config to use fewer connections: https://github.com/ipfs/ipfs-desktop/pull/2055
-
Vielleicht war er nie weg
Cathrin Kahlweit über die Rückkehr von Gerald #Fleischmann #ÖVP
https://www.sueddeutsche.de/medien/sebastian-kurz-gerald-fleischmann-message-control-1.5702725
-
-
-
Watching the #LindaLindas play https://mosswoodmeltdown.com/ with an entire audience of punk doting parents
-
proof that the #rationalists are either time-travellers, or constantly duelling immortals like in Highlander: https://www.lesswrong.com/posts/DtcbfwSrcewFubjxp/the-rationalists-of-the-1950s-and-before-also-called
-
@ruchowdh @KLdivergence welcome, it is great to see you here! 😊
Fediverse can feel a bit disconcerting at first, but you are not alone! Timeline is just reverse chron (no Al… yet 😜). Depending on the client you might see only or mostly users from your instance, so tailoring your network might take a bit.
The pace is a bit different too, no wonder why many slow Internet fans can be found around here (great #slowinternet fan here, ask me about Gemini and Gopher if you are interested 🙃)
-
TIL about #CarloAcutis who may become the first webmaster saint https://en.wikipedia.org/wiki/Carlo_Acutis
-
This weekend I played a bit with #click and click-repl to build a CLI interface for #ch35t: it is still early work but between help, command auto-completion and history it looks pretty cool 🤩 Can’t wait for this to be good enough so I can start sharing new riddles instead… Now the question is: how much is good enough? 🤔
-
All of this to say I pushed an update to #Ch35t and now you also have a README file to understand what the heck it is about 🙂
https://github.com/aittalam/ch35t
(also here’s the first post I wrote about it: https://fosstodon.org/@mala/111087465479136933)