@schmic @mnalis If you have a lot of time, all of our discussions on the topic are even web-public in our Zulip 😅
But to potentially spare you this and give a summary: Part of the reason for why it took us so long with moving forward on making a decision on this was that we worked *really hard* on finding an EU-based home as another option to consider.
But ultimately without much success and felt it was time to not let the perfect be the enemy of the good!
In inference I hit to Rust bottlenecks. It is an enormous effort to make all loops error check free in order to not generate boilerplate. It's really hard to optimize against icache when using Rust TBH.
I spot all the time ".context()" (anyhow) or try-statements and similar cruft. When I work around not having to do any of these checks, inference is suddenly like 1-3% faster. This is ridiculous. Most of my time is now spent on Rust performance issues.
What I will do is to create a C library with the shared primitives (like matrices, tensors, fp16 and similar), which will then be used by ReadSeek and Goosedump to implement inference engines for Qwen3-VL-2B and GPT-OSS-20B.
@schmic @mnalis We had a long list of potential hosts including EU ones, you can see the full list here: https://codeberg.org/comaps/Governance/wiki/Umbrella-Orgs
For many our model of supporting ourselves by having many small donations was not interesting/too much work. And we tried for many months with TCC but never managed to move ahead due to slow response times, which made them a no-go for us.
The majority of the model "market" will focus on inference cost as a metric, not the model itself. Context size, tokens — those will become performance metrics, not price defining. And we will see more and more open-weight and open-data models that are highly domain specific. I will flesh these thoughts out in a longer blog post, but I guess most of you get the gist already. 2/2
My current prediction: In-house and hosted inference of open-weight models will define the market, not "big" AI. This will also lead to a model exchange and less a market. Sure, some models will be proprietary but they will also feel the pressure of being allowed to run on-premise for a fixed (subscription) price, not per token. 1/2
ACN: Aumentano gli eventi cyber e le notifiche nel primo semestre 2026, calano ransomware e DDoS
Nel primo semestre del 2026, l'Agenzia per la Cybersicurezza Nazionale (ACN) ha gestito 2.171 eventi cyber, con un aumento del 47% rispetto allo stesso periodo del 2025.
In celebration of #FursuitFriday, let it be known that it is not only in Germany that furries will party on a tram, we also have a bunch of suspiciously wealthy furries chartering streetcars to party on the streets of Toronto (and this, my friends, is why we absolutely need to bring trams back to Montréal):
"AI" companies buying used books to scan and destroy is a weird flex to avoid copyright problems (at least they hope), but it is not the big scandal some want it to be — in my personal opinion. They are used books that sat in storage. They are bought in bulk. They are not rare books that will be irretrievably lost to humanity.
