someone had to mention @Hetzner_Online to get your attention, huh? i’ve DM’ed all details. your colleague Duarte requested an “urgent escalation” yesterday, which remained pending til Monday working hours. well, a full workday has passed and still no news. btw, your live chat kept me waiting today for 5 minutes and then told me the session expired and to try again later, lol. pure comedy. I guess today is National Croissant Day?

Yes, the right design principles are: - to push as much adversarial logic to the edge, out of the app core - to make it scale independently of the app itself - to make it scale horizontally, so you can dynamically add more rate limiter instances under peak load - to adopt a stateless model if possible: auth token checks and rate limits should avoid db accesses - to make to async and fuzzy: trade 1-2 requests over quota slipping in occasionally to eliminate a sync path

award to the most European company on Earth. timeline reconstructed: - dude scheduled immediate drive replacement on Friday evening - assignee said: nah, I can’t be bothered, i have plans to drink wine and smoke in 1h - they don’t work on weekends - their “24/7 service” is a façade - they just inform you that critical issues wait til Monday kudos @ovh_support_en

lol, @OVHcloud literally ghosted me. so incompetent. - disk controller fails Friday AM - kernel marks disk as dead - i backup and open a ticket to replace the nvme - in 5 min, guy named “Loick” marks as critical - schedules the replacement in 10min - 24h passes, nothing has happened - nobody answers; critical ticket remains open (beautiful SLA!) to be fair, this being France 🇫🇷 I guess they had to rush out to an urgent demonstration, or spontaneously decided to strike

ok, I’ll go against the grain: Effect is quickly turning into Java’s Spring Framework. Premature abstractions are bad, and common interfaces obstruct optimizations. And it makes zero sense in 2026, where software moves at the speed of light and code is free.

kinda ironic that Europe sacrificed its energy production to fight global warming. yet winter is still cold in 2026, and now we’re short of energy. to add insult to injury, we’re surrounded by wars and weaker than ever. your genius must be studied, @vonderleyen, Merkel, and the green lobby frens 👏

2-level tree: sol high/xhigh as the parent (orchestration, integration, etc.) luna max fast as the leaves (impl, scouting, work, etc.) astra high/xhigh in a dotted line to the parent (sol asks it for the initial plan, then continuously resumes at checkpoints to review outputs, and recalibrate the plan forward) fresh ephemeral astra xhigh when something needs a fresh perspective

2-level tree: sol high/xhigh as the parent (orchestration, integration, etc.) luna max fast as the leaves (impl, scouting, work, etc.) astra high/xhigh in a dotted line to the parent (sol asks it for the initial plan, then continuously resumes at checkpoints to review outputs, and recalibrate the plan forward) fresh ephemeral astra xhigh when something needs a fresh perspective

what a bunch of FUD. AI powerful enough to kill us isn’t small, and its footprint doesn’t go unnoticed. sure, AIs can and will plan a takeover, but they still need (a) compute infra to run and (b) network access and pipes to replicate themselves. and it’s not like there is heaps of high-end bare metal or unmonitored networks just lying around waiting for them to take over. in other words, AIs leave a traces as soon as they touch the world because they are physically-contained things. and labs watch those traces closely, more so after the recent bulletin-board “surprises”

the only right way to use Astra is to occasionally escalate to it at xhigh/max: - to review some technical design - to seed fresh, sophisticated ideas when you’re otherwise stuck in a rut - to cry for help to reorganize parts of a shitty codebase - to have dumber agents pull tasks from it

this is the kind of systems intuition I’m afraid we are losing - read/write-through caches tend to be optimizations, not authority - the system can work without them, just slower - we design systems to continue operating even when the cache goes offline (graceful degradation) - therefore, durability should be disabled from day 0, because: - when you recover, your snapshot would be stale anyway, so it’s unsafe to use it - hence it was useless to create it in the first place

as incredible as they are, let’s not forget that LLMs are, ultimately, token-generation machines. two places where this really shows: writing: - they can, and will, write incessantly - they’re built for continuation; extend their budgets, and they’ll go on forever - it’s our job to moderate them, and to know when to stop; aka judgement the beaten path: - everything in context conditions what is generated next - as a session unfolds, the conditioning accumulates, and the space of plausible continuations narrows - this makes it surprisingly difficult to turn a mature session against itself - major corrections, adversarial audits, radically different framings; all become harder to access - the model gets “wedged” in a kind of “basin”: increasingly biased toward a small fan-out of future trajectories - effectively, it becomes “obsessed” with just a region of the solution space - you can sometimes unwedge it by substantially perturbing its context, but at that point you’re better off starting fresh than fighting an uphill battle (literally!) in other words: long sessions buy coherence, but they impose major path dependence once a session takes a turn you don’t like, scrap it. or rewind it to a known-good state, and proceed from there.

Love this! I believe tiny models have a massive place in the future (pun intended). There’s a version of the world in which mega models serve as “root”, tasked with training and birthing tiny hyperspecialized models on-demand, for users. (Imagine if each one of your skills was a model in itself.) In that world, the routing and composability layer becomes the centerpiece. You execute a task by resolving relevant models, chaining them together, and performing cascading, streaming inference. And “made in the EU” is the cherry on the top. You risk becoming irrelevant due to the lack of compute infra? Flip it around and turn frugality into a superpower.