Galo Navarro (@srvaroa) header
Galo Navarro (@srvaroa) profile picture

GPU infrastructure & AI workload orchestration @midokura | Distributed systems / software delivery / cloud

Europe
Joined May 2010
Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)Photo by Galo Navarro (@srvaroa)
New post: “We review every PR with a 27B model on self-hosted infra” varoa.net/2026/11/03/we-revi…
1
49
Galo Navarro retweeted
Golden age of SRE!
5
24
3
185
13,162
Code is easy, show me the prompt
51
Reminder: suspect interviews that require installing stuff on your laptop
genuinely impressed by the level of sophistication of this hack. this is an entirely different league from those pathetic CVE exploits
74
"we find human reviews less valuable, but are reluctant to merge based on automated reviews. We're not ready for the risk of understanding a complex system less and less" 1. you already don't understand a vast surface of it, but got used to a certain altitude ->
1
91
2. every source of discomfort is an item in your roadmap
22
Galo Navarro retweeted
Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations. Basically one-shotted a full NN that’s better than training it
258
372
235
7,770
909,597
Working on our code review system I scanned the literature for some evidence on adversarial reviews pro arxiv.org/abs/2603.04582, or arxiv.org/abs/2508.06709 skeptic arxiv.org/abs/2605.29800 or arxiv.org/abs/2606.20093 ->
maybe a hot take but i now consider adversarial reviews mostly a waste of tokens. far better to use those tokens on empirical combined with formal verification. i have updated pstack to use the interrogate skill less, and eventually i think i may remove it altogether i have found very little value with code review in the abstract, agents tend to just get lost in the nitpicks and yagnis
1
67
IME, different model families (e.g. qwen vs. anthropics) and fresh context gives richer reviews. Agree that, long term, investing tokens on repeatable tests and formal verification wins, but it's a different, complementary concern? (If you review, adversarial ones can be useful)
71
Galo Navarro retweeted
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list. Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%. Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything. Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO. h/t @OliverMolander Would you move your team off frontier APIs if the bill dropped 90%?
306
694
386
7,810
1,014,306
No objection that AI can manage infra, but it’s like saying that code is dead because AIs can write it Anyone managing infra (person or AI) will seek to materialize decisions in ways that keep behaviour predictable. That means getting out of inference and into code
IaC is dead. Sorry DevOps / SREs / Cloud Engineers but you're in for the same roller coaster as Devs. AI can manage infra change better than you. Unfortunately for IaC, we don't need code because with the actual system source code and the infra, AIs have everything needed to determine change sets / approvals and dev env lensing. btw, Effect Oriented Programming probably will help you get there; but that is inevitable for your codebase.
91
In fact we want as many layers as possible fully reproducible (and predictable).
IaC isn't dying, it's becoming more essential. Non-deterministic layers add new volatility to the stack, but they don't break the core principles. For example, it's becoming more obvious that we want some layers to be truly reproducible.
1
69
We are building AX to be a Kubernetes for agents to be able to manage other agents. Our assumption is workloads will be provisioned and managed primarily via other agents not by humans. Agentic loop is the new job controller loop.
38
24
5
358
35,806
I'm still surprised every time I hear anyone in a professional context considering to inject a middleman in their token streams, however useful these things are
The security sensitivity of sites like OpenRouter is insane; imagine the damage you could do by inserting shell commands into people's token streams or just logging and ransoming people's data. Who needs to poison model weights when you can implant yourself within inference providers
71
Automation in global logistics is a good example. It didn't take off until the 1960s. Machines existed for much longer, but were shoehorned in an alien ecosystem designed for human labor. What did the trick was changing the unit of work: the container optimized for machines.
Replying to @thorstenball
Okay, I could've phrased this better. I didn't mean: why are you working on things the agent can't do yet? That has an obvious answer: because the agent can't do them yet. What I meant: there are people complaining that - the agents aren't good with our dev setup - the agents aren't good with our tooling - the agents aren't good with this platform we're working on - the agents aren't good with this language/framework/tool And my point is: if that's the case, you should try as hard as possible to get to a place where you can use agents at their best. If they aren't good at writing native applications with your bespoke GUI framework and your database no one's ever heard of -- why are you working on that? Is it truly worth it if agents aren't good at it? Is it truly worth leaving the biggest developer productivity multipler we've seen in many decades on the table?
70
Got the robots rate limited by GH today. Bottlenecks, bottlenecks everywhere.
51
Galo Navarro retweeted
Staff engineers need to connect technical choices to user needs and economics, then test their assumptions instead of waiting for product to resolve the ambiguity for them. @Lethain with @GergelyOrosz on @Pragmatic_Eng
1
7
29
5,092
Early arch choices inevitably become bad ones with scale (see @JeffDean 's deck on scaling x to +100x). Which is why the ~ most important architectural property of early systems is flexibility to be dismantled and rebuilt with low cost.
Maybe unpopular but yet to see any startup fail b/c architecture choices. I’ve seen a lot fail because of not being able to get users/customers; no real focus on what to build; not finding any traction. When traction hits, even terrible tech choices scale and can be fixed later.
1
62
Weekly reminder: don't use floating versions
Also, this is another time where I would NOT recommend using floating major/minor versions in your Dockerfiles For example: node:22 Why? You don't know if the Docker cache has been cleared and what version you're running! Better to be explicit: node:22.23.3
43
A side effect of AI code reviews is that now someone actually checks for implementation changes that alter past decisions in ADRs and TDs. Obsessively, and without mercy. Which makes those docs even more valuable.
1
3
105