klaudworks

Here to talk about tech: cloud, kubernetes, AI agents. Migrating companies off overpriced cloud offerings: AWS to Hetzner.

📍🇩🇪 living in Bansko 🇧🇬

[Dami-Defi @DamiDefi]

https://t.co/ksvC4bKpZ7

[klaudworks @klaudworks]

I don't think that there will be some kind of deal in the near future because it makes little strategic sense from either side.

I think Anthropic is more incentivized to do some circular hyperscaler deals for medium term compute because it also brings in funding. Or just pay big bucks for capacity that is available right now like SpaceX dc's.

Nebius is short compute and is more incentivized to a) sell it via token factory with higher margins b) distribute it across more smaller customers that grow with them. Both is more in line with their strategy.

Only benefit Nebius would gain is if Anthropic prepays for capacity. Anthropic however, is more concerned about getting compute right now and less about financing medium term compute deals.

I think once Nebius has more spare capacity it may happen but I think the chance of Anthropic prefunding a dedicated buildout at the moment is small.

[klaudworks @klaudworks]

yeah no need to condense it. What I appreciate is that you share authentic thought processes, erratic changes etc. It's pleasant to follow authentic human thought processes and how you arrive at decisions. Feels like I can actually follow what you are doing.

Thats kinda unique because most people just share part of the journey i.e. research ideas and occasionally that they entered /exited a position.

[klaudworks @klaudworks]

for the people who feel stuck micromanaging 1 or 2 claude code agents at the same time. Here a few key learnings on my side:

1. perform all your work inside your coding agent. After you performed any kind of repeatable procedure let your agent write a skill for it. With "repeatable procedure" I mean anything that you would write a playbook in your wiki page for so that a colleague can take over the task next time.

-> next time you have to perform the same procedure your agent will guide you through it or even fully take it over. This reduces the amount of brainpower you have to spend on it.

---

2. for bigger coding sessions look into different tools that help you to run the coding process more autonomously so that you can focus on what actually matters. I usually put a lot of thought into what I actually want to do. Then I write that into some kind of spec / initial plan that I then refine together with my agent. Once I'm happy with the plan, I break it up and feed it into some kind of workflow that iteratively works through it. E.g. for a coding task I use a tool I built called "ralph-meets-rex". It allows me to specify workflows such as: pick next issue from initial plan -> create dedicated plan for component / feature -> implement plan -> review (optionally go to implementation step again to fix issues)

-> I may spend hours to understand the problem and to create the plan. Then afterwards the agent may run for a day and I can do other stuff.

---

3. tests and validation is more important than ever. The better the guardrails for your agent the longer it can run autonomously without fucking up. For coding letting the agent write integrations tests against real testcontainers helped a lot.

-> every time the agent is able to self-correct saves you tokens, time and energy to debug.

---

4. any time you had to intervent because your agent got stuck somewhere tell it to improve the underlying skills, workflows etc.

-> Over time a lot of tasks will be fire and forget. As if you would delegate them to your super reliable senior colleague.

---

5. give your agent the tools it needs to perform the job. E.g. access to Kubernetes clusters, access to perplexity etc. The hard part is giving the agent power to do stuff without the possibility to fuck stuff up.

-> once you figured this out, you can oneshot lots of tasks. I'm an infra guy so a simple task may look like this: container xyz crashed because of oom error -> agent reads relevant information and kubernetes skill -> agent looks up additional information inside kubernetes and via perplexity -> agent creates PR with solution

---

You can use the brainpower you save to focus on running more agents or doing interactive, explorative work that helps you figure out what you actually want to do.

Any other good tips that I missed?

[roon @tszzl]

the frontier labs don’t have “comms problems”. reality right now has a comms problem. what is happening is a little scary and there’s no nice words anyone could say, especially not those profiting from it, that’ll make it feel that much better

[klaudworks @klaudworks]

@daniel_koss accumulated $NOW for 3 months, then sold it the day before it ripped to buy more $NBIS😂

[klaudworks @klaudworks]

@daniel_koss yeah financing would be massive for starting yet another project in parallel. Would be cool if they do something like 1gw in the nordics.

[klaudworks @klaudworks]

running kimi 2.6 via @Eigen_AI_Labs at 300-400tps is a blast if your task doesn't require frontier intelligence e.g. for a straight forward code refactoring

[klaudworks @klaudworks]

@daniel_koss by super easy to integrate EigenAI I meant: integrating the company into Nebius is super easy from a technical perspective. Very little friction in integrating a token as a service provider.

[klaudworks @klaudworks]

I like the EigenAI acquisition for that. I think tokens as a service will be the abstraction layer of choice for most companies in the medium term and thus the largest revenue driver.

I tried EigenAI and they did well with their kimi-2.6 model optimization. Really fun to work with a close to frontier model at 400 tokens per second.

Also super super easy to integrate EigenAI from a technical perspective.

[klaudworks @klaudworks]

you keep mentioning that. What makes you think we don't have truly useful reliable end to end agents already?

your stock research agent is something you can probably build in a day.

the capabilities are there but maybe Leopold Aschenbrenner was right in his famous Dwarkesh interview. His hypothesis was along the lines of: "LLMs will likely overshoot the required capabilities significantly before they will be widely used"

[klaudworks @klaudworks]

@BahloPatrick @daniel_koss bold take