35 hours later, the factory has delivered absolutely nothing of value and also not taught me anything about how to operate a better one.

I bet people on twitter were all "your doing it wrong bro"

In the middle of a conversation the agent ran into “Bad file descriptor” on a test and Astra decided it needs to see if file descriptors can be passed over Unix sockets on macOS in a super compressed manner

(...) it used Bash to run Python to then run Node.js to then use Node.js to invoke PowerShell.

Now that's worthy of a "what the fucking fuck is it even doing"

In the 35 hours it burned around 1B tokens for a total of around 1200 USD in raw API costs. It managed to produce 79 commits, and that comes to a cost of around 15.5 USD per commit, and the agents exchanged around 1400 messages.

I honestly do not need an agent to run for 35 hours on a single prompt. It clearly does not work or result in reasonable outputs.

Yup, it checks out with vibe coding in general

Ladies and gentlemen, the revolutionary technology that will generate 2 trillion dollars in revenue by 2030.

"Developer time savings" (the proxy for output/speed they repeatedly reference) seems to be self-reported time saved by using AI? That makes "AI output accounts for 63% of the variation in developer time savings, suggesting that developers with higher AI output save more time" seem a little questionable, given prior research on how good (or bad, or inconsistent) we are at quantifying the time AI may have saved us on a task. A flip thing to read from that might be "developers who use AI a lot are likely to believe and report that it saves them time." That isn't a very interesting conclusion, though suggests one answer to "What are developers doing with that time?": maybe they aren't actually saving as much time as they think they are.

(presumably their analysis gets into more detail than just that one metric, I know their product measures other things less subject to bias in a survey, but it's hard to know what from this post)

It's biggest output being illegally harvested data and misinformation bots/ malicious foreign trolls

Hell they use these bots to paint AI in a better light in general

Fantastic choice of name.

The LLM community heard harness when it came to debugging programs/datafeeds and immediately went down the “subjugate the LLM” plotline. Time to update language again so the finanically endowed don’t get hard-ons to further burn the world down.

Rewriting legacy applications is never easy, even with agents doing the heavy lifting. We got there by designing and building a strict test harness first, then letting the agent work inside it. Rigorous testing, strict boundary definition, and knowing when to step in as a human are what made this work, and those principles will hold up as the tooling keeps changing.

Holy shit, someone doing it the right way for once.

They are very close to what I'm looking for! Right now the major problems are:

• servers in the US. not sure why they're doing that. some models are in the EU, some in the US, and one might pick a model hosted in the US by mistake
• it requires a google or a github account: we can't use our corporate accounts to manage spending - specifically it seems like finance would have to create a github account to login on nebius. very bizarre

Are you using them by any chance? Do you know if their management interface has controls that allow using normal email addresses or some kind of user management without signing up to some extra service (github, google).

How about a cloud provider and then taking a open model to that?

That would require somebody managing that and there's no way my company would pay for that. From what I could find on Hetzner, it costs 0.4€/hour just for the instance. The rest will come on top.

https://www.experientiallabs.ai/ seems like a self-hosted wrapper around other paid services. You'd need to get an Anslopic subscription, a Goolag subscription, and so on.

I'm looking for an EU company like one of the companies on this list. DeepInfra, fal, basten, and so on. They show you a list of models they have a price, and all you have to do is sign up, whip out the credit card, copy paste the API URL into your harness (pi, opencode, claude code, whatever) and you're a hundreds lighter as a company but with no headache about running the entire thing yourself.

How about a cloud provider and then taking a open model to that? Also difficult to find, but it's better chance... Hetzner comes to mind. Then you may self-host ollama and some gateway to monitor token usage.

Even if it came to API key on a asian or american company, self host makes sense to me to relax on the lock in (e.g. Bifrost) and have some management dashboard. There's a growing ecosystem of self-hosted tools, the lates I saw was experientiallabs.ai that was aimed to save costs.

(6I'm on a similar track tho I am just a 1-man army with lended over pro plans (no apis) setting up omp.sh to combine and decouple tasks among providers and differentially sized models.)

The remnants of Yandex, but with a dutch HQ. They offer hosting regions in France, Finland and the UK among others. In March they raised $4bn to become “the AWS of AI”

https://tokenfactory.nebius.com/ did you see these guys?