I stress-tested Meta Muse until its agent control plane started timing out | Hacker News Reader

I stress-tested Meta Muse until its agent control plane started timing out

blog.cygankiewicz.com

8 pointsmpkc9 comments

This writing style, which reads like an LLM wrote it, is draining to read.

I put up with it when I'm actively using an LLM tool, because I just accept that this is how these tools write, but when reading a blog I expect to read words written in one of the vast varieties of normal human-written blog formats.

Fair point. English isn’t my first language, so I run my writing through tools to polish it. I probably leaned too hard into making this one auditable. The research and conclusions are mine, but yeah, it reads more like a technical report than a blog post.

One of the things that stood out in your article was the high amount of negations or contrasts. This is one of the worst hallmarks of AI writing, because humans tend to write in a more straightforward way, without as many caveats.

My two cents: you should avoid using AI to “polish” what you wrote. At most, instruct it to correct grammar mistakes only, instead of rephrasing your words, changing your arguments, turning paragraphs into bullet lists, “improving” sentences to make them more idiomatic… Nowadays, many technical readers would rather read the occasional non-idiomatic prose than put up with the soulless writing style of AI.

tekst po polsku czyta się niewiele lepiej xd

Yeah, it's exhausting.

> "The next two experiments look like a clean A/B test. They are not."

Fixed, thanks.

I dont fully understand how these black box agent computers will actually work.

One of the huge value adds of the agent computer paradigm is your agent can also fix and customize its own software.

I’m working on a similar agent computer and having the full source code plus a plugin system is the thing that really made it start working.

Something tells me Meta and Grok bots won’t give you this level of control.

More thoughts at this blog:

https://housecat.com/blog/agent-computer-101

I sort of understand DSH arch, and their choices, but still baffled by why you would want the agent to change its own agent loop other than the system prompt (in that extension, memory / soul / tools / skills).

It turns out a lot of the value is in the agent being able to stitch together pre-existing tools. I, too, have a personal agent but my friends using Muse or Instinct approach the functionality I use mine for quite easily.

The insight that enables this is a sophisticated coder can build tools an unsophisticated assistant can use. So our hosted model isn’t as smart but executes at 100+ tok/s single stream across 8 streams on our local hardware at frontier level.

But the code Fable provides is much better. The unsophisticated guy can build little tools but self-modification leads to a tower of cards.

Reply on news.ycombinator.com