[steve_adams_86]

I've been building a harness for qwen3.5:9b lately (to better understand how to create agentic tools/have fun) and I'm not going to use it instead of Opus 4.6 for my day job but it's remarkably useful for small tasks. And more than snappy enough on my equipment. It's a fun model to experiment with. I was previously using an old model from Meta and the contrast in capability is pretty crazy.

I like the idea of finding practical uses for it, but so far haven't managed to be creative enough. I'm so accustomed to using these things for programming.

[tempoponet]

What kind of small tasks do you find it's good at? My non-coding use of agents has been related to server admin, and my local-llm use-case is for 24/7 tasks that would be cost-prohibitive. So my best guess for this would be monitoring logs, security cameras, and general home automation tasks.

[steve_adams_86]

That's about it. The harness is still pretty rudimentary so I'm sure the system could be more capable, and that might reveal more interesting opportunities. I don't really know.

So far I've got it orchestrating a few instances to dig through logs, local emails, git repositories, and github to figure out what I've been doing and what I need to do. Opus is waayyy better at it, but Qwen does a good enough job to actually be useful.

I tried having it parse orders in emails and create a CSV of expenses, and that went pretty badly. I'm not sure why. The CSV was invalid and full of bunk entries by the end, almost every time. It missed a lot of expenses. It would parse out only 5 or 6 items of 7, for example. Opus and Sonnet do spectacular jobs on tasks like this, and do cool things like create lists of emails with orders then systematically ensure each line item within each email is accounted for, even without prompting to do so. It's an entirely different category of performance.

Automation is something I'd like to dabble in next, but all I can think of it being useful for is mapping commands (probably from voice) to tool calls, and the reality is I'd rather tap a button on my phone. My family might like being able to use voice commands, though. Otherwise, having it parse logs to determine how to act based on thresholds or something would also be far better implemented with simple algorithms. It's hard to find truly useful and clear fits for LLMs

[novok]

Oh man you just gave me an idea to use something like qwen 3.5 to categorize a lot of emails. You can keep the context small, do it per email and just churn through a lot of crap.

[nl]

The 0.8B can do this pretty well.

Actually pg's original "A plan for spam" explains how to do this with a Bayesian classifier.

[steve_adams_86]

I've been learning to apply these lately and it has been pretty eye opening. Combined with Fourier analysis (for example) you can do what seems kind of like magic, in my opinion. But it has been possible since long before LLMs showed up.

Totally different categories and different use cases, but the more I learn about LLMs the more I discover there's a powerful, determinsitic, well-established statistical model or two to do the same thing.

Really, LLMs are kind of like convenient, wildly inefficient proxies for useful processes. But I'm not convinced they should often end up as permanent fixtures of logical pipelines. Unless you're making a chat bot, I guess.

[1dom]

> Really, LLMs are kind of like convenient, wildly inefficient proxies for useful processes. But I'm not convinced they should often end up as permanent fixtures of logical pipelines. Unless you're making a chat bot, I guess.

I think I agree with this. It's made me realise LLMs are great for prototyping processes in the same way that 3D printers are great at prototyping physical things. They make it quick and easy to get something close enough to see the unforeseen problems a proper solution might have.

[novok]

IMO the fact that spam detection has devolved into reputation management vs. being able to work on the content themselves makes me think there is a lot of alpha between an llm process vs. the most traditional processes we have now.

[alexpotato]

I was just chatting with a co-worker that wanted to run a LLM locally to classify a bunch of text. He was worried about spending too many tokens though.

I asked him why he didn't just have the LLM build him a python ML library based classifier instead.

The LLMs are great but you can also build supporting tools so that:

- you use fewer tokens

- it's deterministic

- you as the human can also use the tools

- it's faster b/c the LLM isn't "shamboozling" every time you need to do the same task.

[vidarh]

I use Haiku to classify my mail - it's way overkill, but also doesn't require training unlike a classifer. I recieve many dozens of e-mails a day, and it's burned on average ~$3 worth of tokens per month. I'll probably switch that to a cheaper model soon, but it's cheap enough the "payoff" from spending the time optimizing it is long.

[nunodonato]

you can use 4B for that, its quite good