Comment by bitexploder | Hacker News Reader

Comment by bitexploder | Hacker News Reader

Parent Full thread

I have a custom A3B proxy that caps its thinking off. It is a known issue with the model that Qwen themselves documented but is almost never addressed in any harnesses. I also patched up a few other known bugs in the proxy. I highly recommend you shim A3B and when it hits 2K thinking tokens inject (paraphrasing) 'Time to wrap it up bud! Get to work' into its thinking stream. It almost always gets to work. If it needs more time to think there is always next turn.

In my experience it is almost never productively thinking past that point, just spinning in circles. I also reinject all of the thinking. And there are a few tells that it is getting stuck. I have an optional mode that takes the last few turns and tool calls and shoots it off to DSV4 with a prompt to basically understand where it is at and inject better thinking and or planning. It almost always gets it over relatively difficult humps, but some of the time I don't want things going remote. It might end up with 10-30 cents of DSV4 calls over a hours and the quality improvement is remarkable.

The other thing is I trick it into thinking a web_search tool is a web search but it really just asks DSV4 the prompt. DSV4 is a cheap filter to help prevent prompt injection lol. You can give it other models but DSV4 is my cheap-mode default.

edit: oh! My final 35B A3B tip -- use an extremely simple harness. Pi is good. Pi's default tools almost exactly match what Qwen says they tested the model with (likely meaning that tool set is also what they trained it with or something similar). So, in my experience bigger harnesses don't have a noticeable improve ment on tasks.

Fascinating, thank you. I am trying to switch to pi from opencode (my own thinking loops and burnout are a challenge lately).

It had not occurred to me that you could nudge it to stop thinking with a proxy. Nice idea.

Will favourite your comment and come back to it.

ETA: Incidentally you've helped me put into words the difference between the way Muse Glimmer thinks to the way Qwen thinks. There is a clear sense of urgency in Glimmer's thinking traces.

Having spent a good part of the day with it, glimmer reminds me of Rorschach from The Watchmen. No unessential parts of speech, action oriented, brief and to the point. From a token perspective anyway it’s great, and it seems to hold its own well against more verbose models.

I really do feel like it’s effective tok / s is way higher because it doesn’t waste them.

I am very struck by the way open weights LLMs seem to reflect a culture.

I don't really enjoy the way Qwen writes prose, and I find its thinking a bit exhausting, though it clearly writes very good code.

I like the neutral, clear way the Gemma models write, which I sometimes use to get myself a "getting started" document on something I want to understand; it also summarises well. It is neutral, sensible, un-showy. It writes in a way that is fairly close to what I would use for documentation. The 12B and 26B models are also very good for talking about art and photography. Analysing my own photographic work has helped me more than I expected it to.

This model, honestly, has made me smile. It also feels like it is more creative at a given temperature than Gemma. I am trying to motivate myself to do something quite open-ended so I asked it about what other people's considerations might be in my situation, and at the risk of anthropomorphising, the things it has come up with feel like the work of a more curious mind, somehow. More eclectic. I have enjoyed testing it and I really want to test it more, which might help me get over a motivation hump there, too.

(I am also exploring its hard-wired policies by asking it to analyse some studio art nude work I have done; it definitely thinks out loud about its policies in a way I have not seen Gemma do.)

I think we’re going to see a lot more “product“ focus in the future with deliberate attention paid to these kind of properties. Historically though there are some obvious differences, the focus has been on benchmark maximizing. As that saturates, I expect more interesting choices about writing and thinking style designed to be differentiators instead of a side effect. Kudos to the PM here for taking it in a different directions, there’s obviously been thought put into it.

I ran a 9B over my like 100k photo library — it was very good at it. And extracting any text. All local.

It might be an indication that analysis of photography is something of an ideal discipline for an LLM since so much content online involves discussion of pictures.

A lot of what I am trying to do with my photography is sort of meta-photography. I am really interested in early photographic history, pictorialism and its opponents etc., but I try to avoid reproduction, so I try not to simulate processes too closely or to use vintage tropes in props and settings, but I use simple, undercorrected lenses and some vintage lenses, to gain some of the visual language.

Finding out that LLMs (including Gemma-4 12B with its built-in image encoder) understands a lot of my references and influences, could recommend me my (still semi-obscure) favourite historical photographer and other photographers who clearly engage in the same work, is amazing. And sometimes it says stuff I had not thought of, which is what I am looking for, since my photographic journey is somewhat lonely.

And that is just sort of brushing past the fact that these things can describe the contents of photographs with an accuracy that you can almost take for granted.

I would also think about combining the descriptions of photographs with some sort of RAG to essentially help shoot you into the reference references easier.

You really have to get the models to end their thinking. Almost any commercial model serving has safe guards like this to tune how much they think.

Reply on news.ycombinator.com