Hacker News Reader

Comment by petra | Hacker News Reader

Maybe, for some projects, instead of generating code with it, it would be useful to generate a plan and the loop(tests/formal verification),because those take much less tokens than a full project, and than use the loop using the older models ?

View on HN

Yes, I've been using Opus to write a plan and fanout sonnet subagents to implement it. Cheaper and faster

What about quality? Being cheaper and faster, while great and all, is less valuable than quality to me.

All the code an LLM produces is of questionable quality, so I'm not sure why you'd prioritize quality over speed. Speed is their only value add.

Have you compared models across providers? The quality, for the same task, varies tremendously. If you don't prioritize quality you're wasting your own time when you inevitably have to re-do the code...

This is a wild take. All cars can perfectly drive around a track, so why would you ever want an F1 car?

All code is of questionable quality.

There I fixed it

You can always have opus review the result at the end

I've had bad results with this method. If I ask any model to review the same code ten times, then I will get ten different answer.

This is the goal behind Devin Fusion, pretty good results so far I think.

https://cognition.com/blog/devin-fusion

so, pretty much undo the "magic" that the harness is for

Has anyone experimented with Batch Processing? According to https://claude.com/pricing#api using Batch processing cuts the price 50%. So I wonder if any of the harnesses like OpenCode/Pi or similar could be made to use that for planning or similar.

Batch can take up to 24 hours (and often does) and may never complete if it gets cancelled so it’d be hard to build a user workflow around unless you kick off planning on Friday and come back Monday

this is the idea of opusplan https://code.claude.com/docs/en/model-config#opusplan-model-...

Article has a section about context window size settings

I love not getting compacted so often, but 1M context is trash right now, the degradation in speed and quality is too great above ~600k context

Not different than what everyone knows, but the 1M context is masqueraded as an innovation the same way 64k context used to be to 8k context

Yeah, I tried using it before the 'safety' blocks got too much for me. Its pretty good as an 'orchestrator' calling subagents and being the PR reviewer/tast tracker. Seems really not that much better for coding though, yeah it is a bit better, but not worth it.

But when it did work it was making steady progress on a vague git issue backlog and actually following instructions to carefully break into atomic sub issues and carefully assign.

Isn't that the kind thing its best at as well? Art least comparatively with other models. The more agentic stuff. Planning, tool orchestration, etc.

Did that today on a 2 repo affecting project of the kind where I already set the right design for one major use case and I needed Claude to create a superset of that use case that was not substantially different: after plan I had about 10% of 5h context left for fable 5 and this was the only thing I worked on. Hard to generalize this of course.

But wouldn't that still result in higher token usage to scan the code base and figure out the changes and generate the plan? In my experience sometimes Opus launchs a Haiku sub-agent to explore the code base, but it's not gaurenteed.

I think that's the idea, I saw some outrage on reddit about Fable using Opus to do code writing, another comment said exactly my reaction, why do you want to pay double for tool calling when Opus is just fine for the task?

Yes, I do this all the time in Cline. It supports automatic model change when switching from Plan mode to Act (implementation) mode. Opus for planning and Sonnet for implementation. It works great.

Reply on news.ycombinator.com