planckscnst

And you can also rant about it so other people are not harmed in the same way without warning

Is that because every country fails to adequately charge oil/coal/etc for its externalities? Are we essentially subsidizing oil, bringing the costs artificially low such that nuclear can't compete?

"genuine" is another one in the same vein

Yes, it's different. It's a legal doctrine called mosaic theory.

https://share.google/f10FzHcf2jPqSzTRM

Yes, I used it to do a big push of my self-maintaining project All I need next is to push it through enough cycles to trust it (I have high confidence based on results so far) and then I can just setup a cron to do routine maintenance on my project

For people wondering, in some locations, sore and saw are homophones.

I have the $100 plan on OpenCode Zen, which I imagine is similar. I burn through it ridiculously fast. I haven't run the numbers, but it felt about the same as an Anthropic $20 plan

LLMs are already of sufficient capability that if they were super cheap and fast and made no progress on intelligence, we will still make huge leaps in our ability to build systems and solve problems.

I'm just happy there is capital behind the efficiency path because I am confident we can get positive value on that path.

Do you not have many many separate projects happening in parallel because it takes forever for the model to respond, so you give it your feedback and jump to the next one? That context switch is challenging and expensive. Imagine if your feedback was nearly instantly applied and you could just see the result? It would be more like the holodeck metal table scene in Star Trek "Schisms".

You can make something like the link below. I'd say it's a more than marginal improvement. It's still tiring, but it's much better according to my taste. I imagine everyone would have their own version of this for their own preferences.

It's a loop that uses adversarial review to check several dimensions of the writing:

https://github.com/Vibecodelicious/llm-conductor/blob/main/w...

$100/month * 12 months is 1200/year. Yeah, normal working people even in the USA don't have that extra money to spend. And let's be realistic: anyone doing the "greenhouse" method probably has both Anthropic and OpenAI $200/month subscriptions, otherwise, this is going to use up your tokens pretty fast. So $4800/year to have this as a normal practice that you do.

The median household income in the US is $64000/year.

It depends on what you mean by "unsupervised" - I've been strictly working through the agent, through specs the entire time, but it's been very supervised, I just leave the mistakes in-place and have it work from there.

However at this point it can completely maintain itself. When a new version of Claude or OpenCode is released, it updates itself to work on the latest version. It can also add new implementations for harnesses pretty reliably. It's actually pretty fun to watch it at this point. "Make this work on Hermes agent and message me when you're done" and an hour later or so, I can go play with it in Hermes.

I did calculations on the prefix cache effect on costs of sessions where I used it and found that the removal of tokens from context had a much bigger effect on reducing costs than cache busting had on increasing them. I should re-do that and publish it.

I agree. Compaction sucks, so I made tools that let the LLM selectively delete (and recall if needed) chunks of its context. You might want to try context bonsai if you're routinely hitting the auto-compaction wall.

https://github.com/Vibecodelicious/context-bonsai-agents

You don't desire to brush your teeth (at least not in the way that you desire to consume calorie-dense foods). But you manage to still do it anyway. (maybe not you specifically, but people in general) You can make the same choice about nutrition. The lowered desire makes it easier/possible to do this

Hm. I'll have to rethink how I do context management ( https://github.com/Vibecodelicious/context-bonsai-agents#con... )

Yes, also happening more or less continually instead of waiting for the context to bloat up.

You also have some control over it. You can say something like "prune away the work we did on the UI. Make sure to remember that users can upload multiple images now"

I'm working on [Context Bonsai][1] - LLM harness tools that allow the LLM to prune messages out of the context, leaving behind a summary and keywords instead. In addition to a "prune" tool, there is a "retrieve" tool that allows it to recall the messages if needed.

In addition to these tools, I'm also building automation that will port the tools from the reference implementation (OpenCode) to other harnesses (Claude Code, Cline, Pi, Gemini, Kilo, Codex, others to come?). As well as automation that will either cherry-pick or re-implement commits onto the latest head from upstream.

[1]: https://github.com/Vibecodelicious/context-bonsai-agents#con...

[2]: https://blog.vibecodelicio.us/posts/how-i-fixed-context-wind...

I love OpenCode! I wrote a plugin that adds two tools: prune and retrieve. Prune lets the LLM select messages to remove from the conversation and replace with a summary and key terms. The retrieve tool lets it get those original messages back in case they're needed. I've been livestreaming the development and using it on side projects to make sure it's actually effective... And it turns out it really is! It feels like working with an infinite context window.

https://www.youtube.com/live/z0JYVTAqeQM?si=oLvyLlZiFLTxL7p0

I'm using Fedora with KDE and I haven't seen any notifications and had no idea it did this. I'll see if I can figure out what's going on in my system and maybe it will help other people.

I am one of those people who love VR gaming done well. There is a game called Super Rumble built by what I think is a subsidiary of Meta. It's a very well executed arena FPS concept. There are just a couple dozen people in the world who are really skilled and play enough for me to recognize them and be glad they're playing when I'm also online. It's magical when there are good people on this thing playing together.

I hope it's something we can figure out how to propagate despite the seemingly limited interest. I suspect anyone who liked playing quake arena games would love this game if they are not susceptible to motion sickness.

I recently started exploring how to port open source shooters (red eclipse, warsow, nexuiz) to the platform and realized there are several considerations that make games designed for VR special that a pure port wouldn't hit.

I'm working on "context bonsai" which is currently a plugin for OpenCode that allows the LLM to self-edit its own context. It works like compaction, but it can retrieve back the compacted info if needed. And it's not just when the context is completely full, and it doesn't compact the entire context - it picks messages / tool calls where the details are no longer necessary, like a debugging session that is already solved or feature implementation that is complete and you've started on implementing the next feature.

I've also used tweakcc to make this work in Calude Code and plan to also do one for open source coding agents - codex, pi, Gemini, etc. And I'm also doing Livestreams of the development process.

https://github.com/Vibecodelicious/opencode

I mostly use LLMs in a zero-touch way - I never actually edit code and I almost never read it. But I do still dive into the details by exploring it with targeted questions through the LLM. Sometimes I go through ridiculously long sessions to get the LLM to "see" the correct/optimal/simplest/etc solution itself. There are many times when it simply never gets there no matter how close I get the horse to the water. I recently did one of these sessions yesterday and it reinforced my impression that systems like gastown and pure Ralph loop style is just not ever going to have the quality I'm looking for, and it's going to cost a lot of money not to get there.

I've honed a relatively decent flow that requires interaction from me for important parts (mostly) while making its own decisions at the not-important parts (mostly). This results in being able to send the agent off on an hours-long dev cycle and have relatively decent results after that need a few minor fixes. I think this is the best style for the current generation of AI

When they blocked OpenCode, I was in the middle of adding a feature. I don't think it's possible to mimic CC in an undetectable way and have the feature work.

The feature allows the LLM to edit the context. For example, you can "compact" just portions of the conversation and replace it with a summary. Anthropic can see that the conversation suddenly doesn't share the same history as previous API calls.

In fact, I ported the feature to Claude Code using tweakcc, so it literally _is_ Claude Code. After a couple days they started blocking that with the same message that they send when they block third party tools.

They even block Claude Code of you've modified it via tweakcc. When they blocked OpenCode, I ported a feature I wanted to Claude Code so I could continue using that feature. After a couple days, they started blocking it with the same message that OpenCode gets. I'm going to go down to the $20 plan and shift most of my work to OpenAI/ChatGPT because of this. The harness features matter more to me than model differences in the current generation.

"selected" and "highlighted" would also be useful

There is so much work we can do with harnesses that can make the already existing models so much more capable. I definitely feel the author's frustration as I've also been working on some harness stuff. When Anthropic subscriptions got cut off from OpenCode and other third party tools, I was very disappointed because the model I do the most work in is Claude and I was specifically developing a change [1] in the hopes it would make Claude even better. After that, I started implementing the feature in Claude Code directly (using tweakcc) and after a day of working on that, they even block my tweaked Claude Code with the same message. It means I simply won't be able to use this idea with Claude at all

[1]: the README.md describes the Context Bonsai features in my fork here: https://github.com/Vibecodelicious/opencode

Yes - I think there is untapped potential into figuring out how to understand and use the latent space. I'm still at the language layer. I occasionally stumble across something that seems to tap into something deeper and I'm getting better at finding those. But direct observability and actuation of those lower layers is an area that I think is going to be very fruitful of we can figure it out

Yes, I have already made deliberate cache decisions and plan to do more once it's working the way I imagine. I think the trimmed down context will have way bigger effect than the cache stuff, though.

As far as I understand, it's caches are not a "next-turn" thing, but a ttl thing.

I made the "retrieve" tool, which is what pulls back previously removed content, append to the conversation rather than putting it back where it previously was. But it's a but premature to really know if that's a real optimization.

I'm working on lots of projects. My favorite is what I call "context bonsai" where I'm giving LLM harnesses the ability to surgically edit the context. It's available as a tool. You can say "remove that failed debugging session and write a summary of what we learned." Or you can take a more hands-on approach and say "remove messages msg_ID1 through msg_ID2". The removal leaves a summary and keywords, and the original messages can be pulled back into context if the LLM thinks they're useful.

I would really like people to try it out and report bugs, failures, and successes.

https://github.com/Vibecodelicious/opencode/blob/surgical_co...

I'm currently trying to get the LLM to be more proactive about removing content that is no longer useful in order to stay ahead of autocompaction and also just to keep the context window small and focused in general.