sgc

No idea, but yes it seems like good practice to trim context there as well. A good candidate for v2 if it's not present.

At the same time, even if that is not done, the todo is both specification and tests. If it passes those when it should not, then either you approved something you should not have, or else there should only be smaller improvements to be made. I like the concept enough and I am not very concerned about this, so I will give it a test on a project I have had on backburner for a bit.

Of course I have been down the rabbit hole. The software is wrong. Give me an option to bypass re-encoding, and instead of failing based on diagnostics like it does, just play the darn file the best you can. I only have so much patience and I prefer living with a samba share to play it in the exact same mpv player without problems rather than fighting software over something that is so simple and fundamental to a video player.

Don’t make me report you to a submod.

This is pathologically outside scope. I don't think I have ever seen somebody threaten somebody else on hn before. The 'don't make me do it to you' abuser trope is next level.

I have a problem that it constantly tries to re-encode video that could be played perfectly fine in their player or the mpv shim. I wound up just mounting my remote drives and playing in my local mpv. I don't really use it because of that. Unfortunate because otherwise I love a lot of things about it.

Actually it does include an important level of review. They say "init a TODO list with a validation step for each item", and then the last several sections of the execution are within the TODO - where it is validating the code.

To me that was the heart of the article and of what they did - create a self-validating todo list that deterministically forces the agents to stay on task, and then trim out just the right extraneous context (the actual planning) without messing with the read-file context which is where they found subagents burning most of their tokens.

Just a few days ago I was speculating about forking context to implement instead of spawning subagents here on hn. This is similar to that just several more steps more sophisticated. You could even fork the trimmed context here for very large tasks.

I think you missed my point, which is that in the US, people typically described as nationalists tend to be pseudo-nationalists who value pomp and ceremony, but not substantial concrete actions to better their country or actual real care and love for their fellow countrymen. In terms of percentage of the population who value and love their national identity, we are no different than anywhere else.

See https://www.pewresearch.org/global/2026/02/17/what-makes-peo... and especially note that the US is one of the top countries for percentage of population with primarily negative views of their country, at 20%.

I am American who has lived in many countries around the world, and I think this is distinctly wrong and the source of many problems in the US.

It would be more correct to say that the average American values outward displays of nationalism more, and has a more negative perception of those who do not appreciate or want to participate in those displays than people in most other countries. And yes, they conflate this with 'patriotism'. However, this is almost completely performative and lacks real substance, as is proven by the typically far more selfish attitude towards their fellow citizens, and is exemplified by the constant historical failures to provide significant funding for projects designed to help rather than harm others.

Europeans and people from other countries around the world are often fiercely in love with their countries. They just tend not to love the idea of noisily jumping up to gaudily beat their own drum. So yes, the average American thinks they are more nationalistic, when in fact they are just more tribal and crude about their nationalism than what is typically found in other countries around the world. If only our nationalism were taken a bit more seriously than our affiliation with a sports team, which is in theory just for fun and entertainment, that would be an improvement.

A bit late, but I followed up on this useful tip and found a gist that breaks down the opencode methodology:

https://gist.github.com/rmk40/cde7a98c1c90614a27478216cc0155...

The gist led me to the opencode session / prompt control folder:

https://github.com/anomalyco/opencode/tree/dev/packages/open...

It seems like there could be a useful strategy of writing a plan with a main agent, and then instead of spawning subagents to implement, fork the main context to write each part. Then use one last fork to verify the work. That way you keep reusing the same context without polluting your main context for when you are ready to continue.

It seems like they are actually using the subscription providers' respective cli tools and managing context for them. In which case I believe it is not against the ToS any more than invoking codex cli from a custom python script would be? I personally cannot find an openai-hosted clarification, just on hn and x.

Can anybody share a tested system prompt they use for general coding tasks in pi?

I have one core complex task where there are a number of simple errors like this. The easiest thing for me was to just have a post-processing script that performs: lint > mark known fail-early results > fix common errors (all formatting or type errors right now) > relint > final pass / fail > pass to next step (retry with tweaks / user in the loop since I don't have the token budget to just blindly keep trying the same thing).

I don't see raw token counts, just a list of steps and page counts. For example, what is the rough average token count per page in the ocr and in the translation steps for a Greek book?

I have seen Gemini costs change quite a bit when processing very similar books from the same series lately, mainly because thinking tokens have increased about 5x. Has that has happened to you as well?

Edit: for ocr I am using about 15k-25k tokens per page, but I have a complex prompt.

Curious as to what your budget was to get where you are today? That's a lot of tokens. I presume you are using gemini flash?

I ran into a website for work that would let you create a long password, but silently truncate it to 12 characters before saving. Mind boggling.

Sounds more like they are implementing mass surveillance and reporting whatever the US Gov wants for 'security reasons' back to them.

Cruise missiles are not general purpose tools, it's obviously not even remotely similar. Virtually everybody reading this could use Mythos immediately to do real work, collectively in virtually every part of the economy.

It's pretty problematic to not make it more widely available at least to US businesses, and there is not even a vetting process to get approved quickly and easily. If this is the new norm, the intended or unintended consequences of this type of gatekeeping will be an unprecedented consolidation of power amongst the largest corporations. Even more than we have seen over the last 20 years.

What I find impossible to judge is whether me choosing the harness that works best for me and the way I like to work will limit the quality of the LLM output.

In this case, given the complexity of LangChain I don't know if it would burn a lot of tokens or confuse the LLMs with context creep due to the large env and tooling compared to something much simpler, or would something much simpler burn a lot of tokens and stutter in execution compared to LangChain, because they have a lot of middleware optimizations that I would have to relearn the hard way? Or are both those strategies off the mark and there is a better tool for the job I am not even thinking about?

It gets a bit expensive in terms of both time/effort and money to experiment on a full workflow rather than specific steps which is quite easy in openrouter, hence my curiosity as to others' experiences.

I am very curious about this as well. I'm looking for something that does really well with workflows that require 20 plus steps including a couple while loops and user verifications, but also something simple like a chat bot with access to MCP servers and tools. I need to be able to use Gemini directly, openAI directly, and openrouter at various steps.

Right now I am having trouble deciding whether it's better to just write my own harness, or use langchain, or something else.

I think in most scenarios you don't need to worry so much about kvm ram use, since it looks static but actually it's not and you can over-commit [1]. And of course disk allocation can be dynamic as well. I prefer a lot more security for a bit less flexibility. I am not as ram rich as you are, and still every time I think of my few LXCs, my main thought is 'why did I do that?'.

[1] https://docs.redhat.com/en/documentation/red_hat_enterprise_...

Since models just output the the most probable tokens and you can never accuse them of doing anything other than making it all up, I would like to see these tests run with a prompt that attempts to mitigate hallucination and finishes with something like: "Telling me that you don't have the relevant information or that the task is impossible is extremely useful to me and a valid answer", and see how much that changes the scoring - as well as the usefulness of the answers. There are so many skills like context7 that can be tweaked to improve these results as well.

In other words, you shouldn't choose the model that hallucinates the least without detailed prompting, since a well-crafted agents.md clause should go a long way to improving output, and almost certainly the top scoring order will be different. To the point that I don't find this type of raw comparison useful beyond maybe 'make sure you test that one with more explicit prompts'.

The actual research paper:

https://www.sciencedirect.com/science/article/pii/S135041772...

I *literally* cannot read that yellow text on the white background. I even tried changing the brightness to almost 0, but there is just not enough contrast.

To check whether I understand how this all works: Wouldn't a 4 bit quant run reasonably well (for that hardware) with far less ram, something like 1.5x the 476gb, or 714gb+ ram?

This is the first time in terms of model progress where my personal response is: It does not matter to me because the models 6-12 months ago were already good enough for most everything I need to do. I think 95% of dev work is perfectly fine 6 months behind, if that is truly where we are at now with these open models.

As far as I can tell this type of model requires 640GB+ of memory using FP8. So likely can be run using 320GB+ memory if using FP4 or similar. So that would be 3 Nvidia DGX Sparks, or 12k of hardware. Is that correct? If so, it could make perfect sense for a small business.

If you wouldn't mind, could you explain a bit what the 248B model is good for, and where it breaks down and you need something better? I hear this take often, but it is always a fleeting remark so I have no idea what the 'useful' looks like - at all.

You can absolutely do that by using subprocess.run, or use the codex sdk

https://github.com/openai/codex/tree/main/sdk/python

I thought the most interesting part of the post was that they have an mcp endpoint for bring-your-own agents, and they won't be force feeding ai on anybody. In the security context of the post, they mean that you are responsible if your ai is duped into falling a victim, or tricked to send malicious mail.

My first daughter I managed to flip the switch for reading through Tintin and other graphic novels. My younger daughter skipped that entirely. She started reading later than the first, but jumped right in to longer full length books that were captivating for her (they were series she had seen her sister read).

I completely agree that we can encourage but reading needs to come naturally to them. You can't force-feed curiosity and passion, which is what reading is all about for young people.