sgc

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.

I was looking at their docs and Burr has agent cookbooks to get started with this, and it can handle multi-machine workflows. Is this not what you were looking for? I am not sure how it integrates and uses skills etc, but it seems like it should work to me.

https://burr.apache.org/docs/examples/agents/

Yes, I was not sufficiently thinking about copyright as an arbitrary legal construct that can be manipulated at will. I don't think output should have copyright, but I would presume the copyright should it ever exist would belong to the user and not the LLM creator, just like photoshop does not give adobe rights to user output. However much like there is no copyright, the uncertain output from an LLM should never directly create legal liability - the user prompt and intention should, and legal standards regarding recklessness and malice should apply. Otherwise it's a bit like blaming somebody for the output of a roulette wheel.

So I think I like the current decision which is more about presentation, dissemination, application, and claims than content, and there should of course be liability for LLM creators if they are not actively dealing with results like CP, violence, or many other illegal or dangerous things.

The ai legal situation is going to go through growing pains. I am abstracting from the specific laws of any one country, just thinking about the general context:

If ai output is not copyrightable, it should not be considered personal output. So nobody should be responsible for it. Or if it is considered personal output, it should be copyrightable. Or perhaps the ai companies will be liable for all output, and they will therefore all cease to exist in any useful form? This seems like another alternative, where the output legal value is not central, but there will be a thousand different fights about how it is presented to others.

That was a good video, and I also liked the Munro video that does a nice job of explaining how these work: https://www.youtube.com/watch?v=m507ryWhc6c

Not according to the complete comment:

     More like 10%, but my search has not been systematic. I am mostly looking where I know I will find image issues based on image filenames and “Find Similar Images” searches.

They are clearly saying they think this is likely above average.

In my opinion you are hand-waving away a difficult engineering problem and proposing a naive solution as if it would solve a problem that has already been partially solved, by rejecting all the work that has already been done on it. Don't dump on the surface, don't burn millions of tons of fuel a year to do it, study what has been done and improve on it instead.

I think your sense of scale is off. 90% of sea life is on the surface. 0.029% of ocean water is replenished from rainfall annually. Desalination concentrates are absolutely toxic to life. The current daily volume of brine discharge would require more than half the tankers in the world to be filled and discharged every single day. They would of course not last long with such a routine.

I thought they tend to pipe far out and discharge as far below the surface as possible, since there is a lot of surface life and it is less damaging this way.

Ships (with long submerged pipes) would be prone to weather events and generally less reliable than an installed pipe. Perforation would be prone to clogging from build up so a nonstarter I would expect. Adding flex tubing and a relocation robot would be a maintenance headache as well. Not sure there is an easy optimization.

So this is an alternative to using one coding agent with openrouter, changing the models between tasks? I am a neophite in these things, my ai use is more calling apis from scripts right now. Can somebody please explain the pros and cons (beyond to openrouter fees) of each?

It's an in-actionable "question" / comment. The rule does not claim one thing is better than the other. One is easily enforceable, the other is indemonstrable. If the point of this exchange is to better understand and use HN, the reason is because it is not hard to be constructive instead of throwing out non sequiturs.

And I didn't say it's '"often" the case someone asking the obvious question seemingly answered in the article had actually read it'. I said the person pointing it out while refusing to provide receipts or cordially engage is often wrong about what they think is obviously in the article. It's worthless noise regardless.

Just reply with a quote from the article. They will understand they did not read carefully, and you can avoid the low-value 'read the article' snark (that might be false since often it is not actually in the article when somebody does that).

I used to feel that way about prominent banners / cards. Then I tried to get donations on my own site and until I became hyper-aggressive I never received even a dollar. It was frankly disheartening. Now, it is not yet sustainable but at least moving in the right direction. In other words, they really have no choice.

How are you using it? More to complete specific functions or scripts, or for larger architectural design and longer implementation runs?

Agreed. There is actually a lot better control in openoffice / libreoffice than most people know. You just have to set up your styles and be systematic about (virtually) never using direct formatting, instead always applying a pre-configured style. There is a distinct value in seeing your final product as you work, when the final product is visual.

It's not like you aim to do it, you are just in a feedback loop improving results for the tool you are using. It is inherent in any prompt developed through iteration.

Claude has a /product-self-knowledge skill, and I am sure the others have something similar. So yes, it is possible if you work with care, as necessary with all things LLM related. There are hundreds if not thousands of skills on github that were created just this way.

Pretty much every single detailed prompt made after trial, error, and refinement is tailored to a specific LLM. They will all perform worse used with other LLMs than a similar prompt tailored for the second LLM would perform, and at times quite poorly.

They missed the shift key for the $ sign.

Hawaii can only make laws for Hawaii. So I don't even think it could stop a Hawaiian corporation from spending political dollars in another state. But it can probably stop corporations from spending money in Hawaii even if they are not incorporated there. This law will make its way to the supreme Court relatively quickly I expect.