nilsherzig

He even cites the reasons for that over engineering in the post you've linked?

Also, even before reading the comments, I was pleasantly surprised that everything on that page worked and nothing bothered me. There is a good amount of attention to detail (switching of the rounding on codeblocks when switching to light mode for example, avoiding a weird corner on mobile). Nothing moves around on load etc. I know this is a low bar, but it's gotten pretty rare, even for pages linked on here.

I'm trying to say that his 100k lines might actually be spent on attention to detail and aren't just slop.

Fyi, if you're trying to run this under AMD/HIP:

PTQ1_0 has no optimized MMQ-Path in their llama-cpp fork, try running PTQ2_0 (needs a bit more vram, but is about 2x faster on my 6700 XT)

https://gist.github.com/nilsherzig/b8266d001c5c01bdb3d81d209...

Please explain what you mean by "AI assisted usefull content is always fine".

Here is a detailed explanation from Pangram how their scoring works https://www.pangram.com/research/how-it-works

the website of this company is using more resources than their main product is (would really appreciate a button to turn of the animation)

yea fair, probably not

There you go :) https://hnslop.nilsherzig.com/

as an extension so you can continue to comment and use hn as usual

"Made with <3 by ..." Yea it is

added a firefox extension :)

Will do, thanks for hcker.news btw, has been my main mobile client for couple of months. Of all things i like the changelog most, its nice to have a quick view about new stuff, especially if the website changes somewhat faster than usual.

Yea very expensive (hence me siphoning another page haha).

Im trying to get something cheaper to work, Pangram has some nice docs on how to build something like their service https://github.com/pangramlabs/EditLens, they even have training data online.

The harm a missclassification carries is much lower for a hn post than a masters thesis, so we might be fine with a worse model.

nvm sorry, just read the "how this works" comment from OP. https://news.ycombinator.com/item?id=49660301

OPs solution does more than keyword filtering. But its still about filtering out "content about ai" not filtering out "ai written content".

If a Userscript solution is fine for you, see my recent comments :)

Here is an approach to filtering LLM written content: https://hnslop.nilsherzig.com/

Instead of filtering keywords this ranks the content using Pangram. But it's also pretty effective at filtering content about AI. Turns out that a lot of LLM tooling projects have LLM written project descriptions/blog posts.

It's amazing to see how many of the "AI hurts my brain" posts are apparently written by an LLM.

Not going to discuss this further, but wanted to apologize for my message. I was annoyed and should've waited before sending. Sorry.

Pretty decent output, mind sharing the prompt and model used? I struggle getting text i enjoy reading

https://www.pangram.com/history/5a4b90c6-0123-4f97-b50e-9c90...

If you really wanted to optimize something involving living consciousness beings like a machine, at least you should consider things like the need to take a piss as a hard constraint and design the system around it.

Don't get me wrong I love factorio like the next guy and understand the idea behind your thinking, but there should be limits to that thinking in the real world. And if that doesn't convince you consider the efficiency loss of a depressed worker with a urinary tract infection.

Honestly - and I don't mean that as an insult - I hope for you to experience (just once) the loss of dignity that not being able to visit the restroom on your own decision will give you. Might make you realize the hell this kind of thinking leads to.

Weirdly still so much more human (and somewhat relatable) than he ever was in an interview

Maybe... if run locally with full transparency into the prompts (including system etc) and independent benchmarks for the models (i.e. tests on how eager they are about selling me stuff)

But even in this hypothetical world I would still have a hard time teaching an LLM what I find important in a product. Not everything can be captured in easy to compare numbers. "Does the device creak under pressure?", "What's the one weirdly unergonomic thing about it you only notice after weeks of using?"... Stuff like that. Things which are easy to miss even if the model ingests all relevant reviews about the product, because it might not stand out, but might be important for me.

Nothing will ever beat this one https://cf.preview.redd.it/a-photo-of-an-astronaut-riding-a-...

Dogs shadow didn't change

i really enjoyed that one

Still flags as 100% LLM written https://www.pangram.com/history/cc9b6131-c772-4b3b-9a5d-8e81...

update: https://news.ycombinator.com/item?id=49582725 shared another page which already does the parsing and pangram requests. I wrote a little wrapper and userscript based extension around it (https://github.com/nilsherzig/hnslop), hope that helps!

created a little wrapper and userscript based browser extension https://hnslop.nilsherzig.com/

I have (an API, not an extension), but pangram is way to expensive for me to run, current (v4) pricing is $0.05 per 100 words. Should be absolutely doable if we split the cost between users tho.

I tried searching for good opensource / reasonably priced alternatives, pangram themselves even have some of their older architecture and training data on github/hugging face, but i never got that working reliably enough.

checks out:

> In 2006, SoftwareOnline was sued by The Washington State Attorney General's Office for alleged violations of the Consumer Protection Act after complaints were made about two products called "Registry Cleaner" and "InternetShield".

https://en.wikipedia.org/wiki/Dave_Plummer#:~:text=16%5D-,In...

Removing stuff without mentioning the removal.

Asking an LLM to remove an Idea from a Document often times results in an edit which explicitly states that this Idea is not relevant, instead of just removing all references to that Idea.

That might be useful for some form of "evolving" Documentation (so future readers know that this part of the search space was covered and deemed irrelevant) but is just overly verbose and confusing to read in most situations.

You can get better results from less aligned models like Kimi K3. Still not actually funny, but at least it’s able to produce some unhinged stuff and I guess shock and twists are kinda related to humor?

Still missing the human connection of cause, so im not sure if this is a technical / skill issue in the first place.

That's a Claude thing btw. Try using a harness which does not inject half a novel of instructions in combination with a different model. I would recommend Pi + GPT 5.6 Luna for a very capable and cheap test.

After using Claude (paid by work) for a couple of months, I was amazed how well instruction following works in other setups.

oh looks like they killed xcancel :/

https://xcancel.com/Eminem/status/2092367272891310482