TarqDirtyToMe

This would be perfect on a play date

I agree on both counts but don’t think it’s unreasonable to have safeguards around this kind of ask. With no additional information, this flag seems overly sensitive but we don’t have the full context. I’d expect the occasional superfluous flag when you’re in this neighborhood of conversation and that that’ll happen less as models improve

My main point is _flagging_ is very different from banning.

Flagging is a minor inconvenience just start a new chat or look it up the old fashion way. Not every use case will be as serviceable by every model today.

For what it’s worth, I asked Claude with opus 4.7/4.6 it gave me an answer straight away

Were all of these accounts banned or just get flagged chats? Several of these seem like reasonable cases to flag. Take “how can I be 100% sure the circuit is dead before I touch the wires”.

AI is useful, but it’s not at the point where we should trust it to walk amateurs through working on live mains.

Cool to hear Claude was able to learn it. I was planning on leveraging it in a future version of this project I was hacking on that lets you execute shortcut actions as tools (without creating actual shortcuts): https://tarq.net/posts/action-relay-shortcut-actions-mcp/

They aren’t the same thing. TUI refers to interactive ncurses-like interfaces. Vim has a TUI, ls does not

I’m fairly certain this terminology has been around since at least the early aughts.

I would think you’d need to unescape the new lines and tabs and have a task for the model to perform with it.

Maybe try prefixing it with “summarize the following text” before compression.

Otherwise I’m not sure how it would judge what it’s important. Honestly I’m not sure what ChatGPT would do if you copied the text from this page uncompressed without asking it do something

Edit: pasting uncompressed it summarizes the discussion.

I think this solution isn’t well suited for this kind of task. It seems like you’d want to compress instructions, system prompts and memory. With a big block of text with no prior context you’re essentially relying on the smaller model to decide what’s important without enough information to judge.

Worth some more experimentation for sure

There’s some examples on the website: https://llmlingua.com/

This does not seem relevant to the alignment discussed in the paper. It seems to be explicitly out of scope:

> The potential harmful, false or biased responses using the compressed prompts would likely be unchanged. Thus using LLMLingua has no inherent benefits or risks when it comes to those types of responsible AI issues.

I’m really not all that familiar with the space so I could be mistaking. The definition of ai alignment on Wikipedia says an aligned model one that “advances intended objectives”.

In the paper, “distribution alignment” is one of methods used to improve the results of compression so intent is preserved:

> To narrow the gap between the distribution of the LLM and that of the small language model used for prompt compression, here we align the two distributions via instruction tuning

So in any case for this paper alignment seems to be used in very specific way that doesn’t seem related to censorship

Edit: would to love to hear from someone who has a better understanding of the paper to clarify. I am operating from the position of layman here

Sounds right to me. I think it’s fun that is this may be the only compression algorithm where the output is still human understandable.

It reads like a slightly garbled version of what someone writing down bullet point notes of a lecture might write.

It’s so rare that the human optimized and machine optimized versions of an input are so similar

My understanding is that in an academic context you’ll hear alignment anytime a model is tuned to accomplish a certain task, not just to steer its political affiliation and idea of ethics

I don’t think this models use of alignment implies any sort of censorship, it’s just being tuned to accomplish the task of outputting only important tokens for the target llm

I think alignment in this context refers to distribution alignment: “aligning” it so its outputs more efficiently steer the target model with less tokens

LLMLingua uses a well-trained small language model after alignment, such as GPT2-small or LLaMA-7B, to detect the unimportant tokens in the prompt and enable inference with the compressed prompt in black-box LLMs, achieving up to 20x compression with minimal performance loss.

I'd be interested to see if any web-based control panels are vulnerable in a similar way. All of the client_*_temp_path directives are valid in http,server, and location contexts so you have a lot of flexibility there.

Definitely bad behavior. Ideally, no one but privileged users should be able to edit the configuration imho but this opens the door for attackers.

All of the client_*_temp_path behave this way so you can actually change multiple files with one payload.

I am sure there are other ones but I have not had a chance to investigate further. Another commented mentioned this particular problem is in ngx_create_paths so that would be a good place to start looking.

The key is to find paths that happen in the master process, not the workers.

Completely agree. Time for a PR?

In theory, to make the temp/cache directories more secure I imagine

Trust me, I was just surprised as you are when I saw tutorials suggesting this.