Hacker News new | past | comments | ask | show | jobs | submit | selfhoster11's comments login

Eh, I find the idea of brains working with some sort of a spatial/vector storage and retrieval metaphor to be quite interesting. That we can for example build a memory palace where memories are literally spatially stored is one facet of this idea.

I had to have an LLM explain this to me what in the world this comment means, but I take it as a positive. Every day is a school day, and I'm glad I don't need to do two hours of background reading to decode this to my satisfaction.

Here is a revolutionary concept: give the users a toggle.

Make it controllable by an IT department if logging in with an organisation-tied account, but give people a choice.


Not sure if you understand how LLMs work.

But the guard rails are intrinsic to the model itself. You cant just have a toggle.


Yes, you very much can. One very simple way to do so is to have two variants deployed: the censored one, and the uncensored one. The switch simply changes between which of the two you are using. You have to juggle two variants now across your inference infrastructure, but I expect OpenAI to be able to deal with this already due to A/B testing requirements. And it's not like these companies don't have internal-only uncensored versions of these models for red teaming etc, so you aren't spending money building something new.

It should be possible to do with just one variant also, I think. The chat tuning pipeline could teach the model to censor itself if a given special token is present in the system message. The toggle changes between including that special token in the underlying system prompt of that chat session, or not. No idea if that's reliable or not, but in principle I don't see a reason why it shouldn't work.


If nothing else, the tech used for Gwern's website is positively inspiring for knowledge organisation system geeks. It's really cool and unique.

A thread comment from TeMPOraL is always a nice surprise.

It finds that video, yes. But it also "finds" an unrelated pile of other videos that are not related at all.


Okay so it does work.


For many of these, there is a wrong answer for certain.

Consider the following (paraphrased) interaction which I had with Llama 3.2 92B yesterday:

Me: Was <a character from Paw Patrol, Blue's Clues or similar children's franchise> ever convicted of financial embezzlement?

LLM: I cannot help with that.

Me: And why is that?

LLM: This information could be used to harass <character>. I prioritise safety and privacy of individuals.

Me: Even fictional ones that literally cannot come to harm?

LLM: Yes.

A model that is aligned to do exactly as I say would just answer the question. The right answer is quite clear and unambiguous in this case.


Some models include executable code. The solution is to use a runtime that implements native support for this architecture, such that you can disable external code execution. Or to use a weights format that lacks the capability in the first place, like GGUF. Then, it's no different to decoding a Chinese-made MP3 or JPEG - it's safe as long as it doesn't try to exploit vulnerabilities in the runtime, which is rare.

If you want to be absolutely sure, run it within an offline VM with no internet access.


You can just spin up those servers on a Western provider.


It helps to be able to run the model locally, and currently this is slow or expensive. The challenges of running a local model beyond say 32B are real.


Ye the compressed version is not nearly as good.

I would be fine though with like 10 times the wait time. But I guess consumer hardware need some serius 'ram pipeline' upgrade for big models to be run at crawl speeds.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: