Deepseek is like if we banned RAM sticks of more than 2 GB to China and they built a version of Chrome that runs on 256 MB on a Raspberry PI. Simply wild innovation.
Comments
Log in with your Bluesky account to leave a comment
That's my understanding as well. Currently downloading one of the smaller models to run on LM Studio locally.
LLM models are reasonably easy to "jailbreak" if you have enough compute, is my understanding. So it wouldn't be terribly difficult to uncensor with enough compute power. No idea how much.
Sometimes you don't really need much compute to "jailbreak" an LLM. I guess you'd need to explain what you mean by that. You want to fully uncensor the LLM? That's not quite the same thing in my book.
"Uncensor" is probably a more accurate term. What I mean is updating the weights of the model to permanently change it.
In my mind it's all circumventing guardrails put in place, but yes you are correct that prompt injection and other techniques can get around guardrails without modifying the model
Comments
But it will be killed by Nationalism! #MAGA
What good is open source nature of it if the data is still headed overseas to be ingested. Can that be changed by developers?
So the model is OSS but the Deepseek service people use is not from how I understand it
LLM models are reasonably easy to "jailbreak" if you have enough compute, is my understanding. So it wouldn't be terribly difficult to uncensor with enough compute power. No idea how much.
In my mind it's all circumventing guardrails put in place, but yes you are correct that prompt injection and other techniques can get around guardrails without modifying the model
Its not bad at all. Certainly as good as ChatGPT
Now the real test is if it can write 10 sentences ending in the word Apple 🤣. ChatGPt 4o struggles.
Maybe this is again a kind of Linux moment.
One lesson here is that if it takes more than 2 pizza's to feed your team, the project will be bloated and inefficient. :)