this post was submitted on 16 Mar 2025
1022 points (97.7% liked)
Lemmy Shitpost
29875 readers
3320 users here now
Welcome to Lemmy Shitpost. Here you can shitpost to your hearts content.
Anything and everything goes. Memes, Jokes, Vents and Banter. Though we still have to comply with lemmy.world instance rules. So behave!
Rules:
1. Be Respectful
Refrain from using harmful language pertaining to a protected characteristic: e.g. race, gender, sexuality, disability or religion.
Refrain from being argumentative when responding or commenting to posts/replies. Personal attacks are not welcome here.
...
2. No Illegal Content
Content that violates the law. Any post/comment found to be in breach of common law will be removed and given to the authorities if required.
That means:
-No promoting violence/threats against any individuals
-No CSA content or Revenge Porn
-No sharing private/personal information (Doxxing)
...
3. No Spam
Posting the same post, no matter the intent is against the rules.
-If you have posted content, please refrain from re-posting said content within this community.
-Do not spam posts with intent to harass, annoy, bully, advertise, scam or harm this community.
-No posting Scams/Advertisements/Phishing Links/IP Grabbers
-No Bots, Bots will be banned from the community.
...
4. No Porn/Explicit
Content
-Do not post explicit content. Lemmy.World is not the instance for NSFW content.
-Do not post Gore or Shock Content.
...
5. No Enciting Harassment,
Brigading, Doxxing or Witch Hunts
-Do not Brigade other Communities
-No calls to action against other communities/users within Lemmy or outside of Lemmy.
-No Witch Hunts against users/communities.
-No content that harasses members within or outside of the community.
...
6. NSFW should be behind NSFW tags.
-Content that is NSFW should be behind NSFW tags.
-Content that might be distressing should be kept behind NSFW tags.
...
If you see content that is a breach of the rules, please flag and report the comment and a moderator will take action where they can.
Also check out:
Partnered Communities:
1.Memes
10.LinuxMemes (Linux themed memes)
Reach out to
All communities included on the sidebar are to be made in compliance with the instance rules. Striker
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Is this real? On account of how LLMs tokenize their input, this can actually be a pretty tricky task for them to accomplish. This is also the reason why it's hard for them to count the amount of 'R's in the word 'Strawberry'.
The LLM doesn't have to innately implement filtering. You can use a more traditional and concrete filtering strategy on top. So you sneak something problematic by in the prompt and it's too clever to be caught by the input filter, but then on the output the filter can catch that the prompt tricked the LLM into generating something undesired. Another comment specified they tried this and it started to work but then suddenly it seemingly shut out the reply in the middle, presumably the minute the LLM spit something at a more traditional filter and that shut it down.
I think I've seen this sort of approach has been applied to largely mask embarassing answers that become memes, or to detect input known not to work, and to shut it down or redirect it to a better facility (e.g. redirecting math to wolfram alpha).
It’s probably deepseek r1, which is a “reasoning” model so basically it has sub-models doing things like running computation while the “supervisor” part of the model “talks to them” and relays back the approach. Trying to imitate the way humans think. That being said, models are getting “agentic” meaning they have the ability to run software tools against what you send them, and while it’s obviously being super hyped up by all the tech bro accellerationists, it is likely where LLMs and the like are headed, for better or for worse.
Still, this does not quite address the issue of tokenization making it difficult for most models to accurately distinguish between the hexadecimals here.
Having the model write code to solve an issue and then ask it to execute it is an established technique to circumvent this issue, but all of the model interfaces I know of with this capability are very explicit about when they are making use of this tool.
Not really a concern. It's basically translation, which language models excel at. It just needs a mapping of the hex to byte
It is a concern.
Check out https://tiktokenizer.vercel.app/?model=deepseek-ai%2FDeepSeek-R1 and try entering some freeform hexadecimal data - you'll notice that it does not cleanly segment the hexadecimal numbers into individual tokens.
I'm well aware, but you don't need to necessarily see each character to translate to bytes
It's not out of the question that we get emergent behaviour where the model can connect non-optimally mapped tokens and still translate them correctly, yeah.
I’m confused, is the concern when the model doesn’t properly identify when it is using software to identify something like a hex pattern?
The concern is that the model doesn't actually see the world in terms of distinct hexadecimals, but instead as tokens of variable size - you can see this using the tiktokenizer-webapp: enter some text and it will split it into the series of tokens the model actually will process.
It's not impossible for the model to work it out anyway, but it is a reason for this type of task to be a bit harder on LLMs.
I understand how base models tokenize language. What I’m curious about you’re basing your response off a horrendously screenshotted meme image of someone interacting with deepseek. Is your concern that deepseek isn’t showing the code used to approach a hex string? Because that’s certainly a valid concern, though you can ask the model to output the code it is running. That’s definitely an ethics improvement that should be made in the UI, but it’s very clear what the model is doing under the hood
I don't think DeepSeek has the capability of generating code and executing it inline in the context window to support its answers, in the way that ChatGPT does - the "used"-part of that answer is likely a hallucination, while "or would use" more accurately represents reality.