Willison has never claimed to be an expert in the field of machine learning, but you should give more credence to his opinions.
Yeah, I would if he didn’t demonstrate such blatant misconceptions.
Willison is a prominent figure in the web-development scene
🤦 “They know how to sail a boat so they know how a car engine works”
Willison never claims or implies this in his article, you just kind of stuff those words in his mouth.
Reading comprehension. I never implied that he says anything about censorship. It is a correct and valid example that shows how his understanding of how system prompts work is wrong. “Define censorship” is not the argument you think it is lol. Okay though, I’ll define the “censorship” I’m talking about as refusal behavior that is introduced during RLHF and DPO alignment, and no the system prompt will not change this behavior.
I never implied that he says anything about censorship
You did, at least that’s what I gathered originally, you just edited your original comments quite extensively. Regardless,
Reading comprehension.
The provided example was clearly not intended to be taken as “define censorship,” and, again, it is ironic you accuse me of having poor reading comprehension while being incapable or unwilling to give a respectable degree of charitable interpretation to others. You kind of just take what you think is the easiest to argue against reading of others and argue against that instead of what anyone actually said, is a habit I’m noticing, but I digress.
Finally, not that it’s particularly relevant, but if you want to define censorship in this context that way, you’re more than welcome to, but it is a non-standard definition that I am not really sold on the efficacy of. I certainly won’t be using it going forwards.
Anyway, I don’t think we’re gonna get a lot of ground here. I just felt the need to clarify to anyone reading that Willison isn’t a nobody and give them the objective facts regarding his veracity, because again, as I said, claiming he is just some guy in this context is willfully ignorant at best.
if you want to define censorship in this context that way, you’re more than welcome to, but it is a non-standard definition that I am not really sold on the efficacy of. I certainly won’t be using it going forwards.
I just felt the need to clarify to anyone reading that Willison isn’t a nobody
I didn’t say he’s a nobody. What was that about a “respectable degree of chartiable interpretation of others”? Seems like you’re the one putting words in mouths, here.
If he was writing about django, I’d defer to his expertise.
From your own provided source on the arxiv, Noels et al. define censorship as:
Censorship in this context can be defined as the deliberate restriction, modification, or suppression of certain outputs generated by the model.
Which is starkly different from the definition you yourself gave. I actually like their definition a whole lot more. Your definition is problematic because it excludes a large set of behaviors we would colloquially be interested in when studying “censorship.”
Again, for the third time, that was not really the point either and I’m not interested in dancing around a technical scope defining censorship in this field, at least in this discourse right here and now. It is irrelevant to the topic at hand.
I didn’t say he’s a nobody. What was that about a “respectable degree of chartiable interpretation of others”? Seems like you’re the one putting words in mouths, here.
Yeah, this blogger shows a fundamental misunderstanding of how LLMs work or how system prompts work. (emphasis mine)
In the context of this field of work and study, you basically did call him a nobody, and the point being harped on again, again, and again to you is that this is a false assertion. I did interpret you charitably. Don’t blame me because you said something wrong.
EDIT: And frankly, you clearly don’t understand how the work Willison’s career has covered is intimately related to ML and AI research. I don’t mean it as a dig but you wouldn’t be drawing this arbitrary line to try and discredit him if you knew how the work done in Python on Django directly relates to many modern machine learning stacks.
Again, for the third time, that was not really the point either and I’m not interested in dancing around a technical scope defining censorship in this field, at least in this discourse right here and now. It is irrelevant to the topic at hand.
…
Either way, my point is that you are using wishy-washy, ambiguous, catch-all terms such as “censorship” that make your writings here not technically correct, either. What is censorship, in an informatics context? What does that mean? How can it be applied to sets of data? That’s not a concretely defined term if you’re wanting to take the discourse to the level that it seems you are, like it or not.
Yeah, I would if he didn’t demonstrate such blatant misconceptions.
🤦 “They know how to sail a boat so they know how a car engine works”
Reading comprehension. I never implied that he says anything about censorship. It is a correct and valid example that shows how his understanding of how system prompts work is wrong. “Define censorship” is not the argument you think it is lol. Okay though, I’ll define the “censorship” I’m talking about as refusal behavior that is introduced during RLHF and DPO alignment, and no the system prompt will not change this behavior.
You did, at least that’s what I gathered originally, you just edited your original comments quite extensively. Regardless,
The provided example was clearly not intended to be taken as “define censorship,” and, again, it is ironic you accuse me of having poor reading comprehension while being incapable or unwilling to give a respectable degree of charitable interpretation to others. You kind of just take what you think is the easiest to argue against reading of others and argue against that instead of what anyone actually said, is a habit I’m noticing, but I digress.
Finally, not that it’s particularly relevant, but if you want to define censorship in this context that way, you’re more than welcome to, but it is a non-standard definition that I am not really sold on the efficacy of. I certainly won’t be using it going forwards.
Anyway, I don’t think we’re gonna get a lot of ground here. I just felt the need to clarify to anyone reading that Willison isn’t a nobody and give them the objective facts regarding his veracity, because again, as I said, claiming he is just some guy in this context is willfully ignorant at best.
Lol you’ve got to be trolling.
I didn’t say he’s a nobody. What was that about a “respectable degree of chartiable interpretation of others”? Seems like you’re the one putting words in mouths, here.
If he was writing about django, I’d defer to his expertise.
Nope, not trolling at all.
From your own provided source on the arxiv, Noels et al. define censorship as:
Which is starkly different from the definition you yourself gave. I actually like their definition a whole lot more. Your definition is problematic because it excludes a large set of behaviors we would colloquially be interested in when studying “censorship.”
Again, for the third time, that was not really the point either and I’m not interested in dancing around a technical scope defining censorship in this field, at least in this discourse right here and now. It is irrelevant to the topic at hand.
In the context of this field of work and study, you basically did call him a nobody, and the point being harped on again, again, and again to you is that this is a false assertion. I did interpret you charitably. Don’t blame me because you said something wrong.
EDIT: And frankly, you clearly don’t understand how the work Willison’s career has covered is intimately related to ML and AI research. I don’t mean it as a dig but you wouldn’t be drawing this arbitrary line to try and discredit him if you knew how the work done in Python on Django directly relates to many modern machine learning stacks.
…
Lol this you?