Sitting down for a lunch interview is definitely an interesting experience!
-
Sitting down for a lunch interview is definitely an interesting experience! Leaving the roughly two-hour conversation with George Hammond, I felt like we'd covered good ground but had no idea what of that would make it into the piece. In the end, I think this covers it pretty well, though I (of course) have a couple of quibbles/some context to add.
https://www.ft.com/content/9029cc1c-4a3f-42ca-9939-f3ef8e8336ae
>>
-
Sitting down for a lunch interview is definitely an interesting experience! Leaving the roughly two-hour conversation with George Hammond, I felt like we'd covered good ground but had no idea what of that would make it into the piece. In the end, I think this covers it pretty well, though I (of course) have a couple of quibbles/some context to add.
https://www.ft.com/content/9029cc1c-4a3f-42ca-9939-f3ef8e8336ae
>>
My remark that LLMs were "born shitty" was in the context of a discussion of Cory Doctorow's notion of "enshittification". That process involves something that starts of beneficial for consumers (or at least the consumers in focus). LLMs used as synthetic text extruding machines have no legitimate use cases and --- for all the reasons discussed in the stochastic parrots paper --- are prone to harmful outputs to boot.
>>
-
My remark that LLMs were "born shitty" was in the context of a discussion of Cory Doctorow's notion of "enshittification". That process involves something that starts of beneficial for consumers (or at least the consumers in focus). LLMs used as synthetic text extruding machines have no legitimate use cases and --- for all the reasons discussed in the stochastic parrots paper --- are prone to harmful outputs to boot.
>>
@emilymbender Hi professor Bender, thank you for your continued work on this, it is extremely important! I have a terminology question in that I don’t understand how to parse the term “synthetic text extruding machine” (and this may be because I am an L2 speaker of English):
I get “synthetic text” and I get “extrude” (I think - dictionaries say ‘force out’ so I’m thinking something like ‘forcibly extract’?). Put together I read it as ‘machines that forcibly extract synthetic text’. That would, in my ears, imply that the text within the machine is already synthetic. But the texts that the models are trained on a largely human-made, so not synthetic. What am I not getting here?
Thanks again! -
@emilymbender Hi professor Bender, thank you for your continued work on this, it is extremely important! I have a terminology question in that I don’t understand how to parse the term “synthetic text extruding machine” (and this may be because I am an L2 speaker of English):
I get “synthetic text” and I get “extrude” (I think - dictionaries say ‘force out’ so I’m thinking something like ‘forcibly extract’?). Put together I read it as ‘machines that forcibly extract synthetic text’. That would, in my ears, imply that the text within the machine is already synthetic. But the texts that the models are trained on a largely human-made, so not synthetic. What am I not getting here?
Thanks again!@tanyakaroli @emilymbender "extruding" is what happens when frosting leaves a bakers piping bag... Or what happens when hay leaves a horse. Extruding is a bit of a polite euphemism in this context.
-
@emilymbender Hi professor Bender, thank you for your continued work on this, it is extremely important! I have a terminology question in that I don’t understand how to parse the term “synthetic text extruding machine” (and this may be because I am an L2 speaker of English):
I get “synthetic text” and I get “extrude” (I think - dictionaries say ‘force out’ so I’m thinking something like ‘forcibly extract’?). Put together I read it as ‘machines that forcibly extract synthetic text’. That would, in my ears, imply that the text within the machine is already synthetic. But the texts that the models are trained on a largely human-made, so not synthetic. What am I not getting here?
Thanks again!Synthetic Vs Natural (this all comes from 'natural' language processing) refers to the mode of production, not an intrinsic, pre-production characteristic, as text only exists after its production.
"Extruder" here is an image more than a technical descriptor, evoking industrial, mass production, indeed a specific process or pushing stuff though a nozzle to make long... things (think pasta machine).
-
Synthetic Vs Natural (this all comes from 'natural' language processing) refers to the mode of production, not an intrinsic, pre-production characteristic, as text only exists after its production.
"Extruder" here is an image more than a technical descriptor, evoking industrial, mass production, indeed a specific process or pushing stuff though a nozzle to make long... things (think pasta machine).
The shape of the object extruded is based solely on the shape of the nozzle and the amount extruded.
LLM produced synthetic texts' "shape" is the standardised reproduction of the statistical patterns in the training set.
That's the exegesis of the term, for me.
-
The shape of the object extruded is based solely on the shape of the nozzle and the amount extruded.
LLM produced synthetic texts' "shape" is the standardised reproduction of the statistical patterns in the training set.
That's the exegesis of the term, for me.
@MrBerard thanks, that’s helpful! (I would still say that there are texts existing prior to LLM production of synthetic texts, namely the texts in the training data).
-
@tanyakaroli @emilymbender "extruding" is what happens when frosting leaves a bakers piping bag... Or what happens when hay leaves a horse. Extruding is a bit of a polite euphemism in this context.
@OvertonDoors haha, thanks, got it!
@emilymbender
