I am asking people once again to stop describing #AI using human words and attributes.
-
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
@drahardja
I would properly use the output verb, which is expected for a computer program. Minsky did great damage to the whole research sector when he invented the AI allocution in his thesis. -
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
@drahardja
Also, using "AI" for LLMs is misleading. There's no I. -
@drahardja @talexb a client wants me to use the Claude tool, and one thing I am training myself to do is be consistent in my use of “the Claude tool” in any communication including commits. It is not a co-worker in any more meaningful sense than a magic 8 ball I decided to call Ian.
@MikeStok @drahardja I keep coming back to the same there questions.
1. How do you know that the Claude code works?
2. How do you know it's safe?
3. What's the provenance of the code? -
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
ChatGPT, Claude, none of them "know" about the question you're asking or "know" the answer.
Photoshop doesn't "know" what's in the photo you're editing or the digital art you're creating by clicking a portion of the screen to select a tool, then clicking another portion to use it.
They are computer programs. They run instructions to process your input against the data the program is using and deliver output.
You type. You click. It does special math. You see things on-screen.
(@drahardja )
-
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
@drahardja I think regurgitate is still too biological.
I prefer the sentence "AI interpolates swollen data. An interpolation function sometimes matches reality, sometimes not."
-
ChatGPT, Claude, none of them "know" about the question you're asking or "know" the answer.
Photoshop doesn't "know" what's in the photo you're editing or the digital art you're creating by clicking a portion of the screen to select a tool, then clicking another portion to use it.
They are computer programs. They run instructions to process your input against the data the program is using and deliver output.
You type. You click. It does special math. You see things on-screen.
(@drahardja )
@drahardja Ha, actually, that may be a decent start to some new terms folks could try to popularize.
"I needed a cover letter, so I got one of the bots to crank out some special math."
"Ugh, I didn't want to code that bit, so I went to Claude to auto-generate it for me. Unfortunately, it gave me some 'special' math—the REALLY 'special' kind. Wasted twenty minutes sorting through the garbage it produced before just fucking doing it myself anyway."
-
@robin_wren @drahardja At least the plant is a living organism.
@jairajdevadiga @drahardja I think you know what I meant, though.
-
@drahardja seems very much like what actual people do:
People don't answer-they regurgitate plausible output that the educational system taught them was right.
I suppose that depends all on the quality of the person. A very robotic person, who believes everything they are taught is actually very close to the same quality of output we might expect from the AI.
So, if those are the people who are eliminated, maybe it's okay. They are just mediocre anyway, claiming common knowledge as expertise.
@dckim @drahardja
This reply shows that the next paragraph after dehumanization is elimination. -
@MikeStok @drahardja I keep coming back to the same there questions.
1. How do you know that the Claude code works?
2. How do you know it's safe?
3. What's the provenance of the code?In this case I have, as Baldrick might say, a cunning plan.
I have “inherited“ a Rails app with no tests which reflects the state of Rails 10+ years ago.
So I let the Claude tool extrude tests for all the endpoints (lots of reasonable examples in the training data, I hope) and make them pass.
As I don’t allow the Claude tool to do git commits, and I have a copy of the code inaccessible to the host the tool is running on for good measure, I can see the less than 20 lines of the core application the tool has interfered with and review them. My hunch is they are signals of bigger changes I need to make, and I hope I can quickly expunge the tool’s code from the non-test codebase quickly.
I wish I could say the tool’s code was worse than the application code, but that’s a beer chat!
-
@dckim @drahardja
This reply shows that the next paragraph after dehumanization is elimination.@dontreportme @drahardja good point there. My proposed equivalency is only a brutal dehumanization. Well said.
-
@drahardja I think regurgitate is still too biological.
I prefer the sentence "AI interpolates swollen data. An interpolation function sometimes matches reality, sometimes not."
@plantagolabs I also like to use the word “extrude” instead of regurgitate.
-
@drahardja Regurgitate is an animal bodily function. AI doesn’t have a body. I think what I’m getting at is that there aren’t many appropriate words to describe how it works.
@david I also like the word “extrude”
-
@drahardja While I agree with the premise, I do want to point out that we, humans, have a tendency to anthropomorphize anything.
I'll apologize to a potted plant if I knock it off the window sill...@robin_wren @drahardja
Yeah, but don’t anthropomorphize computer programs. They hate that! -
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
But it helps me to do one day of my work in one hour!
-
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
@drahardja That is the correct way to describe GenAI: nothing more than context and probability
-
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
@drahardja Not to disagree with your point, but it's interesting to me how anthropomorphizing cognitive functions (answer, hallucinate, etc.) trips people up but anthropomorphizing biological ones (like regurgitate) doesn't.
-
@drahardja Not to disagree with your point, but it's interesting to me how anthropomorphizing cognitive functions (answer, hallucinate, etc.) trips people up but anthropomorphizing biological ones (like regurgitate) doesn't.
@Thad I think “regurgitate” is not something humans regularly do (because we’re not ruminants!), but the more I think about it maybe the word “extrude” would be better.
-
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
The most resonate description for me is by Dr Emily Bender who compares it to a Magic Eight Ball, a 64 bit Magic Eight Ball.
I remain shocked at how easily, relatively "smart" people embrace AI as a cure-all. How CEOs become cheerleaders to subordinates of an unproven robot to "extrude" managerial "decisions" without question.
Crazy.
-
@drahardja
Also, using "AI" for LLMs is misleading. There's no I.@lehtimaeki @drahardja There's no A either.
-
I am asking people once again to stop describing #AI using human words and attributes.
AI doesn’t answer—it regurgitates plausible output that happens to be right.
AI doesn’t hallucinate—it regurgitates plausible output that happens to be wrong.
AI doesn’t have agency—it regurgitates plausible output that is consumed by a computer program.
AI doesn’t think—it regurgitates plausible output that is fed back into itself to try again.
AI doesn’t create—it regurgitates a mixture of its training data.Remember: the robot is not a person.
@drahardja The worst thing I have heard was a statement "you know who can dothis very well? Copilot!"
