“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
-
@Turre @A_denie @stux @Remittancegirl You're plain wrong, very naive and clearly clueless as to how LLMs are "trained" and here's why:
Any similarities with actual facts is to a very large degree _not_ related to being lucky.
"AI" companies feed a model a lot of data in a way that's not too unlike the lossy compression used to compress a track from a CD into a .mp3 file. An algorithm is then applied to create relations between the data points in the dataset ("tokens") and some basic weights are applied.
The next step is what "AI" companies call "training" which is in reality lots and lots of real people working minimum wage in poor countries who are tasked with asking questions in a script & then they evaluate the output according to a manual/instructions written by a handful of people in Silicon Valley.
You may want to look further into this and really think it through if you still don't get that how the weights are adjusted is, in fact, largely _not_ a matter of "luck" or chance. If the "AI" companies want their LLM to present a narrative when asked about a topic then they absolutely can and will make sure the model presents that narrative every time. Consider the implications as more and more people use LLMs like some kind of all-knowing ministry of truth.
@chaekyung I honestly don't think it is helpful to start a dialogue with the statement "You're plain wrong, very naive and clearly clueless"
Clearly you know a lot about this and your participation is welcome, but the ad hominem is not.
If you want more people to understand the way LLMs are trained, then that's great. Try not to shame them before you educate them.
-
@Turre @A_denie @stux @Remittancegirl You're plain wrong, very naive and clearly clueless as to how LLMs are "trained" and here's why:
Any similarities with actual facts is to a very large degree _not_ related to being lucky.
"AI" companies feed a model a lot of data in a way that's not too unlike the lossy compression used to compress a track from a CD into a .mp3 file. An algorithm is then applied to create relations between the data points in the dataset ("tokens") and some basic weights are applied.
The next step is what "AI" companies call "training" which is in reality lots and lots of real people working minimum wage in poor countries who are tasked with asking questions in a script & then they evaluate the output according to a manual/instructions written by a handful of people in Silicon Valley.
You may want to look further into this and really think it through if you still don't get that how the weights are adjusted is, in fact, largely _not_ a matter of "luck" or chance. If the "AI" companies want their LLM to present a narrative when asked about a topic then they absolutely can and will make sure the model presents that narrative every time. Consider the implications as more and more people use LLMs like some kind of all-knowing ministry of truth.
@chaekyung
Charming. I'm aware, you know.Was merely pointing out that ChatGPT doesn't provide facts, that's not how these things operate.
-
@chaekyung I honestly don't think it is helpful to start a dialogue with the statement "You're plain wrong, very naive and clearly clueless"
Clearly you know a lot about this and your participation is welcome, but the ad hominem is not.
If you want more people to understand the way LLMs are trained, then that's great. Try not to shame them before you educate them.
@Remittancegirl @Turre @A_denie @stux "Any similarities with actual facts is being lucky" is plain wrong & naive. Similarly, if you claim 2+2=5 then I'm going to tell you that you can't do math on a child's level.
Feel free to mute me if you can't deal with people stating the obvious.
-
@Remittancegirl @stux Or how little they know about the education system and it's purpose.
@alpacamale @Remittancegirl @stux "But teacher, when are we ever going to use this stuff in real life?"
-
@Remittancegirl @Turre @A_denie @stux "Any similarities with actual facts is being lucky" is plain wrong & naive. Similarly, if you claim 2+2=5 then I'm going to tell you that you can't do math on a child's level.
Feel free to mute me if you can't deal with people stating the obvious.
It may be plain wrong.
But your appetite and enjoyment of shaming another person is ugly.
I do not wish to see it.
-
“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
Yeah worst of all, gym classes are making them run a mile that Uber can cover in less than a minute@stux related
-
@Turre @A_denie @stux @Remittancegirl You're plain wrong, very naive and clearly clueless as to how LLMs are "trained" and here's why:
Any similarities with actual facts is to a very large degree _not_ related to being lucky.
"AI" companies feed a model a lot of data in a way that's not too unlike the lossy compression used to compress a track from a CD into a .mp3 file. An algorithm is then applied to create relations between the data points in the dataset ("tokens") and some basic weights are applied.
The next step is what "AI" companies call "training" which is in reality lots and lots of real people working minimum wage in poor countries who are tasked with asking questions in a script & then they evaluate the output according to a manual/instructions written by a handful of people in Silicon Valley.
You may want to look further into this and really think it through if you still don't get that how the weights are adjusted is, in fact, largely _not_ a matter of "luck" or chance. If the "AI" companies want their LLM to present a narrative when asked about a topic then they absolutely can and will make sure the model presents that narrative every time. Consider the implications as more and more people use LLMs like some kind of all-knowing ministry of truth.
@stux @Remittancegirl @chaekyung @Turre @A_denie oh delusionary bullshit rants get well deserved blocks
-
“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
Yeah worst of all, gym classes are making them run a mile that Uber can cover in less than a minute@stux astonishing how many people have apparently long derided learning and thinking, and were wholly relieved when a personal 'thinking machine' that knows everything finally arrived.
-
@Turre @A_denie @stux @Remittancegirl You're plain wrong, very naive and clearly clueless as to how LLMs are "trained" and here's why:
Any similarities with actual facts is to a very large degree _not_ related to being lucky.
"AI" companies feed a model a lot of data in a way that's not too unlike the lossy compression used to compress a track from a CD into a .mp3 file. An algorithm is then applied to create relations between the data points in the dataset ("tokens") and some basic weights are applied.
The next step is what "AI" companies call "training" which is in reality lots and lots of real people working minimum wage in poor countries who are tasked with asking questions in a script & then they evaluate the output according to a manual/instructions written by a handful of people in Silicon Valley.
You may want to look further into this and really think it through if you still don't get that how the weights are adjusted is, in fact, largely _not_ a matter of "luck" or chance. If the "AI" companies want their LLM to present a narrative when asked about a topic then they absolutely can and will make sure the model presents that narrative every time. Consider the implications as more and more people use LLMs like some kind of all-knowing ministry of truth.
@chaekyung @Turre @A_denie @stux @Remittancegirl LLMs are not trained using facts, they are trained using text, and text is not facts. LLMs can infer facts about how text usually works, but not facts about the world.
-
“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
Yeah worst of all, gym classes are making them run a mile that Uber can cover in less than a minute@stux Aside from multiplication tables and maybe poetry and drama, I don't think US schools have actually required rote memorization in generations, aside from a few reactionary teachers trying to make a stupid point.
Even in the 80s, we were being taught how to use Boolean logic in digital catalog searches and how to use scientific calculators.
-
@Remittancegirl @Turre @A_denie @stux "Any similarities with actual facts is being lucky" is plain wrong & naive. Similarly, if you claim 2+2=5 then I'm going to tell you that you can't do math on a child's level.
Feel free to mute me if you can't deal with people stating the obvious.
@chaekyung @Remittancegirl @Turre @A_denie @stux
We are supposed to be friends here. If you can't be kind you should go back to twitter.
-
“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
Yeah worst of all, gym classes are making them run a mile that Uber can cover in less than a minute@stux @briankrebs ChatGPT does NOT retrieve facts. It generates “fact shaped responses”
-
@econads @landelare @stux This is exactly the same reason the same people who were yelling at kids "don't trust what you read on the internet" are now trusting every piece of propaganda and slop in their Facebook feeds.
Then, "on the internet" meant queer anarchist hackers. Now it means "right wing disinformation machines".
@dalias
I hear what you're saying and there was definitely an element of gatekeeping. It can be both though, especially for physics, which, not *widely* understood. I tend to trust in crowdsourcing myself but I'm not sure most of my professors would have been sharing conspiracy theories on Facebook.. any large enough group is varied enough to have different motivations. -
@Remittancegirl @stux I'm sick and tired of apologism for torturous "education".
Especially the mass delusion of ill-defined nebulous positive side effects of torturous tedious memorization somehow making it the best way.
Those effects can be brought by any subject matter, there is no need for it to be bullshit that "doesn't matter" and sucks the soul out of every participant.@ariarhythmic @Remittancegirl @stux
So you're calling to end all learning?Or is there nuance to your argument like what you're missing in the rest of the thread?
-
@Remittancegirl @stux Yes, exactly this! Holding two opposing thoughts generally creates cognitive dissonance, so it should be incredibly important to help kids early to understand what that feeling is and how to think critically through it in order to hold those thoughts and then respond instead of react negatively. I like this idea a lot!
@ManWhoTalksToHisHand THis. so much! @stux
-
@ariarhythmic @Remittancegirl @stux
So you're calling to end all learning?Or is there nuance to your argument like what you're missing in the rest of the thread?
@cshlan I cannot see the original post because the poster has blocked me.
I think they equated the taught skill of memorisation with all rote learning. Maybe they had a very bad experience with it. -
@blotosmetek @stux This. Having been a lecturer at a university, I can say without question - it is not what information is retained that matters, but that it is retained, so it can be applied critically to decision-making.
@Remittancegirl @blotosmetek @stux What happens when the grid goes down and/or loss of internet access?
A lot of helpless, hopelessly stupid people who can't function.I have purposefully learned manual, low tech ways of making things:
Handsewing, non-powered woodworking, knitting, spinning, weaving, and handwriting notes*. I don't need a calculator because I can still do math on paper (adding long lists of numbers does get checked with a calculator).
*Due to the way my brain works, I can write significantly faster than I can type.
So, friends, while the adle-brained sit in the dark, I shall continue on.
-
“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
Yeah worst of all, gym classes are making them run a mile that Uber can cover in less than a minute@stux This is EXACTLY the problem with so-called AI getting anywhere near the education system.
Learning and knowledge are not just the accumulation of the tokens we use to represent facts. They are a whole-body *practice* that is integral to an individual functioning as part of the human collective. And if we don't *practice* that, we have nothing, exactly the same as not moving and feeding and using our bodies ends up with us ... not being able to move and use our bodies so well at all.
-
“We’re sending kids to school to memorize facts that ChatGPT can retrieve in under a second.”
Yeah worst of all, gym classes are making them run a mile that Uber can cover in less than a minute@stux Actually it can't retrieve those facts in under a second. Oftentimes you get "Thinking..." or something like it for a while, and then a response. And the response may be wrong.
-
@Turre @A_denie @stux @Remittancegirl You're plain wrong, very naive and clearly clueless as to how LLMs are "trained" and here's why:
Any similarities with actual facts is to a very large degree _not_ related to being lucky.
"AI" companies feed a model a lot of data in a way that's not too unlike the lossy compression used to compress a track from a CD into a .mp3 file. An algorithm is then applied to create relations between the data points in the dataset ("tokens") and some basic weights are applied.
The next step is what "AI" companies call "training" which is in reality lots and lots of real people working minimum wage in poor countries who are tasked with asking questions in a script & then they evaluate the output according to a manual/instructions written by a handful of people in Silicon Valley.
You may want to look further into this and really think it through if you still don't get that how the weights are adjusted is, in fact, largely _not_ a matter of "luck" or chance. If the "AI" companies want their LLM to present a narrative when asked about a topic then they absolutely can and will make sure the model presents that narrative every time. Consider the implications as more and more people use LLMs like some kind of all-knowing ministry of truth.
@chaekyung @Turre @A_denie @stux @Remittancegirl
Fascinating…
The key point is still that factuality of LLM output is correlational, not causational.
