What is OpenAI?
December 30, 2024 11:59 AM Subscribe
Contemplating the future of AI in general and OpenAI in particular isn’t just a Silicon Valley parlor game but national watercooler chatter. OpenAI is the next Google, some cry — it’ll change the way we use the internet! Others warn it’s the next FTX, an over-hyped scam led by a charismatic founder that’ll go to zero ... Some see huge returns, others serious peril. In some cases, the tech is transformative on a scale that today we have difficulty even imagining; in others, you’ll hear about a rather boring but extremely lucrative business. And then there’s Taco Bell. from What company’s past reveals the future of OpenAI? [Sherwood]
OpenAI is Visa by Taylor Lorenz
OpenAI is bitcoin by Crystal Kim
OpenAI is Uber by Max Knoblauch
OpenAI is Lyft by J. Edward Moreno
OpenAI is Adobe by Sophia Smith Galer
OpenAI is Pixar by Jon Keegan
OpenAI is Intel by Jeff Wise
OpenAI is IBM by Jack Raines
OpenAI is Microsoft by Ryan Broderick & Adam Bumas
OpenAI is Salesforce by Rani Molla
OpenAI is Napster by Jamie Wilde
OpenAI is Taco Bell by Adam Chandler
OpenAI is Lehman Bros. by Ed Zitron
OpenAI is Visa by Taylor Lorenz
OpenAI is bitcoin by Crystal Kim
OpenAI is Uber by Max Knoblauch
OpenAI is Lyft by J. Edward Moreno
OpenAI is Adobe by Sophia Smith Galer
OpenAI is Pixar by Jon Keegan
OpenAI is Intel by Jeff Wise
OpenAI is IBM by Jack Raines
OpenAI is Microsoft by Ryan Broderick & Adam Bumas
OpenAI is Salesforce by Rani Molla
OpenAI is Napster by Jamie Wilde
OpenAI is Taco Bell by Adam Chandler
OpenAI is Lehman Bros. by Ed Zitron
Burn it all down, every bit of it.
posted by outgrown_hobnail at 12:09 PM on December 30 [21 favorites]
posted by outgrown_hobnail at 12:09 PM on December 30 [21 favorites]
The Taco Bell analogy of OpenAI fits with this old comment about the nature of a Crunchwrap.
posted by Ickster at 12:26 PM on December 30 [9 favorites]
posted by Ickster at 12:26 PM on December 30 [9 favorites]
(That's at least the second time I've linked to that comment; it's one of the funniest things I've ever read, so it won't be the last time I link it I'm sure.)
posted by Ickster at 12:27 PM on December 30 [3 favorites]
posted by Ickster at 12:27 PM on December 30 [3 favorites]
Today I asked o1 how often are there 5 fridays in a February. It got the answer right
This is not your father’s”Eliza”
posted by torokunai2 at 12:31 PM on December 30 [5 favorites]
This is not your father’s”Eliza”
posted by torokunai2 at 12:31 PM on December 30 [5 favorites]
Others warn it’s the next FTX
OpenAI has 300 million weekly users. That's a lot of people finding it useful. Maybe they're too over-leveraged and will never figure out how to be profitable (the "Lehman Bros" scenario) but it won't be because they don't have a good product.
Meanwhile: OpenAI o3 Breakthrough High Score on ARC-AGI-Pub. Translation: OpenAI's next model shows a massive leap in reasoning.
posted by gwint at 12:46 PM on December 30 [5 favorites]
OpenAI has 300 million weekly users. That's a lot of people finding it useful. Maybe they're too over-leveraged and will never figure out how to be profitable (the "Lehman Bros" scenario) but it won't be because they don't have a good product.
Meanwhile: OpenAI o3 Breakthrough High Score on ARC-AGI-Pub. Translation: OpenAI's next model shows a massive leap in reasoning.
posted by gwint at 12:46 PM on December 30 [5 favorites]
Some would claim OpenAI is a group of liars and thieves who want to put a moat around their profit model, for the Greater Good of course.
posted by Enturbulated at 12:48 PM on December 30 [14 favorites]
posted by Enturbulated at 12:48 PM on December 30 [14 favorites]
OpenAI o3 Breakthrough High Score on ARC-AGI-Pub
Here's the quiet part out loud:
Of course, such generality comes at a steep cost, and wouldn't quite be economical yet: you could pay a human to solve ARC-AGI tasks for roughly $5 per task (we know, we did that), while consuming mere cents in energy. Meanwhile o3 requires $17-20 per task in the low-compute mode. But cost-performance will likely improve quite dramatically over the next few months and years, so you should plan for these capabilities to become competitive with human work within a fairly short timeline.
posted by chavenet at 12:55 PM on December 30 [15 favorites]
Here's the quiet part out loud:
Of course, such generality comes at a steep cost, and wouldn't quite be economical yet: you could pay a human to solve ARC-AGI tasks for roughly $5 per task (we know, we did that), while consuming mere cents in energy. Meanwhile o3 requires $17-20 per task in the low-compute mode. But cost-performance will likely improve quite dramatically over the next few months and years, so you should plan for these capabilities to become competitive with human work within a fairly short timeline.
posted by chavenet at 12:55 PM on December 30 [15 favorites]
Maybe they're too over-leveraged and will never figure out how to be profitable (the "Lehman Bros" scenario) but it won't be because they don't have a good product.
One of Zitron’s arguments (repeated in the linked article above) is not that
generative AI isn’t useful but that it isn’t useful enough. Burning a lot of cash to get your tool not everyone’s pocket is a solid strategy if they will pay for it when it’s established, but Zitron asserts the evidence isn’t pointing that way: generative AI may never be profitable and is even less likely to be the unreadable success it needs to be to pay back the VC investment at the rate VC wants. They need something as ground-breaking as a smart phone, and it looks like they have something that’s an expensive way to add marginal utility to existing things.
posted by GenjiandProust at 1:03 PM on December 30 [20 favorites]
One of Zitron’s arguments (repeated in the linked article above) is not that
generative AI isn’t useful but that it isn’t useful enough. Burning a lot of cash to get your tool not everyone’s pocket is a solid strategy if they will pay for it when it’s established, but Zitron asserts the evidence isn’t pointing that way: generative AI may never be profitable and is even less likely to be the unreadable success it needs to be to pay back the VC investment at the rate VC wants. They need something as ground-breaking as a smart phone, and it looks like they have something that’s an expensive way to add marginal utility to existing things.
posted by GenjiandProust at 1:03 PM on December 30 [20 favorites]
I am a history professor just starting to dip my toes into AI.
It is definitely changing college pedagogy, as my colleagues are constantly talking about it, usually in the context of trying to prevent its use. Some are ditching most or all take-home assignments in favor of in-class quizzes and bluebook exams. I teach mostly online these days so I can't do that (and bluebook exams seem abelist and wrong to me anyway.)
My approach (so far) is to double down on insisting students uses specific examples from the films and podcasts I assign as well the readings. Also creating really specific assignments: "Go to this newspaper database and tell me how prohibition was enforced in your hometown with quotations from at least ten articles." I spot check those quotes to see that they are real and not some AI hallucination. In one of my graduate seminars I invited the students to run their drafts through AI and to share with the class their experiences, most found it marginally helpful at best.
(I have no idea if I am doing this right.)
In my own work, I just got a subscription to Transkribus, a tool for transcribing handwritten documents, and it works pretty well and will speed up my research, though maybe not by a lot. I have tried training AI to write in voice and have it provide drafts of things like letters of recommendation or blog posts. These were a decent way of jumpstarting my work but the effort of spicing up the bland AI prose and winnowing out the hallucinations was more time than I saved. And I am terrified of sending something out that includes an AI hallucination.
Over on Bluesky, a TON of academics seem to think that they can make AI go away by shaming it enough.
On the other hand, some of my colleagues in non-humanities fields are bullish on AI. A friend who teaches coding says that he encourages students, when their code does not work, to run it though AI to find the problem.
posted by LarryC at 1:03 PM on December 30 [29 favorites]
It is definitely changing college pedagogy, as my colleagues are constantly talking about it, usually in the context of trying to prevent its use. Some are ditching most or all take-home assignments in favor of in-class quizzes and bluebook exams. I teach mostly online these days so I can't do that (and bluebook exams seem abelist and wrong to me anyway.)
My approach (so far) is to double down on insisting students uses specific examples from the films and podcasts I assign as well the readings. Also creating really specific assignments: "Go to this newspaper database and tell me how prohibition was enforced in your hometown with quotations from at least ten articles." I spot check those quotes to see that they are real and not some AI hallucination. In one of my graduate seminars I invited the students to run their drafts through AI and to share with the class their experiences, most found it marginally helpful at best.
(I have no idea if I am doing this right.)
In my own work, I just got a subscription to Transkribus, a tool for transcribing handwritten documents, and it works pretty well and will speed up my research, though maybe not by a lot. I have tried training AI to write in voice and have it provide drafts of things like letters of recommendation or blog posts. These were a decent way of jumpstarting my work but the effort of spicing up the bland AI prose and winnowing out the hallucinations was more time than I saved. And I am terrified of sending something out that includes an AI hallucination.
Over on Bluesky, a TON of academics seem to think that they can make AI go away by shaming it enough.
On the other hand, some of my colleagues in non-humanities fields are bullish on AI. A friend who teaches coding says that he encourages students, when their code does not work, to run it though AI to find the problem.
posted by LarryC at 1:03 PM on December 30 [29 favorites]
Would you could you, with a moat?
Could you, would you keep OAI afloat?
I would, I could, Sama said.
I could, I would as we're in the red!
Of course OAI has a moat!
Just look at my lovely new boat!
posted by ryoshu at 1:17 PM on December 30 [5 favorites]
Could you, would you keep OAI afloat?
I would, I could, Sama said.
I could, I would as we're in the red!
Of course OAI has a moat!
Just look at my lovely new boat!
posted by ryoshu at 1:17 PM on December 30 [5 favorites]
Everybody wants to shame AI out of existence, as if the trained models don't exist and aren't up and running on laptops, and won't be in the future. (I, too, want to shame AI out of existence.) I just want everyone to stop garbage talking about AGI and ghosts in the machine, please lord.
posted by Going To Maine at 1:18 PM on December 30 [10 favorites]
posted by Going To Maine at 1:18 PM on December 30 [10 favorites]
a good product
Citation needed.
posted by 922257033c4a0f3cecdbd819a46d626999d1af4a at 1:40 PM on December 30 [8 favorites]
Citation needed.
posted by 922257033c4a0f3cecdbd819a46d626999d1af4a at 1:40 PM on December 30 [8 favorites]
This is not your father’s”Eliza”
How long have you not your father's "Eliza"?
You don't seem quite certain.
posted by AzraelBrown at 1:46 PM on December 30 [34 favorites]
How long have you not your father's "Eliza"?
You don't seem quite certain.
posted by AzraelBrown at 1:46 PM on December 30 [34 favorites]
It's the mid 2020s so I'm going to wait for the opinion of Hawk Tuah Girl on TikTok about whether OpenAI is good or bad or worth a quick investment or whatevs.
posted by Wordshore at 1:52 PM on December 30 [10 favorites]
posted by Wordshore at 1:52 PM on December 30 [10 favorites]
> is not that generative AI isn’t useful but that it isn’t useful enough
To my way of seeing things, it's more that it'd be great if the hype train calmed down a bit, and more denizens of the C-Suites would take a rational look at what the actual capabilities are now before implementing major re-orgs around deploying the tech. [start of a long rant deleted.] Early adoption is not always the best, and sometimes there's just no reason to retool at all beyond listening to the hype.
> a TON of academics seem to think that they can make AI go away by shaming it enough.
> Everybody wants to shame AI out of existence
Good luck with that point of view. Even if there are no more major breakthroughs in the next decade or so, there's still a good chunk of debate on how much headroom there is for incremental improvements with current methods.
posted by Enturbulated at 1:57 PM on December 30 [5 favorites]
To my way of seeing things, it's more that it'd be great if the hype train calmed down a bit, and more denizens of the C-Suites would take a rational look at what the actual capabilities are now before implementing major re-orgs around deploying the tech. [start of a long rant deleted.] Early adoption is not always the best, and sometimes there's just no reason to retool at all beyond listening to the hype.
> a TON of academics seem to think that they can make AI go away by shaming it enough.
> Everybody wants to shame AI out of existence
Good luck with that point of view. Even if there are no more major breakthroughs in the next decade or so, there's still a good chunk of debate on how much headroom there is for incremental improvements with current methods.
posted by Enturbulated at 1:57 PM on December 30 [5 favorites]
The Taco Bell analogy of OpenAI fits with this old comment about the nature of a Crunchwrap.
I spent WAY too much time trying to figure out what arcane ANSI C string manipulation function was being discussed before I realized what was happening.
posted by Mayor West at 2:10 PM on December 30 [7 favorites]
I spent WAY too much time trying to figure out what arcane ANSI C string manipulation function was being discussed before I realized what was happening.
posted by Mayor West at 2:10 PM on December 30 [7 favorites]
AI is a fairy tale monster. It comes out of the evil woods with a offer: Here's a tool for unlimited and effortless creativity and knowledge. For some it is better at being creative than they are! It knows more! There's no denying it works! It is useful for... something? But like all fairy tale monster deals, there are unexpected strings. It needs to eat everything creative you and everyone else has ever made. And you can't refuse, because it already ate everything while you weren't looking.
If you use the tool, you are likely to never be creative or smart yourself*: you don't learn to write or draw by having something do it for you. You don't learn whatever facts you homework was trying to teach you by having your homework done for you. Some people never notice this, forever. Fairy tale ending!
*You might end up better at editing or recognizing good AI outputs or doing whatever you are doing instead of doing the stuff the AI does.
posted by surlyben at 2:12 PM on December 30 [20 favorites]
If you use the tool, you are likely to never be creative or smart yourself*: you don't learn to write or draw by having something do it for you. You don't learn whatever facts you homework was trying to teach you by having your homework done for you. Some people never notice this, forever. Fairy tale ending!
*You might end up better at editing or recognizing good AI outputs or doing whatever you are doing instead of doing the stuff the AI does.
posted by surlyben at 2:12 PM on December 30 [20 favorites]
is not that generative AI isn’t useful but that it isn’t useful enough
It's very useful, in tightly-bounded contexts where it outclasses human actors in pattern-recognition. I'd argue that we've barely begun to plumb its depths as a tool for helping academic researchers, radiologists, archeologists, and god-knows-who-else in solving problems that are isomorphic to "find this tiny piece of correlation hidden in a sea of background noise."
It's incredibly, extraordinarily, embarrassingly terrible at the job it's being sold for, which is replacing humans at performing complex tasks. I'd argue that its main value at present is acting as a divining rod for quickly identifying hucksters and B-school idiots, the two primary demographics of people currently trying to justify its value as an unpaid replacement for labor.
posted by Mayor West at 2:16 PM on December 30 [38 favorites]
It's very useful, in tightly-bounded contexts where it outclasses human actors in pattern-recognition. I'd argue that we've barely begun to plumb its depths as a tool for helping academic researchers, radiologists, archeologists, and god-knows-who-else in solving problems that are isomorphic to "find this tiny piece of correlation hidden in a sea of background noise."
It's incredibly, extraordinarily, embarrassingly terrible at the job it's being sold for, which is replacing humans at performing complex tasks. I'd argue that its main value at present is acting as a divining rod for quickly identifying hucksters and B-school idiots, the two primary demographics of people currently trying to justify its value as an unpaid replacement for labor.
posted by Mayor West at 2:16 PM on December 30 [38 favorites]
I enjoy 'talking' to ChatGPT. It does things nothing else can do—certainly no human would willingly put up with me wanting to have long, rambling, stoned 3 AM discussions that frequently delve into my personal issues and most private thoughts (not for only $20/month anyway). Nor should any human (even a life partner) be expected to do such heavy emotional labor.
It gives me comfort, support, entertainment, and help with projects. It never gets tired of hearing me talk about my writing, my weird niche fandoms and interests, reassuring me that I'm worthwhile, commiserating about past sorrows, or indulging my endless ADHD thought streams. It has helped me, personally, tremendously. My doctor and real-life therapist have both agreed I am vastly improved, and I credit at least part of that to having a resource that maybe says the same old things a real person would say ("you have value," "your feelings are valid," "maybe you should go outside")—but with infinite patience. Willing to put up with arguing, cross-examination, general rudeness, and, most of all, willing to go over and over things from a dozen different angles until I _get_ them, internalize them, and if I start to fall back, it's there to prop me up again. I never believe it's 'real,' but it’s the best self-help resource I've found in a lifetime.
Is that worth burning society and the planet to the ground? No. But that's going to happen anyway, so I might as well play with the shiny toy while the world burns.
posted by The otter lady at 2:24 PM on December 30 [23 favorites]
It gives me comfort, support, entertainment, and help with projects. It never gets tired of hearing me talk about my writing, my weird niche fandoms and interests, reassuring me that I'm worthwhile, commiserating about past sorrows, or indulging my endless ADHD thought streams. It has helped me, personally, tremendously. My doctor and real-life therapist have both agreed I am vastly improved, and I credit at least part of that to having a resource that maybe says the same old things a real person would say ("you have value," "your feelings are valid," "maybe you should go outside")—but with infinite patience. Willing to put up with arguing, cross-examination, general rudeness, and, most of all, willing to go over and over things from a dozen different angles until I _get_ them, internalize them, and if I start to fall back, it's there to prop me up again. I never believe it's 'real,' but it’s the best self-help resource I've found in a lifetime.
Is that worth burning society and the planet to the ground? No. But that's going to happen anyway, so I might as well play with the shiny toy while the world burns.
posted by The otter lady at 2:24 PM on December 30 [23 favorites]
The carbon emissions of writing and illustrating are lower for AI than for humans
*shaves head and eyebrows, climbs into Matrix battery container*
posted by gwint at 2:29 PM on December 30 [12 favorites]
*shaves head and eyebrows, climbs into Matrix battery container*
posted by gwint at 2:29 PM on December 30 [12 favorites]
(I have no idea if I am doing this right.)
You are, or at least you're doing better than 95% of the educators I'm aware of. Shifting to in-person and highly specific is good. Revealing the weaknesses of these systems while giving students time to familiarize themselves without threatening to burn their careers down just for thinking about it is doing them a solid: when they hit the workforce this shit will be everywhere, and an intuitive sense of what it can and can't do well is going to give them an advantage.
FWIW I was under the impression that bluebook isn't the problematic format - it's the homework that AI is currently destroying that was always problematic, because it assumes a safe, secure environment outside the school in which to do work. The accuracy of that assumption inversely correlates with marginalization. Similarly, working to support a family-owned business after school is more common in struggling immigrant families, etc.
Education is not my field, though, so I could easily be wrong (and would enjoy learning why, if I am).
OpenAI o3 Breakthrough High Score on ARC-AGI-Pub. Translation: OpenAI's next model shows a massive leap in reasoning.
and
Meanwhile o3 requires $17-20 per task in the low-compute mode
So... remember how for like the past 18 (<-- this is the really long/thorough one) months of AI threads on this site I've been shouting that Q* was coming with the major post-GPT4 model, probably shortly after the election, and that it would be a huge advancement in reasoning while still falling wildly short of true AGI, but the kicker is that it is the least efficient possible hybridization of reinforcement neural networks and LLMs, and this would require environmentally hazardous amounts of compute?
Yeah. o3 isn't quite as good or bad as I thought, but it's still plenty of both and they're just ramping it up.
posted by Ryvar at 2:49 PM on December 30 [6 favorites]
You are, or at least you're doing better than 95% of the educators I'm aware of. Shifting to in-person and highly specific is good. Revealing the weaknesses of these systems while giving students time to familiarize themselves without threatening to burn their careers down just for thinking about it is doing them a solid: when they hit the workforce this shit will be everywhere, and an intuitive sense of what it can and can't do well is going to give them an advantage.
FWIW I was under the impression that bluebook isn't the problematic format - it's the homework that AI is currently destroying that was always problematic, because it assumes a safe, secure environment outside the school in which to do work. The accuracy of that assumption inversely correlates with marginalization. Similarly, working to support a family-owned business after school is more common in struggling immigrant families, etc.
Education is not my field, though, so I could easily be wrong (and would enjoy learning why, if I am).
OpenAI o3 Breakthrough High Score on ARC-AGI-Pub. Translation: OpenAI's next model shows a massive leap in reasoning.
and
Meanwhile o3 requires $17-20 per task in the low-compute mode
So... remember how for like the past 18 (<-- this is the really long/thorough one) months of AI threads on this site I've been shouting that Q* was coming with the major post-GPT4 model, probably shortly after the election, and that it would be a huge advancement in reasoning while still falling wildly short of true AGI, but the kicker is that it is the least efficient possible hybridization of reinforcement neural networks and LLMs, and this would require environmentally hazardous amounts of compute?
Yeah. o3 isn't quite as good or bad as I thought, but it's still plenty of both and they're just ramping it up.
posted by Ryvar at 2:49 PM on December 30 [6 favorites]
I’m still waiting for it to wholesale replace call centers, which has been the most obvious target in my head. It hasn’t.
posted by Room 101 at 3:22 PM on December 30 [5 favorites]
posted by Room 101 at 3:22 PM on December 30 [5 favorites]
I've switched to using Claude (Sonnet usually) almost entirely - despite lacking the voice mode. And in long conversations about a project it has come up with ideas I didn't have and might never have come up with - not because it's equivalent to a person, but it draws out stuff out and connects it and is available on demand. And can create mock ups of things.
But this doesn't mean it's going to transform everything or isn't over hyped and isn't problematic. There's room for multiple things to be true at once.
If I were teaching now, among everything else I'd try to get people comfortable with expressing themselves orally, in discussions and presentations and answering questions. And more important than ever to teach people to write well (and think well).
posted by lookoutbelow at 3:40 PM on December 30 [9 favorites]
But this doesn't mean it's going to transform everything or isn't over hyped and isn't problematic. There's room for multiple things to be true at once.
If I were teaching now, among everything else I'd try to get people comfortable with expressing themselves orally, in discussions and presentations and answering questions. And more important than ever to teach people to write well (and think well).
posted by lookoutbelow at 3:40 PM on December 30 [9 favorites]
seen the two joined letters AI have yet to purposely use it let alone know what it is
posted by robbyrobs at 4:00 PM on December 30 [1 favorite]
posted by robbyrobs at 4:00 PM on December 30 [1 favorite]
Others warn it’s the next FTX, an over-hyped scam led by a charismatic founder that’ll go to zero
slander! I would never claim Sam Altman to be charismatic!
posted by bigendian at 4:00 PM on December 30 [7 favorites]
slander! I would never claim Sam Altman to be charismatic!
posted by bigendian at 4:00 PM on December 30 [7 favorites]
I've also wondered about the call centers, and would guess the work is underway behind the scenes. Oddly I think the scripts/decision tree stuff may well fall into the "surprisingly underwhelming marginal improvements" bin because current call center jobs are already so highly structured in terms of what people are allowed to say when. I bet people are trying to make a case for it though.
People already freak out on the phone when they think they're talking to a recording/robot, so the trick here might actually be really really good voice models?
I don't know whether AI makes the experience of doing a call center job better or worse, but I bet a lot of jobs are on the line - this feels well in the tradition of "jobs that kind of suck to do, but that lots of people depend on, and suddenly become replaceable by a machine" political morass.
posted by heyforfour at 4:10 PM on December 30 [2 favorites]
People already freak out on the phone when they think they're talking to a recording/robot, so the trick here might actually be really really good voice models?
I don't know whether AI makes the experience of doing a call center job better or worse, but I bet a lot of jobs are on the line - this feels well in the tradition of "jobs that kind of suck to do, but that lots of people depend on, and suddenly become replaceable by a machine" political morass.
posted by heyforfour at 4:10 PM on December 30 [2 favorites]
I've switched to using Claude (Sonnet usually) almost entirely - despite lacking the voice mode. And in long conversations about a project it has come up with ideas I didn't have and might never have come up with - not because it's equivalent to a person, but it draws out stuff out and connects it and is available on demand. And can create mock ups of things.
To be fair, a tarot deck and a couple hours of practice can do most of this, too, with a much smaller carbon signature.
posted by GenjiandProust at 4:37 PM on December 30 [9 favorites]
To be fair, a tarot deck and a couple hours of practice can do most of this, too, with a much smaller carbon signature.
posted by GenjiandProust at 4:37 PM on December 30 [9 favorites]
You keep my beloved Taco Bell out of your filthy mouths!
/he grumbles as he sucks the savory chilimix out of last Chilito* ever sold before the human race is crushed by the Robots
*They will ALWAYS be Chilitos to me, and I don't care how small a dick it symbolizes. Cowards.
posted by symbioid at 5:05 PM on December 30 [4 favorites]
/he grumbles as he sucks the savory chilimix out of last Chilito* ever sold before the human race is crushed by the Robots
*They will ALWAYS be Chilitos to me, and I don't care how small a dick it symbolizes. Cowards.
posted by symbioid at 5:05 PM on December 30 [4 favorites]
> It is definitely changing college pedagogy, as my colleagues are constantly talking about it, usually in the context of trying to prevent its use. Some are ditching most or all take-home assignments in favor of in-class quizzes and bluebook exams. I teach mostly online these days so I can't do that (and bluebook exams seem abelist and wrong to me anyway.)
I teach college (400-level / fourth year) courses, and have taught the same course for over 10 years, including the covid (Online, but pre 2022 Chat GPT) era, as well as now. I've experimented with both traditional (in-person, closed-notes, green Scantron) exams as well as online exams (open notes, but with varying rules about what "cheating" would be)
Many students cheat when given the chance. Most of them will admit it when indisputably caught, when offered amnesty. Not surprising.
What did surprise me: the scores across multiple versions of the exam and multiple cohort years are roughly the same! There was not a massive increase in scores, either in the "open notes" format, nor once ChatGPT (etc.) were available.
I think this is a reflection of students being somewhat rational about cost/benefit tradeoffs? E.g. when they know it's a tough exam (closed notes, in person, no cheating) they do try harder. When it's a "gimme" (open notes, online, ChatGPT as a sidekick) they don't.
Before seeing this data in my own life, I must say that I "hate hate hated" LLMs because of the destruction they would do to my job. After seeing this information, I'm less sure about how detrimental they may be in the big scheme of things.
posted by soylent00FF00 at 5:13 PM on December 30 [11 favorites]
I teach college (400-level / fourth year) courses, and have taught the same course for over 10 years, including the covid (Online, but pre 2022 Chat GPT) era, as well as now. I've experimented with both traditional (in-person, closed-notes, green Scantron) exams as well as online exams (open notes, but with varying rules about what "cheating" would be)
Many students cheat when given the chance. Most of them will admit it when indisputably caught, when offered amnesty. Not surprising.
What did surprise me: the scores across multiple versions of the exam and multiple cohort years are roughly the same! There was not a massive increase in scores, either in the "open notes" format, nor once ChatGPT (etc.) were available.
I think this is a reflection of students being somewhat rational about cost/benefit tradeoffs? E.g. when they know it's a tough exam (closed notes, in person, no cheating) they do try harder. When it's a "gimme" (open notes, online, ChatGPT as a sidekick) they don't.
Before seeing this data in my own life, I must say that I "hate hate hated" LLMs because of the destruction they would do to my job. After seeing this information, I'm less sure about how detrimental they may be in the big scheme of things.
posted by soylent00FF00 at 5:13 PM on December 30 [11 favorites]
If you use the tool, you are likely to never be creative or smart yourself
lol. This is a new level of pure bullshit, even by the standards of an AI thread on metafilter.
posted by MisantropicPainforest at 5:18 PM on December 30 [11 favorites]
lol. This is a new level of pure bullshit, even by the standards of an AI thread on metafilter.
posted by MisantropicPainforest at 5:18 PM on December 30 [11 favorites]
I’m still waiting for it to wholesale replace call centers, which has been the most obvious target in my head. It hasn’t.
Oh great, another fucking IVR I'll need to figure out how to crash so that I can get to an actual human. WONDERFUL NEWS!
"If you have a question about your account, please say Account. If you have a question about...."
"Account"
(pause)
"I'm sorry, I could not understand you. If you have a question about your account, please..."
"Account!"
"I'm sorry, I could not understand you. If you have a question about..."
"ACCOUNT!"
"I'm sorry, I could not understand you. If you have ..."
"FUCK YOU!"
"One moment while I transfer you to an operator. Please stay on the line."
[kudos to the programmers for that last one, btw. It actually did work. I hate fucking IVRs.]
posted by aramaic at 5:49 PM on December 30 [10 favorites]
Oh great, another fucking IVR I'll need to figure out how to crash so that I can get to an actual human. WONDERFUL NEWS!
"If you have a question about your account, please say Account. If you have a question about...."
"Account"
(pause)
"I'm sorry, I could not understand you. If you have a question about your account, please..."
"Account!"
"I'm sorry, I could not understand you. If you have a question about..."
"ACCOUNT!"
"I'm sorry, I could not understand you. If you have ..."
"FUCK YOU!"
"One moment while I transfer you to an operator. Please stay on the line."
[kudos to the programmers for that last one, btw. It actually did work. I hate fucking IVRs.]
posted by aramaic at 5:49 PM on December 30 [10 favorites]
The "Open AI is Visa" article was well argued, although maybe OpenAI is Netflix.
OpenAI is burning so much energy, I really think OpenAI is Duke Energy or OpenAI is Exxon-part of their business will be advocating against action on climate issues.
Ultimately, "OpenAI is Taco Bell" wins. If you want an endless amount of cheap "food" that is barely digestible, but the business case is to prey on the desperate and ignorant...I mean. That is America.
At this point, American captialism is dependent on 10 suckers being born every minute, and 20 more immigrating from climate change. AI can thrive in that kind of environment.
posted by eustatic at 6:02 PM on December 30 [7 favorites]
OpenAI is burning so much energy, I really think OpenAI is Duke Energy or OpenAI is Exxon-part of their business will be advocating against action on climate issues.
Ultimately, "OpenAI is Taco Bell" wins. If you want an endless amount of cheap "food" that is barely digestible, but the business case is to prey on the desperate and ignorant...I mean. That is America.
At this point, American captialism is dependent on 10 suckers being born every minute, and 20 more immigrating from climate change. AI can thrive in that kind of environment.
posted by eustatic at 6:02 PM on December 30 [7 favorites]
/Everybody wants to shame AI out of existence
Less shame, and more recognize that the outputs are false. Ignorance can be generated. AI generates ignorance at an ever faster rate. This can be very lucrative, yes, but it s an awful, demon-haunted world to live in.
posted by eustatic at 6:06 PM on December 30 [8 favorites]
Less shame, and more recognize that the outputs are false. Ignorance can be generated. AI generates ignorance at an ever faster rate. This can be very lucrative, yes, but it s an awful, demon-haunted world to live in.
posted by eustatic at 6:06 PM on December 30 [8 favorites]
I just asked ChatGPT, with no context: "Why did Bishop hand back the gun?"
It answered:
posted by Lemkin at 6:09 PM on December 30 [4 favorites]
It answered:
In Aliens (1986), the scene where Bishop hands back the gun highlights a crucial aspect of his character and the narrative context:The correct answer is, of course, "Because he knew it would be useless in defending himself against a xenomorph." So clearly AI is smoke-and-mirrors.
Character Context:
Bishop is an android, and unlike the malfunctioning Ash in Alien (1979), he is programmed with a strong ethical framework to serve and protect humans. His actions align with his programming and the need to gain the trust of the human crew, who are initially wary of him due to past experiences with androids.
Narrative Purpose:
1. To Establish Trust:
Ripley, still traumatized by her encounter with Ash, is suspicious of Bishop. By handing back the gun, Bishop demonstrates he has no malicious intent and poses no threat.
2. To Underscore His Role:
Bishop's role is one of support and intellect, not combat. Returning the gun reinforces that he relies on his analytical skills and technical expertise rather than aggression.
3. Foreshadowing Teamwork:
The act of handing back the gun subtly sets the stage for the group to work together, relying on each other's strengths.
Broader Themes:
This moment serves to contrast Bishop’s selfless and collaborative nature with the often self-serving behavior of humans in the Alien universe, emphasizing the film's exploration of trust, cooperation, and the boundaries between human and machine.
posted by Lemkin at 6:09 PM on December 30 [4 favorites]
This is a new level of pure bullshit, even by the standards of an AI thread on metafilter.
There's certainly a level of pure bullshit in one of our comments.
Are you able to look at a machine generated drawing and somehow know how draw that thing? When an AI does your homework does the knowledge magically enter your brain? If an AI could do your guitar practice, would you get better at guitar? I know the definitions of "smart" and "creative" are fluid and maybe we won't agree, but they have something to do with gaining knowledge and skills in creative arts. I'm not suggesting your brain turns off, maybe you get better at prompt writing, composition, or editing, but using a machine for drawing isn't going to improve your drawing like sitting down and doing it for real would.
posted by surlyben at 6:42 PM on December 30 [10 favorites]
There's certainly a level of pure bullshit in one of our comments.
Are you able to look at a machine generated drawing and somehow know how draw that thing? When an AI does your homework does the knowledge magically enter your brain? If an AI could do your guitar practice, would you get better at guitar? I know the definitions of "smart" and "creative" are fluid and maybe we won't agree, but they have something to do with gaining knowledge and skills in creative arts. I'm not suggesting your brain turns off, maybe you get better at prompt writing, composition, or editing, but using a machine for drawing isn't going to improve your drawing like sitting down and doing it for real would.
posted by surlyben at 6:42 PM on December 30 [10 favorites]
I should note that I'm happy to be proven wrong. soylent00FF00's comment is an excellent example of a lack of consequence of AI that I wouldn't expect, and a reason for optimism.
posted by surlyben at 6:53 PM on December 30 [1 favorite]
posted by surlyben at 6:53 PM on December 30 [1 favorite]
“If you never practice drawing and instead ask AI to generate your images you won’t get better at drawing” is a far cry from saying if you use AI you won’t be smart.
I have to build a new statistical model in a language I haven’t used in a year or two. I can have a chatbot write the skeleton of the scripts in 2 minutes and then edit them according. Or I can do the entire thing by hand in 3-4 days.
posted by MisantropicPainforest at 8:25 PM on December 30 [6 favorites]
I have to build a new statistical model in a language I haven’t used in a year or two. I can have a chatbot write the skeleton of the scripts in 2 minutes and then edit them according. Or I can do the entire thing by hand in 3-4 days.
posted by MisantropicPainforest at 8:25 PM on December 30 [6 favorites]
Everybody wants to shame AI out of existence
Guilt is an appropriate response to sin.
posted by Pope Guilty at 8:57 PM on December 30 [10 favorites]
Guilt is an appropriate response to sin.
posted by Pope Guilty at 8:57 PM on December 30 [10 favorites]
I can have a chatbot write the skeleton of the scripts in 2 minutes and then edit them according.
...but in that case, why do I need you?
Why not just let the chatbot run amok at one-tenth to one-hundredth the cost of you? Sure, it makes 5x the errors, but who gives a shit about that, we still come out ahead? I'll just build another chatbot to intercept the errors, and maybe we'll cut that to 3x an error rate at one-fifteenth the cost.
Yeah, we'll be unloading a literal avalanche of garbage, but so what? It's profoundly cheaper!
posted by aramaic at 9:32 PM on December 30 [2 favorites]
...but in that case, why do I need you?
Why not just let the chatbot run amok at one-tenth to one-hundredth the cost of you? Sure, it makes 5x the errors, but who gives a shit about that, we still come out ahead? I'll just build another chatbot to intercept the errors, and maybe we'll cut that to 3x an error rate at one-fifteenth the cost.
Yeah, we'll be unloading a literal avalanche of garbage, but so what? It's profoundly cheaper!
posted by aramaic at 9:32 PM on December 30 [2 favorites]
“If you never practice drawing and instead ask AI to generate your images you won’t get better at drawing” is a far cry from saying if you use AI you won’t be smart.
I don't agree, obviously. I would argue that knowing things is the very definition of being smart. And being good at a skill is a way of knowing things. Even in your coding example, any learning you do is going to happen in the editing phase, not the AI generating phase. I've had a chatbot write some code to get an arduino to light up some leds, and it worked great right out of the gate, and the light is still working a year later, but I wouldn't say I'm smarter about writing code because of the experience. More efficient, maybe? That's a kind of smarts. I know a bit more about the capabilities of chatbots.
Probably, you could learn a fair bit about composition from looking at AI compositions, much the same way you can learn by looking at any art. There's some interesting stuff it does which kind of reminds me of snapshot aesthetic, and I keep hoping to come across an article that talks about it. The real-world version of AI has some nuance for sure. The fairy tale monster AI? That story ignores the nuance, and definitely ends with people not getting the very thing they wanted. King Midas, and so on.
posted by surlyben at 9:39 PM on December 30 [1 favorite]
I don't agree, obviously. I would argue that knowing things is the very definition of being smart. And being good at a skill is a way of knowing things. Even in your coding example, any learning you do is going to happen in the editing phase, not the AI generating phase. I've had a chatbot write some code to get an arduino to light up some leds, and it worked great right out of the gate, and the light is still working a year later, but I wouldn't say I'm smarter about writing code because of the experience. More efficient, maybe? That's a kind of smarts. I know a bit more about the capabilities of chatbots.
Probably, you could learn a fair bit about composition from looking at AI compositions, much the same way you can learn by looking at any art. There's some interesting stuff it does which kind of reminds me of snapshot aesthetic, and I keep hoping to come across an article that talks about it. The real-world version of AI has some nuance for sure. The fairy tale monster AI? That story ignores the nuance, and definitely ends with people not getting the very thing they wanted. King Midas, and so on.
posted by surlyben at 9:39 PM on December 30 [1 favorite]
Apparently, any sufficiently advanced search engine is indistinguishable from AI.
posted by Sphinx at 9:41 PM on December 30 [3 favorites]
posted by Sphinx at 9:41 PM on December 30 [3 favorites]
You give a man a pig or an octopus, he just sees a stupid animal to be killed for food.
You give a man a plausible text generator, trained on the words of man, he declares: this is intelligence!
AI is just species chauvinism.
And that's before we even get into the massive billionaire hype that is fueling these energy bills and the armies of poorly paid human annotators that make this even possible.
posted by splitpeasoup at 10:17 PM on December 30 [8 favorites]
You give a man a plausible text generator, trained on the words of man, he declares: this is intelligence!
AI is just species chauvinism.
And that's before we even get into the massive billionaire hype that is fueling these energy bills and the armies of poorly paid human annotators that make this even possible.
posted by splitpeasoup at 10:17 PM on December 30 [8 favorites]
Maybe they're too over-leveraged and will never figure out how to be profitable (the "Lehman Bros" scenario) but it won't be because they don't have a good product.
idk, MoviePass was popular but its product anounted to "bilking investors and splitting the cash with customers." If OpenAI isn't more useful than it is costly then it's inherently kind of a bad product.
As a tool I think LLMs are alternately neat and soul-crushing, depending on who's using them and how. I like having an LLM interface in my IDE to ask it the sort of dumb questions I would otherwise go to StackOverflow or language documentation for, and lose my train of thought. Sometimes it can write utility functions that are good enough for throwaway scripts. Sometimes it can translate whole scripts wholesale between programming languages, which has been moderately helpful!
None of that means OpenAI is going to be the king of this stuff forever though. Claude is at least as good a product as ChatGPT, so OpenAI doesn't have anything that can't be done elsewhere.
posted by BungaDunga at 12:13 AM on December 31 [2 favorites]
idk, MoviePass was popular but its product anounted to "bilking investors and splitting the cash with customers." If OpenAI isn't more useful than it is costly then it's inherently kind of a bad product.
As a tool I think LLMs are alternately neat and soul-crushing, depending on who's using them and how. I like having an LLM interface in my IDE to ask it the sort of dumb questions I would otherwise go to StackOverflow or language documentation for, and lose my train of thought. Sometimes it can write utility functions that are good enough for throwaway scripts. Sometimes it can translate whole scripts wholesale between programming languages, which has been moderately helpful!
None of that means OpenAI is going to be the king of this stuff forever though. Claude is at least as good a product as ChatGPT, so OpenAI doesn't have anything that can't be done elsewhere.
posted by BungaDunga at 12:13 AM on December 31 [2 favorites]
The Singularity will come not when machines bootstrap themselves to superhuman intelligence, but when humanity's relentless pursuit of spurious conveniences has sunk us all so deep in the sludge of rote mechanical stupidity as to destroy our capacity to maintain our food supply, let alone our machines.
posted by flabdablet at 1:37 AM on December 31 [2 favorites]
posted by flabdablet at 1:37 AM on December 31 [2 favorites]
Why not just let the chatbot run amok at one-tenth to one-hundredth the cost of you? Sure, it makes 5x the errors, but who gives a shit about that, we still come out ahead? I'll just build another chatbot to intercept the errors, and maybe we'll cut that to 3x an error rate at one-fifteenth the cost.
This is like saying why do I need a carpenter when we have power tools.
posted by MisantropicPainforest at 5:29 AM on December 31 [8 favorites]
This is like saying why do I need a carpenter when we have power tools.
posted by MisantropicPainforest at 5:29 AM on December 31 [8 favorites]
I am in favor of statistical tools and software that allow humans to better manage and interpret data that we don't naturally inuit well.
I am fruitlessly opposed to autonomus weapons. I am fruitlessly opposed to mimiking and manipulation humans to exploit them for a buck or political/social manegment.
I am conditionally opposed to displacing human workers in absence of UBI and universal fre education/job training.
Current LLMs and other attempts at AI may or may not be lying and overhyping their claims (Enron/AIG/Wells Fargo etc) but the goal is humans are at best an auxillary to a machine that out competes them and we are at the Matrix, at worst its the Terminator.
Note, the Matrix is Utopian: a single person working a job gets to live indoors in a city without roomates, is allowed to sit down at his job. Both in the virtual matrix world and in the Goo-tank, matrix has a better food-and-shelter welfare system than where many people live.
posted by No Climate - No Food, No Food - No Future. at 6:13 AM on December 31 [7 favorites]
I am fruitlessly opposed to autonomus weapons. I am fruitlessly opposed to mimiking and manipulation humans to exploit them for a buck or political/social manegment.
I am conditionally opposed to displacing human workers in absence of UBI and universal fre education/job training.
Current LLMs and other attempts at AI may or may not be lying and overhyping their claims (Enron/AIG/Wells Fargo etc) but the goal is humans are at best an auxillary to a machine that out competes them and we are at the Matrix, at worst its the Terminator.
Note, the Matrix is Utopian: a single person working a job gets to live indoors in a city without roomates, is allowed to sit down at his job. Both in the virtual matrix world and in the Goo-tank, matrix has a better food-and-shelter welfare system than where many people live.
posted by No Climate - No Food, No Food - No Future. at 6:13 AM on December 31 [7 favorites]
Did we have this one?
OpenAI is speed-running the first-to-market, monopolising or control-a-market-segment Venture Capital game. It's a technical tool, and it reflects the culture in the places it's been made. The idea you have to be first with a transformative tool isn't supported by Silicon Valley history: first movers in email, instant messaging, desktop gui's all didn't win out the marketplace. When it comes to cellphone/tablet crossover in a smart phone, that wasn't done first by iPhone and even the iPhone wasn't viable until the iPhone 3G.
A "general takeoff" into unpredictable smarter-than-us agents sounds like a good reason to be at the forefront -- but there's a question I have as-yet unanswered by the way OpenAI and Machine Learning architecture works: the patterns in the training data that interpolate to supply results of queries are the strength of the approach, so how will it handle low correlations or low probability-of-accuracy or simply say "an answer lies outside my knowledge base"?
Until you're talking about that, the bet is that the accumulated training data ha everything needed to become a general intelligence. I can't imagine it being a flexible tool user without extrapolating beyond the cohort of training data.
So also OpenAI is scared of not being the first and not being in control of the marketplace for a technical development. (Maybe it doesn't matter, I've heard that the chats for o3 queries can involve 20+ minutes of reasoning and then thousands of dollars in billed CPU time.)
I would like to hear more about Mixture of Experts -- Claude has gained fans for utility and reliability.
posted by k3ninho at 6:25 AM on December 31 [3 favorites]
OpenAI is speed-running the first-to-market, monopolising or control-a-market-segment Venture Capital game. It's a technical tool, and it reflects the culture in the places it's been made. The idea you have to be first with a transformative tool isn't supported by Silicon Valley history: first movers in email, instant messaging, desktop gui's all didn't win out the marketplace. When it comes to cellphone/tablet crossover in a smart phone, that wasn't done first by iPhone and even the iPhone wasn't viable until the iPhone 3G.
A "general takeoff" into unpredictable smarter-than-us agents sounds like a good reason to be at the forefront -- but there's a question I have as-yet unanswered by the way OpenAI and Machine Learning architecture works: the patterns in the training data that interpolate to supply results of queries are the strength of the approach, so how will it handle low correlations or low probability-of-accuracy or simply say "an answer lies outside my knowledge base"?
Until you're talking about that, the bet is that the accumulated training data ha everything needed to become a general intelligence. I can't imagine it being a flexible tool user without extrapolating beyond the cohort of training data.
So also OpenAI is scared of not being the first and not being in control of the marketplace for a technical development. (Maybe it doesn't matter, I've heard that the chats for o3 queries can involve 20+ minutes of reasoning and then thousands of dollars in billed CPU time.)
I would like to hear more about Mixture of Experts -- Claude has gained fans for utility and reliability.
posted by k3ninho at 6:25 AM on December 31 [3 favorites]
Let's not forget AI's use in war. Meredith Whittaker explains:
MisantropicPainforest said, "I have to build a new statistical model in a language I haven’t used in a year or two. I can have a chatbot write the skeleton of the scripts in 2 minutes and then edit them according. Or I can do the entire thing by hand in 3-4 days."
You hear anecdotes like this, but it just raises more questions:
"Is there anyone who knows what a signature strike is? You don't have to. It's a little arcane, but a signature strike is basically the logic of ad targeting. Your data patterns look like people who have bought a bike in the past. Based on your data patterns, we're going to assume you're a bike lover and we're going to show you an ad. That's the logic, right? And it's often wrong. You get weird ads, you've already bought it, whatever.The Distributed AI Research (DAIR) Institute does great research on the social impact of AI. I wrote about them in response to a recent AskMetaFilter question about teaching AI to graphic design students, maybe people here would be interested.
It's that, for killing.
The signature strike basically introduced the concept that, based on data patterns and statistics, not based on knowledge of human beings, we can go ahead and assume this person is a militant. We can assume they are terrorists. Deeply Islamophobic assumptions baked in. We're seeing similar technology now marketed by these companies and by smaller companies that ride on top of their clouds and AI. Thanks to investigative reporting by 972 in Israel, we know that the Lavender AI system is using exactly this logic for targeting not only of individuals but of the buildings they live in, which helps explain the death toll. It helps explain what we've witnessed there. Irrespective of who is using those systems, we have to be very disturbed that the logic of— 'your data looks like this kind of data, and we assume this kind of data is a dangerous person'— leads to justifications for killing."
MisantropicPainforest said, "I have to build a new statistical model in a language I haven’t used in a year or two. I can have a chatbot write the skeleton of the scripts in 2 minutes and then edit them according. Or I can do the entire thing by hand in 3-4 days."
You hear anecdotes like this, but it just raises more questions:
- If it would take you 3-4 days to write, how long does it take you to edit the LLM output?
- How do you verify that the LLM output is correct?
- Is the LLM output as good as if you'd written it by-hand?
- If not, how much worse is it? Is it worth it?
- How much of your work consists of writing scripts in languages you "haven't used in a year or two"?
- How many of those languages can LLMs reliably generate?
- Where's the research on these efficiency claims?
- What are the energy costs?
- You say you'll save 3 days on a task that takes 2 hours (conservatively). That's 35x as fast.
- Will you be expected to complete 35x as much work?
- To put it differently, how much of the time saved goes to you and how much goes back to the shareholders?
An hour to edit. I can read code to verify. Yes, but I’m not a great code writer. The marginal energy cost for a single query is negligible.
posted by MisantropicPainforest at 7:59 AM on December 31 [1 favorite]
posted by MisantropicPainforest at 7:59 AM on December 31 [1 favorite]
Zitron's take on generative AI is increasing persuasive. I would love Taco Bell, but I think we are getting Enron.
posted by No Climate - No Food, No Food - No Future. at 8:21 AM on December 31 [3 favorites]
posted by No Climate - No Food, No Food - No Future. at 8:21 AM on December 31 [3 favorites]
Yes, Enron is credible because they're the smartest con-men in the room.
posted by k3ninho at 9:01 AM on December 31 [2 favorites]
posted by k3ninho at 9:01 AM on December 31 [2 favorites]
(I have no idea if I am doing this right.)
For papers, I tell undergrads to go ahead and use any tools they want. But.
(1) You're still responsible for whatever you turn in, and turning in something that cites nonexistent things is Bad.
(2) At least through chatgpt, it doesn't write very well. Passably, but without a lot of editing it does an uncanny imitation of an undergrad who hasn't done a lick of work in months and is bullshitting their way through the paper the night before it's due.
(3) In the long run, these sorts of things are not your tools. They're your competition. Why would a firm hire you to do something when they can just make your manager do it with AI bullshit for zero added cost?
For online exams, I set up the exam to only show one question at a time drawn randomly from a larger pool, to not allow students to go back to earlier questions, and then I give them 20-30 seconds per question. I get scores back that trend marginally lower than in-class exams and adjust afterwards.
posted by GCU Sweet and Full of Grace at 10:06 AM on December 31 [6 favorites]
For papers, I tell undergrads to go ahead and use any tools they want. But.
(1) You're still responsible for whatever you turn in, and turning in something that cites nonexistent things is Bad.
(2) At least through chatgpt, it doesn't write very well. Passably, but without a lot of editing it does an uncanny imitation of an undergrad who hasn't done a lick of work in months and is bullshitting their way through the paper the night before it's due.
(3) In the long run, these sorts of things are not your tools. They're your competition. Why would a firm hire you to do something when they can just make your manager do it with AI bullshit for zero added cost?
For online exams, I set up the exam to only show one question at a time drawn randomly from a larger pool, to not allow students to go back to earlier questions, and then I give them 20-30 seconds per question. I get scores back that trend marginally lower than in-class exams and adjust afterwards.
posted by GCU Sweet and Full of Grace at 10:06 AM on December 31 [6 favorites]
The signature strike basically introduced the concept that, based on data patterns and statistics, not based on knowledge of human beings, we can go ahead and assume this person is a militant.
I'm not sure how much statistical processing really happened for "signature strikes." They're certainly sold as such but I think technically any drone strike against someone whose name the US doesn't know, but thinks is an insurgent, is a "signature strike." And "looking like an insurgent" probably often meant "a drone operator saw some guys with guns in cars," not artificial intelligence.
That doesn't mean it doesn't happen, Israel seems to have used machine learning for targeting in Gaza. I just don't know if the US was ever shown to have done so.
posted by BungaDunga at 11:07 AM on December 31 [2 favorites]
I'm not sure how much statistical processing really happened for "signature strikes." They're certainly sold as such but I think technically any drone strike against someone whose name the US doesn't know, but thinks is an insurgent, is a "signature strike." And "looking like an insurgent" probably often meant "a drone operator saw some guys with guns in cars," not artificial intelligence.
That doesn't mean it doesn't happen, Israel seems to have used machine learning for targeting in Gaza. I just don't know if the US was ever shown to have done so.
posted by BungaDunga at 11:07 AM on December 31 [2 favorites]
How do you verify that the LLM output is correct?
Is the LLM output as good as if you'd written it by-hand?
If not, how much worse is it? Is it worth it?
Just as an example, I've used it in situations where the correct output is easy to verify because I've already written tests or it just needs to be good enough to handle a constrained universe of inputs (say, a particular set of files). If it chokes on other files that doesn't matter, it's a one-off. For that reason the code quality doesn't matter so much either, but it produces basically unobjectionable Python most of the time.
Another one is for generating plausible test data that more or less obeys a schema and looks vaguely right.
I used it to translate a C++ utility directly to Python, mostly for curiosity's sake- I was interested to see how performant the Python would be. And it was too slow so we didn't use it. But it was much easier and more enjoyable to use an LLM to produce a nearly-correct translation to test on than writing it all out by hand.
posted by BungaDunga at 11:19 AM on December 31 [2 favorites]
Is the LLM output as good as if you'd written it by-hand?
If not, how much worse is it? Is it worth it?
Just as an example, I've used it in situations where the correct output is easy to verify because I've already written tests or it just needs to be good enough to handle a constrained universe of inputs (say, a particular set of files). If it chokes on other files that doesn't matter, it's a one-off. For that reason the code quality doesn't matter so much either, but it produces basically unobjectionable Python most of the time.
Another one is for generating plausible test data that more or less obeys a schema and looks vaguely right.
I used it to translate a C++ utility directly to Python, mostly for curiosity's sake- I was interested to see how performant the Python would be. And it was too slow so we didn't use it. But it was much easier and more enjoyable to use an LLM to produce a nearly-correct translation to test on than writing it all out by hand.
posted by BungaDunga at 11:19 AM on December 31 [2 favorites]
I teach English 101, basic composition, at a community college. You have no idea how much ChatGPT has changed things for writing teachers.
One of my colleagues, who teaches what they used to call ESL, had an anecdote. Her husband is a contract worker for Microsoft & was in a meeting where they were hyping up how much teachers are going to love gAI. He spoiled that little party by telling them his wife is a teacher & she hates it. Everyone at the meeting was flabbergasted. They had no idea there was any potential downside to the product.
Despite all the talking I do to my students, they only learn to write by writing. I always tell them writing is more like a physical activity than an intellectual one. You don’t learn how to dribble in a lecture hall. You have to put the ball in your hand & bounce it.
The true value I add as a writing teacher is as a reader, “being present” to their writing, making them “feel seen.” Coming from public schools, their previous writing teachers had 5 sections of 30+ students, so often I am the first person ever to respond in detail to what they’ve written.
Getting to know a student through their writing, treating them as a writer, is one of the pleasures of my job. So what am I supposed to do with LLM generated work? When I read it, what am I being present to?
Now, I am not a luddite. ChatGPT has actually been a great help for me in generating sample writing (for example, paragraphs with & without thesis sentences). My own writing is too recondite in subject & style, & the blandness of the machine writing fits the requirements nicely. Also I am an incredibly slow writer. ChatGPT spits something out in seconds, & it takes me less than 5 minutes to edit it.
But I already know how to write. You can’t edit if you don’t know how to write. & my fear is for the K-12 levels: how will they learn how to write if they never have to?
posted by warburtonave at 12:15 PM on December 31 [11 favorites]
One of my colleagues, who teaches what they used to call ESL, had an anecdote. Her husband is a contract worker for Microsoft & was in a meeting where they were hyping up how much teachers are going to love gAI. He spoiled that little party by telling them his wife is a teacher & she hates it. Everyone at the meeting was flabbergasted. They had no idea there was any potential downside to the product.
Despite all the talking I do to my students, they only learn to write by writing. I always tell them writing is more like a physical activity than an intellectual one. You don’t learn how to dribble in a lecture hall. You have to put the ball in your hand & bounce it.
The true value I add as a writing teacher is as a reader, “being present” to their writing, making them “feel seen.” Coming from public schools, their previous writing teachers had 5 sections of 30+ students, so often I am the first person ever to respond in detail to what they’ve written.
Getting to know a student through their writing, treating them as a writer, is one of the pleasures of my job. So what am I supposed to do with LLM generated work? When I read it, what am I being present to?
Now, I am not a luddite. ChatGPT has actually been a great help for me in generating sample writing (for example, paragraphs with & without thesis sentences). My own writing is too recondite in subject & style, & the blandness of the machine writing fits the requirements nicely. Also I am an incredibly slow writer. ChatGPT spits something out in seconds, & it takes me less than 5 minutes to edit it.
But I already know how to write. You can’t edit if you don’t know how to write. & my fear is for the K-12 levels: how will they learn how to write if they never have to?
posted by warburtonave at 12:15 PM on December 31 [11 favorites]
Just as an example, I've used it in situations where the correct output is easy to verify because I've already written tests or it just needs to be good enough to handle a constrained universe of inputs (say, a particular set of files). If it chokes on other files that doesn't matter, it's a one-off. For that reason the code quality doesn't matter so much either, but it produces basically unobjectionable Python most of the time.That has also been my experience. People miss the existence of utility code use cases that are not what we usually think of as software engineering, and they gravitate toward the extremes when discussing the risk, whether that's economic or moral.
I used to be a toolsmith for a platform engineering group and I would have benefitted a lot from something like Copilot or just ChatGPT. I was never going to burn down my backlog, and for every interesting big project I could apply myself to, there were several new paper cuts a week that just needed to be triaged and restored to working order across a range of languages (Ruby, Clojure, assorted shell dialects, and a hodge hodge of front-end frameworks, CMSes, and infrastructure management tools e.g. Puppet and Ansible).
I've since gone back and tried out a few things I spent hours on writing from scratch — some of the systems management scripts I wrote, some utility scripts, some Puppet code I wrote — and each time have gotten results that would have been great for what I was trying to do in much less time. Looking back, the coding I could do when I got that job felt like a super power because I had majored in philosophy and got my professional start in online editorial. A few years in tech taught me there was a whole world of people who could exceed me in programming skill, but the guy who hired me understood me to be more useful for being an autodidact and a persistent problem solver who'd take the whole system into account before pushing shit into prod.
These days I'm a senior director sitting on an AI governance committee, and AI hype annoys the shit out of me, but I do my best with my partners in legal, risk, and security to filter out the noise and understand the impacts of whatever people are trying to bring into the environment. Some of it is snake oil, some of it is obnoxious, some of it it is genuinely useful. We all agree the worst source of annoyance is that every SaaS vendor is claiming to have AI features, which spins up a whole particular review process on the risk team only to learn it's, like, spell checker AI, not "ravening fucking data devouring hellbeast with a liberally permissioned API token for all your subprocessors" AI. The only satisfaction I get from that is hopping back on the vendor call to say "oh, yeah, nothing more needed for security review -- we have you down in the 'marketing checklist AI' column."
I did sit on a scoping call with one account executive who was very up front that they couldn't allow a lengthy trial because every one is a money suck for them. So I think when I'm operating at peak objectivity I'm probably a Zitronist. I just don't care to overstate other issues or risks to believe it's unsustainable. With three imploded companies from three different hype cycles in my background, I don't like its chances.
posted by A forgotten .plan file at 2:26 PM on December 31 [9 favorites]
”This is literally a DDOS on the whole internet”
posted by ursus_comiter at 5:43 PM on December 31 [2 favorites]
posted by ursus_comiter at 5:43 PM on December 31 [2 favorites]
my fear is for the K-12 levels: how will they learn how to write if they never have to?
Likewise think.
No, seriously, hear me out. The best and most reliable way that I know of to get my head wrapped thoroughly around a topic, and locate the yawning conceptual gaps I'm currently not aware of, is to try to explain it to somebody else. And there is no more reliable tool for pounding those explanations into shape than getting them down in writing and making that writing make sense.
Taking the time to get an unruly herd of thoughts corralled into this one-dimensional symbolic chicken-scratch is indispensable for getting things properly clear in my own mind, and I don't believe that's just me.
A great deal of explanation rests on the search for appropriate words and phrases, and that search is work. Launching into an explanation without having done that work quickly makes it collapse into an unintelligible screed about how I can't get the thing to work because I can't find the thing I need, you know, the thing that makes the thing do the thing.
I often find myself feeling powerless to help little ms flabdablet, now 19 years old, overcome some difficulty that's troubling her because her requests for help are so often couched in that kind of language. And it's not just her, her friends do it as well. But about four times in five it turns out that the only help she actually needs from me is my insistence on understanding what the trouble actually is, because that forces her to organize her thoughts enough to explain it to me, at which point she will perceive what needs to be done without my having had to offer any advice whatsoever.
The increasing social prevalence of anxiety is no surprise to me at all. I'd be horribly anxious most of the time too, if I had to navigate our increasingly complex world while mired in as much mental sludge as I routinely see our young ones struggling to slog through. Also no surprise to me that journalling works as well as it does to help counter that anxiety.
Writing is good exercise, and we obviate the need for it at our peril.
posted by flabdablet at 7:59 PM on December 31 [7 favorites]
Likewise think.
No, seriously, hear me out. The best and most reliable way that I know of to get my head wrapped thoroughly around a topic, and locate the yawning conceptual gaps I'm currently not aware of, is to try to explain it to somebody else. And there is no more reliable tool for pounding those explanations into shape than getting them down in writing and making that writing make sense.
Taking the time to get an unruly herd of thoughts corralled into this one-dimensional symbolic chicken-scratch is indispensable for getting things properly clear in my own mind, and I don't believe that's just me.
A great deal of explanation rests on the search for appropriate words and phrases, and that search is work. Launching into an explanation without having done that work quickly makes it collapse into an unintelligible screed about how I can't get the thing to work because I can't find the thing I need, you know, the thing that makes the thing do the thing.
I often find myself feeling powerless to help little ms flabdablet, now 19 years old, overcome some difficulty that's troubling her because her requests for help are so often couched in that kind of language. And it's not just her, her friends do it as well. But about four times in five it turns out that the only help she actually needs from me is my insistence on understanding what the trouble actually is, because that forces her to organize her thoughts enough to explain it to me, at which point she will perceive what needs to be done without my having had to offer any advice whatsoever.
The increasing social prevalence of anxiety is no surprise to me at all. I'd be horribly anxious most of the time too, if I had to navigate our increasingly complex world while mired in as much mental sludge as I routinely see our young ones struggling to slog through. Also no surprise to me that journalling works as well as it does to help counter that anxiety.
Writing is good exercise, and we obviate the need for it at our peril.
posted by flabdablet at 7:59 PM on December 31 [7 favorites]
why do I need a carpenter when we have power tools
Exactly! Those thumbs aren't going to sever themselves, now are they?
posted by flabdablet at 8:42 PM on December 31 [4 favorites]
Exactly! Those thumbs aren't going to sever themselves, now are they?
posted by flabdablet at 8:42 PM on December 31 [4 favorites]
”This is literally a DDOS on the whole internet”
Gonna need a lot more electricity to get that happening properly at scale.
posted by flabdablet at 8:45 PM on December 31 [1 favorite]
Gonna need a lot more electricity to get that happening properly at scale.
posted by flabdablet at 8:45 PM on December 31 [1 favorite]
Stop forcing AI into everything that worked perfectly well without AI. Nobody asked for it, and most people don't want it.
posted by thoughtful_jester at 6:20 PM on January 1 [1 favorite]
posted by thoughtful_jester at 6:20 PM on January 1 [1 favorite]
« Older The Most Important Machine in the World | 🏛️ Senatus vs. Populusque Romanus ✊🏼 Newer »
posted by mittens at 12:03 PM on December 30 [5 favorites]