My mental shorthand for the low-stakes version of the "trust" legal example is a food one. DIfferent languages divide up the world along different axes sometimes. Take mandarin "jiaozi' and "baozi," usually translated as "dumplings" and "buns." And usually that's true. But the English distinction is mostly about the type of dough used for the skin, while the Mandarin distinction is mostly about the shape, how it's formed and closed.
LLMs have the advantage, in principle, of having read every word ever written and put online (or every word ever spoken and recorded and put online). That would give them the maximum possible amount of context across all languages for choosing the right word. But in practice, we're not at that point, which may or may not happen soon.
I would like to see AI devalue organizational skills and effort. The personality tests say I’m highly creative and analytical. I’m not structured and I often miss or forget things because of it. An AI handling these details with minimal effort on my part would be great.
Tim's point that AI will likely replace low-end translators while the specialty/high-end ones will not be endangered reflects the same point the always entertaining Jeff Maurer made about AI writing scripts (https://imightbewrong.substack.com/p/ai-spells-doom-for-incompetent-hacks). AI will probably never be able to replace a Shakespeare (or a Billy Wilder) but it will probably produce schlock at least as good as the mediocre writers in Hollywood -- which is the large majority of them. Tough luck for you hacks, but you probably shouldn't be a writer in the first place.
(Speaking of Billy Wilder, I recently saw his grave in the tiny celebrity-filled cemetery tucked away off Wilshire Blvd here in LA, with the epigraph "I'm a writer but then nobody's perfect" (https://commons.wikimedia.org/wiki/File:Billy_Wilder_Grave.jpg) For lovers of "Some Like It Hot," what a great sendoff.)
This is a great guest post! I know some of the initial guest posts got a rocky reception because they were lower quality hot takes, but I would love to see more of this kind of analytic post.
The taxi cab analogy would seem to undermine the “don’t worry too much” thesis--a lot of cabbies ended up committing suicide after the arrival of smartphone apps! Taxi medallion debt might be considered a special case, but a lot of knowledge workers enter the market deeply in debt from college loans.
The thing about these kinds of technologies is that they help people with fewer skills and they hurt people with specialized skills. Unfortunately, we know that most knowledge workers have an above average skill level relative to their peers in knowledge work, so the adjustment is going to be quite rough.
Accepting arguendo that the number of cabbie suicides triggered by competition from Uber, etc. qualifies as "lots" (this is one of these things where it is completely unclear what the "normal rate" is, i.e., how many cabbies commit suicide over a comparable length of time when ridesharing smartphone apps weren't available, so we can't actually say whether there's been a statistically significant increase), it appears that the victims were overwhelmingly, if not exclusively, immigrants to the US: https://www.nytimes.com/2018/12/02/nyregion/taxi-drivers-suicide-nyc.html
Given the paucity of native-born Americans committing suicide in this context, it seems reasonable to think that the suicides were driven as much by a lack of understanding of the US bankruptcy system, US tax system, and/or possibly fear that loss of their business might not just have financial consequences, but lead to them being deported if they aren't naturalized US citizens. I'm skeptical the same pressures apply to most knowledge workers, especially when income-based repayment plans for student loans and public service loan deferrals/forgiveness are already an established thing.
I don't want to claim (and don't think I did claim) that nobody is going to face hardship as a result of AI progress. Economic change is always stressful for people and I don't want to minimize that. I just think a lot of people are overestimating its likely severity.
What I wrote was "workers in other industries don’t need to worry about AI taking over their jobs overnight." Certainly I think it's reasonable for workers to worry a little bit and to plan accordingly. And yes, I think people should think twice before accumulating a lot of debt, though I don't think the value of a college degree is going to crash the way a taxi medallion's value did.
Agree that most college graduates shouldn’t worry, though I do wonder about translators. The AI shock may not happen all at once, but the tech also keeps getting better. An industry can appear to be declining slowly and then experience something like a “cliff moment.” Probably in this case it will take care of itself through attrition--hard to imagine many people are entering college today with the express ambition of becoming a translator.
The solution is not to manage the decline, the solution is to ban all AI stronger than GPT-3 in order to preserve human worth, dignity, and control of the future. I have never seen a satisfactory explanation for how we are to control something orders of magnitude smarter than us.
I'm not sure whether Wendigo will outright say it, but the idea is that defection would be a legit casus belli. (I say this with doomer sympathies, but not along the economics disruption axis).
Just want to note that we’re already seeing some of the real-world benefits of cheap translation: it’s frequently used on reality shows like “90-Day Fiancé”. It’s made international matchmaking a solid order of magnitude easier.
So the thesis is that this is merely the next step of automation. Nothing more nothing less. If so it’s still pretty bad for the workers imo, for reasons others laid out. However I’d like to point out something else. What automation does- I think- is hollow out the middle ground. It takes a service previously available to the middle class, or to the average person in moderate amounts, and converts it to a shittier but good-enough product available in abundance, whereas the previous, higher-quality product which was “normal” for middle class people becomes a status symbol of the super rich. There are pros and cons for this process but we certainly lose out something, esp the “we” who are in the middle and upper-middle (but not super rich) strata of society.
What do you mean by "previously available to the middle class?" I don't see any reason that the introduction of AI would dramatically raise the cost of an old-fashioned translation. It just makes good-enough AI cheap enough that the high cost of a human translation is no longer worth it for most purposes. But people can still pay it if they want to.
Isn’t that what follows from your own prediction? If translation will die as a massive industry it will be much harder for the middle class person to find a human translator at a decent price, especially one who knows what they’re about. It’s easy to buy pretty good furniture from ikea but I bet it’s far more expensive to have custom made high quality furniture by a highly skilled carpenter than it was a century ago. Cooking for yourself at home and eating out have both become cheaper thanks to technology but how many middle class people can afford their own full time cook? Having a private secretary, fully qualified to write your correspondences etc used to be standard for professionals. Doubtless far more people have access to chat gpt and phone calendars than ever had to secretaries, but a highly skilled, human, “personal assistant” (or however they’re called nowadays), which is still obviously better than all the tech put together, is on the trend to becoming the mark of the upper echelons Etc
P.S.
I’m not nostalgic for the past. I think that- thus far at least- rise in productivity has been a *net* good on the long term. All I’m pointing out is that even if it’s a net good for society, some of us are permanently losing out on some things, and that’s true even from the consumer’s perspective, not just the worker’s.
"I bet it’s far more expensive to have custom made high quality furniture by a highly skilled carpenter than it was a century ago."
Is this actually true in real terms? I think the difference is that 100 years ago you could buy really crappy products or super expensive well made products, but there wasn't very much decent stuff in the middle.
You can still buy the super expensive stuff now, but most people don't want to. Similar to other parts of fashion, people now often will want to switch to new looks or designs after a decade or so instead of buying super expensive furniture and having it for life.
It's more expensive to hire a carpenter than 100 years ago because wages in general have gone up, and so carpenters' incomes have gone up along with everyone else's. Ditto for cooks, secretaries, etc. But this isn't because living standards have gone down. Quite the contrary. The "middle class" of 100 years ago that could afford maids and cooks were in the top 5 percent if not the top 1 percent of the income distribution. They were able to afford this kind of labor because most people had very low wages and so there was a lot of surplus labor around.
The people we call the middle class today are a completely different slice of the income distribution, from say the 40th to 90th percentile. People in that portion of the income distribution in 1923 would not have had servants. They just had a much worse standard of living due to the lack of washing machines, vaccuum cleaners, Ikea, etc.
So yes, rising wages have made life worse in some ways for people in the 95th to 99th percentile of the income distribution because they have to get by with fewer servants. But it's been good for people in the bottom 90 percent of the income distribution who couldn't afford servants in 1923 and can't afford them now. Overall it seems like clear progress to me.
In 1920 only 22 percent of white people aged 25 to 29 had high school diplomas and only 4 percent had a college degree. So people look back, read that that it was common for people with college degrees to have servants, and conclude that living standards have fallen. But in reality people who had college degrees in 1923 were a totally different slice of the income distribution than people with college degrees today. Having a college degree in 1920 made you a member of a tiny elite, and they could easily afford servants because pay for non-college graduates was very low.
Today, many more people have college degrees (35 percent of all young American adults) and median wages are a lot higher. So unsurprisingly most college graduates today can't afford to hire servants. That's because prosperity is far more widely shared than it was a century ago.
You’d notice that you didn’t answer my question, unless you assume that *only* people with college degrees had servants? Anyway you keep harking back to an uncontested point (aka straw man). I ask again, did you read my ps? We’re in agreement that society today is better off. That’s hardly the point.
Most, or perhaps all of your examples are cases where the middle class was priced out by Baumol's cost disease, which is triggered by productivity improvements in _other_industries.
But, in order for translation to die out, that means AI translation has to get a lot better.
It won't be "you have to settle for current gen AI translation because you can't afford one of the translator specialists that you previously could have afforded", it'll be some much better translation.
Basically, when you say: " It takes a service previously available to the middle class, or to the average person in moderate amounts, and converts it to a shittier but good-enough product available in abundance"
If it was good enough to kill most of the translation industry then I'm not convinced it will be meaningfully worse than what you could afford now, so I think the word "shittier" is wrong there.
I found listening to some of the ai generated audio instructive. I get that the use of AI here might have been a sort of “meta” demonstration but as such it really highlights the shortcomings. It’s understandable but pretty bad. Hearing eg “you-ber” (Uber) is distracting, and the reading is monotone, making comprehension marginally harder and satsifacito somewhat lessened. The whole thing is less than 13 min at normal speed. Even allowing for some editing and error correction, it shouldn’t have taken the author more than say 30min to record himself actually reading it. I think it would have been worth the effort to do so in terms of the impression left on the listener. I imagine this kind of “penalty” in terms of user experience can be modeled by economists. I wonder whether that will limit the full takeover if AI or if we are merely going to gradually lower our standards and expectations of quality from a whole bunch of services.
Thanks for the feedback. I think it would take me more than 30 minutes to produce a high-quality recording because every time I misspeak (which I do fairly often) I have to go back and re-record a sentence. Also a significant portion of the prep time was finding and copying over the source quotes. Plus it was my first time using the software. I expect that once I know the tools well I'll be able to do it in about an hour.
Still I agree about the audio quality. I also noticed the You-ber problem. The question is whether we'll continue to see quality improve. This stuff is a lot better than it was five years ago, so in five more years perhaps You-ber type gaffes will be a thing of the past.
I think that’s the key question. And also I should say that I appreciate the “meta” quality of using the ai for this particular piece, but generally speaking my hunch is that at least for shorter pieces like this actually recording yourself might be the better choice- at least for now.
Coincidentally, I just finished a novel in which the main character is a simultaneous translator (sounds like a stressful job!), mostly for medical presentations. This article makes me realize that the type of human translators that are still needed will be very skilled and specialized. But if real wages are dropping in the industry, will the skill level naturally fall? Or will there be niche jobs in the field that pay a lot?
This also makes one wonder that if all the low-level translation jobs are filled by AI, how does one become a very skilled and specialized translator. Isn't there some kind of career ladder for reaching the higher levels?
The video game example demonstrates just how more serious makers have become at getting translations correct. It's a far cry from comparing the mistranslations of games I played as a kid in the 1980s and 1990s, a world in which phrases like "All your base are belong to us" were tolerated.
"If you put ‘trust’ in ChatGPT it's going to translate it to confianza,” Leon said. “But that's not what it means.” In reality, Leon says, there are 20 or 30 different ways to translate the legal concept of a trust to Spanish. Figuring out which meaning to use in any given sentence requires a sophisticated understanding of American law, Spanish law, and the context of the specific document she is translating."
Isn't context pretty much what AI does? Isn't it coming up with text based on the statistics of situations that the AI has trained on? If you trained the AI on Spanish and US legal documents that have been translated into the other system, wouldn't it get really good at those translations itself?
Apparently not! It's possible that general ChatGPT can't do it but a fine-tuned model specifically focused on legal translation could do a better job. But often you need information beyond the four corners of the document—information you'd get by interviewing the client or reviewing other documents. AI might be able to do that kind of thing eventually but it'll take a while to develop it.
As someone who works overseas regularly, I can’t wait until AI real time translation happens.
But... I wouldn’t be optimistic for these translators. Every time they edit a machine translation, the AI is learning that much more. They think they are editing, but really they are just teaching.
Back to real time actual translation. It is going to make the dating game so much easier. Everyone’s potential partners will expand by billions.
Re dating, since most people meet via dating apps these days and a lot of initial communication starts with exchanging text messages, it won't be long before a variant of ChatGPT* will be generating witty, charming banter for you to copy and paste into the text box.
* Yes, of course the trademarked name for this will be "Cyrano."
"since most people meet via dating apps these days..."
I don't think this is correct. Pew says that only 53% of under people under 30 report having EVER USED a dating site or app. I suspect the majority of ways people meet is still someone you knew from school, work, church, met through mutual friends, etc.
As someone who has been married for a while, I've not used a dating app or site, but anecdotally I know plenty of people who have and success is "mixed" at best. Even for people using such tools, many of them ended up with someone they met through other ways.
There’s a case to be made that changing the dating pool from the few dozen people you see in bars and work and through friends, to everyone within a ten mile gps radius on a dating app, may have made people’s dating experience worse. When you have more people to choose between, your standards yet higher - and perhaps more importantly, *their* standards get higher too, so everyone takes far longer to find a reasonable match, and ends up with one not much better. Expanding to the whole world could make that dynamic that much worse.
Being on Hinge right now, I think this is very accurate.
Dating apps also expose more of us quickly. Things that shouldn't matter (or that might matter but can be overcome by other positive attributes) are now front and center; examples include job title, hometown, astrological sign, and marijuana use. This encourages snap judgments.
This sounds to me similar to the internet optimism of the early 2000s. By contrast I think it might exacerbate the current problems of the internet. If we can all talk we can all fight.
That’s something I confidently predicted (in a discussion forum pretty much like this one) twenty years ago. Responses to the idea were almost entirely skeptical.
It’ll happen one day. But imagine the potential bumps in the road when a less-than-perfect translator chooses the wrong word or idiom!
"Every time they edit a machine translation, the AI is learning that much more."
What's the mechanism by which the data is being fed back into the machine? For legal docs/etc, or video game text, we don't put these out on public networks.
I mean, it'll get better over time anyway but unless you're feeding back the translated text with the source text how is it improving based on _your_ translation?
The pros often use integrated software where the human corrections are made inside the translation UI. Therefore the company that makes the software can use these corrections to improve its algorithm.
Uber is highly subsidized by investors and runs at a deep, deep loss. The prices they charged were never realistic.
LLMs like ChatGPT are also highly subsidized. The training is extremely expensive and even the per-query costs are quite high. Maybe the chips will eventually get cheap enough to break even on ads, but I doubt it.
I don't really agree with this. I don't know if these neural networks are breaking even right now, but I don't think there's much doubt that Moore's law will make them profitable at scale. Amazon Web Services has been obscenely profitable for several years now despite steadily cutting their prices.
Moore’s law is going to run out of steam in the middle of this decade (and may already have run out of steam—Nvidia’s Jensen Huang, who would probably know, pronounced it dead in 2022) because of limits imposed by physics— you can only cram so many circuits into a silicon wafer. We’ll still get improvements in compute from improved parallelization and specialized chip designs optimized for performing specific tasks, but the late 20th century’s exponential improvements can’t scale.
I think that with the right set of optimizations, running LLM instances will be cost-effective for a lot of tasks, but the “just 10x the number of parameters” strategy for improving performance will stop being viable (because of both training and operating costs), so we’ll probably see a ceiling on their sophistication until there’s some sort of major paradigm shift. At the moment, I think LLMs are on track to be a useful and commercially important but not world-shattering tech.
All way beyond my pay grade, but I keep reading claims we'll be able squeeze more time out of Moore's law because of better software and better materials. Also quantum computing?
QC is completely irrelevant here, anyone who says this is BSing. Software is sort of orthogonal to Moore - the point of Moore* is you don't need to pay programmers to optimize things, everything gets better automatically
AWS is profitable because web dev is insanely cheap. I run a medium sized news site and CPU and bandwidth are free. All the money we spend is for database hosting and image resizing. So far at least, LLMs take orders of magnitude more computing power. A good person to talk to about this is Tim Bray. He used to be an engineering VP at AWS. He had some throw away line on Mastodon about how you can feel all the compute being burned by ChatGPT. The fact is that even for ChatGPT 3.5, responses are super-slow. The reason that responses are slow is that there is a huge and economically unviable amount of compute being thrown at them. In the long run, yes, Moore’s law will probably make it viable, but if the minimum LLM experience people expect is even more expensive, it might all just wash out. As it is, the CPU used by bigger LLMs is scaling up faster than Moore’s law. Again, everything can change, but it’s also wrong to just assume that it will all take care of itself.
My understanding was that it takes enormous computing power to train LLM, but not so much to run it after the fact. So it could be unprofitable to train it initially, but then its quite reasonable to use it.
I'm sure there is more to it than that, but would be interested to see you lay out more details.
The training vs inference very much depends on the nature of the product itself, and how popular it is. Training is expensive, but inference does not yet have "zero marginal cost" economics. That makes a big difference when compared to the traditional cloud model.
In Nov 2022, MidJourney was running into issues with cloud capacity to keep up. 90% of their cloud costs were from inference, and only 10% from training. On top of gating usage by price, they still had to (and continue to) rate-limit usage of premium users.
In the 8 months since then, both the number of "users" and "online now" (both include lurkers) has increased 4x -- to 16.65M and 1.44M respectively. Despite the increase in traffic, I have heard from some longtime users that user experience, especially wait times, have improved. To me this indicates that their inference costs are going down. But they are still at the order of 100x smaller than Twitter and 1000x smaller than FB.
Here's what Bard says about operating costs for ChatGPT.
If we assume that ChatGPT uses 8 GPUs to operate, and that each GPU costs $3 an hour, then each word generated on ChatGPT costs $0.0003. At least 8 GPUs are in use to operate on a single ChatGPT, and each question typically generates around 30 words. This means that the per question cost of ChatGPT is around $0.009, or 9 cents.
However, the actual per question cost may be lower than this. For example, if ChatGPT is able to reuse some of the computing resources from previous questions, then the per question cost will be lower. Additionally, if ChatGPT is able to be more efficient in its use of computing resources, then the per question cost will also be lower.
My mental shorthand for the low-stakes version of the "trust" legal example is a food one. DIfferent languages divide up the world along different axes sometimes. Take mandarin "jiaozi' and "baozi," usually translated as "dumplings" and "buns." And usually that's true. But the English distinction is mostly about the type of dough used for the skin, while the Mandarin distinction is mostly about the shape, how it's formed and closed.
LLMs have the advantage, in principle, of having read every word ever written and put online (or every word ever spoken and recorded and put online). That would give them the maximum possible amount of context across all languages for choosing the right word. But in practice, we're not at that point, which may or may not happen soon.
Can we get this audio on the Slow Boring feed? I can't find how to get it to my podcaster
I would like to see AI devalue organizational skills and effort. The personality tests say I’m highly creative and analytical. I’m not structured and I often miss or forget things because of it. An AI handling these details with minimal effort on my part would be great.
Tim's point that AI will likely replace low-end translators while the specialty/high-end ones will not be endangered reflects the same point the always entertaining Jeff Maurer made about AI writing scripts (https://imightbewrong.substack.com/p/ai-spells-doom-for-incompetent-hacks). AI will probably never be able to replace a Shakespeare (or a Billy Wilder) but it will probably produce schlock at least as good as the mediocre writers in Hollywood -- which is the large majority of them. Tough luck for you hacks, but you probably shouldn't be a writer in the first place.
(Speaking of Billy Wilder, I recently saw his grave in the tiny celebrity-filled cemetery tucked away off Wilshire Blvd here in LA, with the epigraph "I'm a writer but then nobody's perfect" (https://commons.wikimedia.org/wiki/File:Billy_Wilder_Grave.jpg) For lovers of "Some Like It Hot," what a great sendoff.)
This is a great guest post! I know some of the initial guest posts got a rocky reception because they were lower quality hot takes, but I would love to see more of this kind of analytic post.
Tim has a long history of writing high quality articles, glad to see Matt introduce him to all Slow Borers with a guest article.
Thank you both!
The taxi cab analogy would seem to undermine the “don’t worry too much” thesis--a lot of cabbies ended up committing suicide after the arrival of smartphone apps! Taxi medallion debt might be considered a special case, but a lot of knowledge workers enter the market deeply in debt from college loans.
The thing about these kinds of technologies is that they help people with fewer skills and they hurt people with specialized skills. Unfortunately, we know that most knowledge workers have an above average skill level relative to their peers in knowledge work, so the adjustment is going to be quite rough.
Accepting arguendo that the number of cabbie suicides triggered by competition from Uber, etc. qualifies as "lots" (this is one of these things where it is completely unclear what the "normal rate" is, i.e., how many cabbies commit suicide over a comparable length of time when ridesharing smartphone apps weren't available, so we can't actually say whether there's been a statistically significant increase), it appears that the victims were overwhelmingly, if not exclusively, immigrants to the US: https://www.nytimes.com/2018/12/02/nyregion/taxi-drivers-suicide-nyc.html
Given the paucity of native-born Americans committing suicide in this context, it seems reasonable to think that the suicides were driven as much by a lack of understanding of the US bankruptcy system, US tax system, and/or possibly fear that loss of their business might not just have financial consequences, but lead to them being deported if they aren't naturalized US citizens. I'm skeptical the same pressures apply to most knowledge workers, especially when income-based repayment plans for student loans and public service loan deferrals/forgiveness are already an established thing.
I don't want to claim (and don't think I did claim) that nobody is going to face hardship as a result of AI progress. Economic change is always stressful for people and I don't want to minimize that. I just think a lot of people are overestimating its likely severity.
What I wrote was "workers in other industries don’t need to worry about AI taking over their jobs overnight." Certainly I think it's reasonable for workers to worry a little bit and to plan accordingly. And yes, I think people should think twice before accumulating a lot of debt, though I don't think the value of a college degree is going to crash the way a taxi medallion's value did.
Agree that most college graduates shouldn’t worry, though I do wonder about translators. The AI shock may not happen all at once, but the tech also keeps getting better. An industry can appear to be declining slowly and then experience something like a “cliff moment.” Probably in this case it will take care of itself through attrition--hard to imagine many people are entering college today with the express ambition of becoming a translator.
The solution is not to manage the decline, the solution is to ban all AI stronger than GPT-3 in order to preserve human worth, dignity, and control of the future. I have never seen a satisfactory explanation for how we are to control something orders of magnitude smarter than us.
I'm not sure whether Wendigo will outright say it, but the idea is that defection would be a legit casus belli. (I say this with doomer sympathies, but not along the economics disruption axis).
Edit - spacing.
Yes, I take that view.
I think doing it in secret would be more challenging than you are accounting for, but I agree.
The electricity signatures would be obvious, as would the outputs.
Just want to note that we’re already seeing some of the real-world benefits of cheap translation: it’s frequently used on reality shows like “90-Day Fiancé”. It’s made international matchmaking a solid order of magnitude easier.
So the thesis is that this is merely the next step of automation. Nothing more nothing less. If so it’s still pretty bad for the workers imo, for reasons others laid out. However I’d like to point out something else. What automation does- I think- is hollow out the middle ground. It takes a service previously available to the middle class, or to the average person in moderate amounts, and converts it to a shittier but good-enough product available in abundance, whereas the previous, higher-quality product which was “normal” for middle class people becomes a status symbol of the super rich. There are pros and cons for this process but we certainly lose out something, esp the “we” who are in the middle and upper-middle (but not super rich) strata of society.
What do you mean by "previously available to the middle class?" I don't see any reason that the introduction of AI would dramatically raise the cost of an old-fashioned translation. It just makes good-enough AI cheap enough that the high cost of a human translation is no longer worth it for most purposes. But people can still pay it if they want to.
Isn’t that what follows from your own prediction? If translation will die as a massive industry it will be much harder for the middle class person to find a human translator at a decent price, especially one who knows what they’re about. It’s easy to buy pretty good furniture from ikea but I bet it’s far more expensive to have custom made high quality furniture by a highly skilled carpenter than it was a century ago. Cooking for yourself at home and eating out have both become cheaper thanks to technology but how many middle class people can afford their own full time cook? Having a private secretary, fully qualified to write your correspondences etc used to be standard for professionals. Doubtless far more people have access to chat gpt and phone calendars than ever had to secretaries, but a highly skilled, human, “personal assistant” (or however they’re called nowadays), which is still obviously better than all the tech put together, is on the trend to becoming the mark of the upper echelons Etc
P.S.
I’m not nostalgic for the past. I think that- thus far at least- rise in productivity has been a *net* good on the long term. All I’m pointing out is that even if it’s a net good for society, some of us are permanently losing out on some things, and that’s true even from the consumer’s perspective, not just the worker’s.
"I bet it’s far more expensive to have custom made high quality furniture by a highly skilled carpenter than it was a century ago."
Is this actually true in real terms? I think the difference is that 100 years ago you could buy really crappy products or super expensive well made products, but there wasn't very much decent stuff in the middle.
You can still buy the super expensive stuff now, but most people don't want to. Similar to other parts of fashion, people now often will want to switch to new looks or designs after a decade or so instead of buying super expensive furniture and having it for life.
It's more expensive to hire a carpenter than 100 years ago because wages in general have gone up, and so carpenters' incomes have gone up along with everyone else's. Ditto for cooks, secretaries, etc. But this isn't because living standards have gone down. Quite the contrary. The "middle class" of 100 years ago that could afford maids and cooks were in the top 5 percent if not the top 1 percent of the income distribution. They were able to afford this kind of labor because most people had very low wages and so there was a lot of surplus labor around.
The people we call the middle class today are a completely different slice of the income distribution, from say the 40th to 90th percentile. People in that portion of the income distribution in 1923 would not have had servants. They just had a much worse standard of living due to the lack of washing machines, vaccuum cleaners, Ikea, etc.
So yes, rising wages have made life worse in some ways for people in the 95th to 99th percentile of the income distribution because they have to get by with fewer servants. But it's been good for people in the bottom 90 percent of the income distribution who couldn't afford servants in 1923 and can't afford them now. Overall it seems like clear progress to me.
Did you read my ps before responding ? Also- I’d like some data on the percentages in 1923 if you have it. Thanks!
In 1920 only 22 percent of white people aged 25 to 29 had high school diplomas and only 4 percent had a college degree. So people look back, read that that it was common for people with college degrees to have servants, and conclude that living standards have fallen. But in reality people who had college degrees in 1923 were a totally different slice of the income distribution than people with college degrees today. Having a college degree in 1920 made you a member of a tiny elite, and they could easily afford servants because pay for non-college graduates was very low.
Today, many more people have college degrees (35 percent of all young American adults) and median wages are a lot higher. So unsurprisingly most college graduates today can't afford to hire servants. That's because prosperity is far more widely shared than it was a century ago.
You’d notice that you didn’t answer my question, unless you assume that *only* people with college degrees had servants? Anyway you keep harking back to an uncontested point (aka straw man). I ask again, did you read my ps? We’re in agreement that society today is better off. That’s hardly the point.
Most, or perhaps all of your examples are cases where the middle class was priced out by Baumol's cost disease, which is triggered by productivity improvements in _other_industries.
But, in order for translation to die out, that means AI translation has to get a lot better.
It won't be "you have to settle for current gen AI translation because you can't afford one of the translator specialists that you previously could have afforded", it'll be some much better translation.
Basically, when you say: " It takes a service previously available to the middle class, or to the average person in moderate amounts, and converts it to a shittier but good-enough product available in abundance"
If it was good enough to kill most of the translation industry then I'm not convinced it will be meaningfully worse than what you could afford now, so I think the word "shittier" is wrong there.
It will be “good enough” but not quite as good, and we’ll get used to it, like in so many cases.
I found listening to some of the ai generated audio instructive. I get that the use of AI here might have been a sort of “meta” demonstration but as such it really highlights the shortcomings. It’s understandable but pretty bad. Hearing eg “you-ber” (Uber) is distracting, and the reading is monotone, making comprehension marginally harder and satsifacito somewhat lessened. The whole thing is less than 13 min at normal speed. Even allowing for some editing and error correction, it shouldn’t have taken the author more than say 30min to record himself actually reading it. I think it would have been worth the effort to do so in terms of the impression left on the listener. I imagine this kind of “penalty” in terms of user experience can be modeled by economists. I wonder whether that will limit the full takeover if AI or if we are merely going to gradually lower our standards and expectations of quality from a whole bunch of services.
Thanks for the feedback. I think it would take me more than 30 minutes to produce a high-quality recording because every time I misspeak (which I do fairly often) I have to go back and re-record a sentence. Also a significant portion of the prep time was finding and copying over the source quotes. Plus it was my first time using the software. I expect that once I know the tools well I'll be able to do it in about an hour.
Still I agree about the audio quality. I also noticed the You-ber problem. The question is whether we'll continue to see quality improve. This stuff is a lot better than it was five years ago, so in five more years perhaps You-ber type gaffes will be a thing of the past.
I think that’s the key question. And also I should say that I appreciate the “meta” quality of using the ai for this particular piece, but generally speaking my hunch is that at least for shorter pieces like this actually recording yourself might be the better choice- at least for now.
Coincidentally, I just finished a novel in which the main character is a simultaneous translator (sounds like a stressful job!), mostly for medical presentations. This article makes me realize that the type of human translators that are still needed will be very skilled and specialized. But if real wages are dropping in the industry, will the skill level naturally fall? Or will there be niche jobs in the field that pay a lot?
This also makes one wonder that if all the low-level translation jobs are filled by AI, how does one become a very skilled and specialized translator. Isn't there some kind of career ladder for reaching the higher levels?
The video game example demonstrates just how more serious makers have become at getting translations correct. It's a far cry from comparing the mistranslations of games I played as a kid in the 1980s and 1990s, a world in which phrases like "All your base are belong to us" were tolerated.
"If you put ‘trust’ in ChatGPT it's going to translate it to confianza,” Leon said. “But that's not what it means.” In reality, Leon says, there are 20 or 30 different ways to translate the legal concept of a trust to Spanish. Figuring out which meaning to use in any given sentence requires a sophisticated understanding of American law, Spanish law, and the context of the specific document she is translating."
Isn't context pretty much what AI does? Isn't it coming up with text based on the statistics of situations that the AI has trained on? If you trained the AI on Spanish and US legal documents that have been translated into the other system, wouldn't it get really good at those translations itself?
Apparently not! It's possible that general ChatGPT can't do it but a fine-tuned model specifically focused on legal translation could do a better job. But often you need information beyond the four corners of the document—information you'd get by interviewing the client or reviewing other documents. AI might be able to do that kind of thing eventually but it'll take a while to develop it.
As someone who works overseas regularly, I can’t wait until AI real time translation happens.
But... I wouldn’t be optimistic for these translators. Every time they edit a machine translation, the AI is learning that much more. They think they are editing, but really they are just teaching.
Back to real time actual translation. It is going to make the dating game so much easier. Everyone’s potential partners will expand by billions.
Re dating, since most people meet via dating apps these days and a lot of initial communication starts with exchanging text messages, it won't be long before a variant of ChatGPT* will be generating witty, charming banter for you to copy and paste into the text box.
* Yes, of course the trademarked name for this will be "Cyrano."
"since most people meet via dating apps these days..."
I don't think this is correct. Pew says that only 53% of under people under 30 report having EVER USED a dating site or app. I suspect the majority of ways people meet is still someone you knew from school, work, church, met through mutual friends, etc.
As someone who has been married for a while, I've not used a dating app or site, but anecdotally I know plenty of people who have and success is "mixed" at best. Even for people using such tools, many of them ended up with someone they met through other ways.
South Park did it. 😂
https://www.youtube.com/watch?v=hEk0Tas7xgE
I’ve already tried this. “Please generate a sincere living apology for my wife”
And . . . ?
(I assume you meant "loving" unless the infraction to be apologized for was subject to capital punishment.)
There’s a case to be made that changing the dating pool from the few dozen people you see in bars and work and through friends, to everyone within a ten mile gps radius on a dating app, may have made people’s dating experience worse. When you have more people to choose between, your standards yet higher - and perhaps more importantly, *their* standards get higher too, so everyone takes far longer to find a reasonable match, and ends up with one not much better. Expanding to the whole world could make that dynamic that much worse.
Being on Hinge right now, I think this is very accurate.
Dating apps also expose more of us quickly. Things that shouldn't matter (or that might matter but can be overcome by other positive attributes) are now front and center; examples include job title, hometown, astrological sign, and marijuana use. This encourages snap judgments.
You don’t think there’s a pretty strong advantage to looking for partners who are within your general geographic area?
You obviously have never been to Colombia.
This sounds to me similar to the internet optimism of the early 2000s. By contrast I think it might exacerbate the current problems of the internet. If we can all talk we can all fight.
“…the dating game…”
That’s something I confidently predicted (in a discussion forum pretty much like this one) twenty years ago. Responses to the idea were almost entirely skeptical.
It’ll happen one day. But imagine the potential bumps in the road when a less-than-perfect translator chooses the wrong word or idiom!
"Every time they edit a machine translation, the AI is learning that much more."
What's the mechanism by which the data is being fed back into the machine? For legal docs/etc, or video game text, we don't put these out on public networks.
I mean, it'll get better over time anyway but unless you're feeding back the translated text with the source text how is it improving based on _your_ translation?
The pros often use integrated software where the human corrections are made inside the translation UI. Therefore the company that makes the software can use these corrections to improve its algorithm.
Ahh, did not know that, thank you.
On the analogy of Uber:
Uber is highly subsidized by investors and runs at a deep, deep loss. The prices they charged were never realistic.
LLMs like ChatGPT are also highly subsidized. The training is extremely expensive and even the per-query costs are quite high. Maybe the chips will eventually get cheap enough to break even on ads, but I doubt it.
"deep, deep loss" seems extreme. The market has baked in reaching operating profitability this year.
I don't really agree with this. I don't know if these neural networks are breaking even right now, but I don't think there's much doubt that Moore's law will make them profitable at scale. Amazon Web Services has been obscenely profitable for several years now despite steadily cutting their prices.
Moore’s law is going to run out of steam in the middle of this decade (and may already have run out of steam—Nvidia’s Jensen Huang, who would probably know, pronounced it dead in 2022) because of limits imposed by physics— you can only cram so many circuits into a silicon wafer. We’ll still get improvements in compute from improved parallelization and specialized chip designs optimized for performing specific tasks, but the late 20th century’s exponential improvements can’t scale.
I think that with the right set of optimizations, running LLM instances will be cost-effective for a lot of tasks, but the “just 10x the number of parameters” strategy for improving performance will stop being viable (because of both training and operating costs), so we’ll probably see a ceiling on their sophistication until there’s some sort of major paradigm shift. At the moment, I think LLMs are on track to be a useful and commercially important but not world-shattering tech.
All way beyond my pay grade, but I keep reading claims we'll be able squeeze more time out of Moore's law because of better software and better materials. Also quantum computing?
QC is completely irrelevant here, anyone who says this is BSing. Software is sort of orthogonal to Moore - the point of Moore* is you don't need to pay programmers to optimize things, everything gets better automatically
*Moreso Dennard but whatever
AWS is profitable because web dev is insanely cheap. I run a medium sized news site and CPU and bandwidth are free. All the money we spend is for database hosting and image resizing. So far at least, LLMs take orders of magnitude more computing power. A good person to talk to about this is Tim Bray. He used to be an engineering VP at AWS. He had some throw away line on Mastodon about how you can feel all the compute being burned by ChatGPT. The fact is that even for ChatGPT 3.5, responses are super-slow. The reason that responses are slow is that there is a huge and economically unviable amount of compute being thrown at them. In the long run, yes, Moore’s law will probably make it viable, but if the minimum LLM experience people expect is even more expensive, it might all just wash out. As it is, the CPU used by bigger LLMs is scaling up faster than Moore’s law. Again, everything can change, but it’s also wrong to just assume that it will all take care of itself.
My understanding was that it takes enormous computing power to train LLM, but not so much to run it after the fact. So it could be unprofitable to train it initially, but then its quite reasonable to use it.
I'm sure there is more to it than that, but would be interested to see you lay out more details.
The training vs inference very much depends on the nature of the product itself, and how popular it is. Training is expensive, but inference does not yet have "zero marginal cost" economics. That makes a big difference when compared to the traditional cloud model.
In Nov 2022, MidJourney was running into issues with cloud capacity to keep up. 90% of their cloud costs were from inference, and only 10% from training. On top of gating usage by price, they still had to (and continue to) rate-limit usage of premium users.
In the 8 months since then, both the number of "users" and "online now" (both include lurkers) has increased 4x -- to 16.65M and 1.44M respectively. Despite the increase in traffic, I have heard from some longtime users that user experience, especially wait times, have improved. To me this indicates that their inference costs are going down. But they are still at the order of 100x smaller than Twitter and 1000x smaller than FB.
Here's what Bard says about operating costs for ChatGPT.
If we assume that ChatGPT uses 8 GPUs to operate, and that each GPU costs $3 an hour, then each word generated on ChatGPT costs $0.0003. At least 8 GPUs are in use to operate on a single ChatGPT, and each question typically generates around 30 words. This means that the per question cost of ChatGPT is around $0.009, or 9 cents.
However, the actual per question cost may be lower than this. For example, if ChatGPT is able to reuse some of the computing resources from previous questions, then the per question cost will be lower. Additionally, if ChatGPT is able to be more efficient in its use of computing resources, then the per question cost will also be lower.
“…around $0.009, or 9 cents”
I keep seeing articles about how AI will revolutionize the accounting profession.
Here's an interesting article that describes fast-approaching limits to scaling:
https://asteriskmag.com/issues/03/the-transistor-cliff
Does anyone here knows where i can find Timothy's email?
Thanks a lot
tim@fullstackeconomics.com
I think this is the same guy:
tim@fullstackeconomics.com
timothy.lee@arstechnica.com
tim@vox.com
Not this TimBL: https://en.wikipedia.org/wiki/Tim_Berners-Lee
It took me a while a few years back to realize that this was a different guy, because there is some natural overlap in subject area.