ChatGPT Can't Kill Anything Worth Preserving
biblioracle.substack.com
biblioracle.substack.com
To take an analogy: bad voice recognition software abounds everywhere, not because it is better than what it replaced in terms of UX, but because it works just enough and allows massive cost-savings on hiring people to do customer service jobs.
A world where most marketing copy is written by mediocre AI, and more and more written and visual content are generated by big models that are technically impressive but intellectually hollow is going to be one where the quality of everything sucks just a bit more, but it's so cheap that it becomes pervasive.
(This trend is already apparent and not created by, or limited to, ChatGPT.)
Drives me nuts as someone who likes using subtitles.
It's nice to have access sooner, especially for less popular shows that might never get any kind of release outside of Japan, but it's a lot worse then most fan subs and sometimes misses entire sections of dialog.
I've been away from all that for rather a long time, so I don't know. Still, I'd guess these robo-subtitles are probably replacing the "fast and bad" niche. The "slow and good" niche seems like it could survive.
Back then there were multiple fansubber camps - notably Artic Animation which fell into "janky but really fast translation" releasing many shows and purist fansubbers who released maybe a few titles a year but with extremely high quality subs.
Then you had extremists who demanded no subs whatsoever and then those who preferred commercial dubs...
(I’m almost ashamed I watch Star Trek so often that I know this offhand…)
The subtitles are literally the worst, short of being incomprehensible. Like so bad that until someone experiences them, you can’t really grok how terrible they are. They’re a solid D minus.
Maybe it‘s an iOS thing. But every so often, the positioning codes sneak in to the subtitles themselves, so you start seeing x,y coordinates. And any time anything is italicized, the italicized part jumps to the opposite side of the screen, meaning the caption gets split in half. It’d be comical if it wasn’t frustrating.
(Thanks for the opportunity to vent about how sad it is that Star Trek is trapped behind such a bad streaming service. And I even pay for no ads, yet they still show ads every 8th episode or so.)
Subtitles are not something I normally have on since I know almost every word verbatim for stng and voyager... so maybe they are even worse than I realize!
I only pay for this service because I want to support my beloved star trek.. glad it's been getting some new life in recent years at least!
But yes.. it's sad that they have such a crappy service.
I always have YT's subtitles turned on and while they aren't perfect they are way better than the alternative (none).
Network/Streaming TV/Movies should absolutely be paying someone to, at a minimum, clean up the first pass by AI and we should demand that but I'm not at all ready to throw the baby out with the bathwater.
I realise it is not good HN form to talk about the comment instead of adding to the conversation but the subtlety with which you injected this was brilliant.
I was ready to ignore the change as human error/typo until it hit me.
Point about AI driven subtitling very well made!
Public numbers for prime is around 60,000 titles in 2021. Those are most likely going to be in four languages. And there's going to be at least two versions of each of those depending on what regions are playing in. That also assumes that a title is only one piece of content not a TV show if we assume that 50% of the titles are TV shows and each of those has a minimum of 10 episodes and we assume every title is around an hour including movies so that'll average out with shorter shows longer shows and longer movies. That ends up being around 4.8 million hours worth of content.
Let's just assume that the rate of title entry is one-to-one with the length of content though it's much more likely 1.5 to 1 or 2 to 1 given that people have to pause go back and fix things. That gives us with the average worker working 2,000 hours a year 2400 person years to data enter the entire catalog. Manual entry also obviously leaves rooms for poor workers or fat fingering. So if you wanted really high quality you would spot check other workers and so you might bloat that up to 3,000 people years.
So if you hired a brand new team of unskilled workers train them for 6 months with 3,000 people and then spent them for a year you would be able to backfill all of Prime's catalog right now.
But what happens when you on board say 10,000 titles from getting a new licensing deal with searchlight films?
You want that content up as fast as possible and people actually like it less if it doesn't have titles on it then if they're bad as some other comments have said.
Also just running the numbers let's say you pay someone around $30,000 a year to do this data entry at a very low wage double that for facilities and support HR and all that crap. 3000 employees at that wage for a year is 180 million. That training alone is 90M.
Should each of the streaming services take a large chunk of their budget just to make sure that a human reviews the subtitles possibly at an accuracy rate only slightly higher than the machine learning can do currently. Would you rather have 5% human coverage or 100% 90% accurate AI coverage.
From my understanding most content doesn't actually have subtitles on it unless it was on a premiere TV network that was required for government regulations to subtitle their shows such as a BBC TV show. That means that the streaming networks are actually doing this out of customer interest as opposed to being required to do so so they're actually backfilling work for the people that produce the videos in the first place. And from the little bit that I've worked in broadcast subtitles sharing isn't all that standard and Netflix for instance may have added subtitles but that doesn't mean that prime video or Hulu is going to get those subtitles if they take on that content later the video producers aren't that interested in pulling the information back into their catalog they don't have the tech support to do stuff like that.
Also almost all of this was dictated via Google voice AKA subtitles via ai and the only mistake I noticed was it didn't understand tangently it instead put 10 generally.
I had to go through hoops talking to the Doordash bot to get a change to my order. And I pray for the poor soul that runs afoul of Google…
This one really puzzles me.
Voice recognition replaced phone trees. And from what I can tell, it's just worse. In this particular use case I don't think it really replaced tier 1 support. Either I'm missing something that it's a lot better for some group(s) of people, or people adopted it because of promises that failed to materialize.
I could see it saying "I understand you want to reach customer representative", "Please select from these options while I connect you to a representative" while circling you around a graph of options or just automatically hanging up after 45 minutes to an hour of being on the call.
It's 100% internal metrics.
When I was in the military, there was some metric that the clinic was graded on - I think it was number of days between when an appointment was made and when the visit happened. So the clinics just didn't allow unimportant people to make appointments more than 2 weeks out.
Can get on with life until they ring you back then.
A call deferred is a call denied.
Convincing the customer to give up is one way to reduce workload.
See also: automated systems that play a prerecorded message along the lines of "we're currently experiencing higher than normal call volume" regardless of how long wait times even are.
A manager with misaligned incentives that reduced call center costs by 50% and got promoted also wins.
Customer service, be it commercial or tech support, is always treated horribly.
It took a global pandemic for the California DMV to finally start an email ticket system. It is hands down so much better than going in to the DMV and has to be way more efficient for them. You simply fill out a form and you get an email with a ticket number. They respond within 3 business days asking you to send images of documents or whatever is missing. A comparatively pleasant experience.
This would provide the customer with a written record which could be used against the corporation. Phone is generally favorable to the corporations because the corporation can decide whether/how to retain/destroy records, but the customer will almost never have a transcript or recording to use for their own evidence.
Many corporations have moved to a system where you can either call an automated line which does its best to make you hang up...or you can fill out a browser form which provides the customer no ticket/confirmation number/email/any record that you ever filled it out.
So the companies save money by then cutting the size of their call centers.
Hit the nail on the head. ChatGPT will replace the wrong things, and probably create a world where we still need to solve problems ChatGPT has become attached to.
It's all noise, I can't perceive a signal at this point.
But give it context and it flies. Ask it to write a cover letter for an engineering role, it will turn out mediocre crap.
Give it the actual role and your CV in the prompts, you're pretty much good to go. It's unique, you still control the narrative (please highlight my ability to work under pressure) and it's done in 10 secs.
I really don't see what isn't to love about that.
Also it usually writes better if you tell to write well.
My analogy is that it is to writing what the spell checker is to spelling. Very few people in the world have to be good at spelling anymore. Yeah everyone is okay, but it's not as valued a skill as it used to be. ChatGPT is doing the same for writing. Yes you need to be able to write to an okay level but you don't need to be able to write well, ChatGPT does that.
ChatGPT, stable diffusion and I am sure upcoming music models will enable this mechanization in the arts, where great artists will be unaffected but making it nearly impossible for “good enough” artists to compete while also pushing up the floor of what is an acceptable competency level that merits being paid for making it more difficult for people to support themselves while improving their skills.
Now AI's coming for most of those who survived that first culling. And not just the middling-talent folks this time.
My grandfather and group sang. From what tapes I have heard of them, they didn't sound worse than any other group of Appalachian gospel/blue grass singers.
My parents' wedding photos weren't 'professional' and I think they came out pretty well, perhaps even charming.
Sometimes I think people are putting too much stock in the notion of a 'perfect' life, rather than a lived one.
This is going to push us back to making our own music like we use to. Singing/playing should be a fun group activity rather a performance given to a group of non performers. We are all artists.
Printing, photography, radio, film, television, etc have all increased the availability and 'quality' of entertainment available at the same time as reducing the number of creators involved. (Obviously there is some debate possible around quality)
It also assumes that they were constrained to local creators.
It is possible that people simply did other activities that don't involve a creator. Another possibility is that most people consumed content from a small set of non-local creators, like authors with wide distribution.
Being a half-competent folk musician or good storyteller or being able to sketch pretty well used to be super valuable to your family and community. Not so much anymore. Expressions of those sorts of skills are more often tolerated than genuinely looked-forward-to, now. The need is gone.
People on this site complain about folks being consumers and not producers, not being creative—well, for a large swath of the arts, that's where it started. Recording and photography. Took it from something that was strongly socially encouraged & rewarded to something private. You can't fix that with "maker" movements—not in any major way.
Technology changed the social context and wiped out the external motivation & encouragement for a bunch of kinds of creative expression that were accessible to & achievable by the masses. AI is more of the same.
There's a popular youtuber named Joel Haver who films videos, then uses an AI tool to convert them into animations, so they can be magically put into a space or fantasy setting.
AI dungeon is an text adventure where the content is AI-generated.
I also imagine tools similar to Github Co-pilot for other markets. Some AI could generate music or video games levels based on inputs from a user, then the user can take or modify the best bits. The goal hopefully being that they get something no human could have thought up, instead of just generating a bunch of mediocre content.
Hang on, AI tool? Joel Haver converts every frame manually. His method is the literal opposite of using an AI tool. It's incredibly painstaking.
No, they cost exactly as much as they always had. Before machines, people didn't have wardrobes full of tailor-made clothes. Each person had, give or take one, exactly as many tailor-made clothes as we have today.
Many, perhaps most, people don't have any tailor made clothes because they can't afford a tailor. Which has indeed always been the case. But before machines, people who couldn't afford a tailor had no other option for clothes except homemade. Now, with machines, they do.
Not impossible, but definitely a raising of the bar.
Reminds of the Empire in Asimov's Foundation where they knew just enough to keep the current tech running, but not enough to fix it if something major breaks or create new tech.
When the AI breaks something, we will be missing the people who knew how all this s$%T works.
well, no, you probably didn't listen to much music or see many stories at all because the average person couldn't afford to experience such things more than a few times in their life. shocking that Luddites are still a thing in the 21st century, especially here.
Sorry that was a ramble, please take no offense.
Whereas now we have the complete opposite problem, we're inundated by (much) lesser variants of these things to the point of saturation, addiction, and emotional depletion. I'm not sure either one of the extremes is particularly desirable, nor how acknowledging issues that mass-market consumerist societies currently visibly suffer from makes one a Luddite.
Means that many, many more people can have something of passable quality. For example:
> not that long ago all clothes you wore were custom made by a tailor
If you could afford a tailor; otherwise you had to make do with homemade rags that constantly needed mending and looked terrible.
> all the music you listened to was played live by musicians
If you could afford to go to concerts.
> all stories were brought to life in front of you by theatre actors
If you could afford to go to the theatre.
> now all of those while still available are significantly more expensive and niche than the mechanized (production or reproduction) equivalents
Which the vast majority of people can afford, and which significantly improves their quality of life. Now they can buy clothes at Walmart or Target instead of having to wear homemade; sure, not the same as a custom tailored suit, but good enough. Now they can buy digital recordings of world class musicians and theatre actors for much, much less than it would cost to see them live.
> ChatGPT, stable diffusion and I am sure upcoming music models will enable this mechanization in the arts
That already happened decades ago, as soon as mass produced recordings became widely available. It's already next to impossible for any artist who isn't world class (or, more precisely, is not publicized so that people think they're world class) to make a living at their art. ChatGPT and the equivalent in other arts aren't going to affect that much.
Not to mention the degree of control over what music is actually being played.
Sure, people with more disposable income could enjoy more formal, and perhaps more professional, musical performances. The growing middle class created demand for music and other forms of entertainment.
Music was taught in the schools. Even small towns had a bandstand in the park. Music was used in public ceremonies, including church. Any tavern was likely to have a musician, paid or not. Sometimes the pay consisted of food and lodging.
> If you could afford a tailor; otherwise you had to make do with homemade rags that constantly needed mending and looked terrible.
Or you bought clothes infrequently and maintained them, and they lasted longer because the quality was much better.
People are so used to fast fashion these days that they assume that the poor in previous eras suffered in this regard way more than they actually did, because people assume the poor quality of clothing today is representative of how it always was, rather than a recent phenomenon.
The quality of clothing was much, much higher - even for the relatively poor - and clothing wasn't treated as disposable, so people would maintain and repair it, so it would last much longer.
> Which the vast majority of people can afford, and which significantly improves their quality of life. Now they can buy clothes at Walmart or Target
Buying clothes at Walmart or Target is a step down, not a step up. (And ironically, it's not even necessarily cheaper!).
If we're talking about clothing, it's very clear that these trends have served to the detriment of the average person, not to their benefit. The ones who actually benefit are the ones capturing the profit - the Sam Waltons of the world.
And it's nice that when my clothes are stained or irreparably damaged I can afford to replace them quickly.
This is a preference that has been demonstrated time and again throughout the world as it became an option.
So no, I don't believe it is clear at all how this is to the detriment of the average person.
This is extremely incorrect. First of all, fast fashion isn't actually necessarily cheaper. The real difference is that fast fashion exists in an industry that has gutted the entire infrastructure for alternatives, so people actually don't have the alternative options anymore.
It's like saying that "people prefer to drive cars, which is evident if you look anywhere in the US". Sure, almost everyone drives cars, but that's because the previous rail infrastructure was literally ripped up and destroyed by the auto industry, so now there's no alternative.
If you look at the developing world, it's very clear how incorrect your statement is, because there, fast fashion is more expensive, and other forms of clothing are far cheaper, far better quality, and far more common.
All of that is part of fast fashion - including Walmart and Target; it's just the lower end or very far down the pipeline. You can't separate one from the other, because they are economically codependent.
> And the quality is perfectly fine
Fast fashion is indubitably much lower quality, and it's weird to see people here denying that, when the fashion industry itself is completely in agreement about that fact. It's not a secret - they discuss it openly.
Those articles of clothing I'm talking about are almost never worth mending, but they still can last for years.
It's all fast fashion, just different points along the line.
Yes, broccoli and kale look different, and one is a fancier class signifier than the other, but at the end of the day they're still the same species.
So there's only high quality tailor-made clothes... and fast fashion?, of which there may be a spectrum that goes from something like "very fast fast fashion", all the way to "slow fast fashion"?
> Yes, broccoli and kale look different, and one is a fancier class signifier than the other, but at the end of the day they're still the same species.
So you're against... clothes? And possibly also agnostic when it comes to the existence of taste buds?
I'm sorry but your line of reasoning is making zero sense to me. This is getting to the point where I have to ask: Have you ever bought clothes yourself or does someone else buy them for you? Can you put into words of your own (without googling) what makes an article of clothing like a shirt not fast fashion?
They made econmic trade offs depending on their level of poverty.
> Means that many, many more people can have something of passable quality.
There is also a feedback loop where automation puts people out of work, and even though it helps make products cheaper, it also creates less of a market for the products. Your post makes it seems like there are only positives.
Looks very much like a race to the bottom dominated by a few big players to me. Consumers don't seem to care that much.
100 year houses are only livable because someone put a ton of money into retrofitting things like plumbing, electric, and HVAC. Most of that has had major rework done several times as well since it was first added. Even at that, the fundamentals of those houses mean that they cannot be retrofitting for good insulation, and it can be argued they should all be scrapped for that reason.
I think there is an ethos of buying new and overpriced (and also buying "experiences") to show off that you have money for that and aren't some poor nerd trying to optimize their budget at the expense of consumerism. Often you don't actually have that much money for actual conspicuous consumption, so at least you're trying to imitate the Apple aesthetic. I am getting this vibe from talking to a distinct set of people of my age group, of course not all.
Office furniture seems to have been incredibly well-built, though, by modern standards. Kinda like how office and government buildings used to be built a lot better. Institutions seem to have cared a lot more about that kind of thing back then. They built desks for ordinary low-paid clerks or typists like they expected the office to still be running, and still using that same desk, in 200 years.
[EDIT] Also, part of what makes older furniture seem so nice is that excellent-quality wood was abundant. It seems so nice because they used wood that'd be ultra-expensive luxury-grade today, for, like, interior structural parts you never even see. Same story for houses. I've seen wooden support beams that'd be almost impossible to find at any price, today, the quality's so high and the piece is so large. Floor underlayment (again, not even intended to be seen) so knot-free it'd probably be used for veneer, now. That kind of thing.
edit removed: Leonard Part 6.
100 years ago, how much did the family dinner table cost? Would you measure it in hours or days of salary? When that family in 1923 went table shopping, they were either going to buy from a company like Sears Roebuck or from a local store and there probably wasn't much variety from either source.
Today, there's lots of cheap stuff available, but there's also lots of high end custom stuff available as well. We have more of everything.
There's a moat, and it's very evident just by looking at family pictures from the sixties and eighties.
Cheap, mass-produced furniture has taken the bottom of the market but there really is a wide variety available today.
It's been like a 1.5 months since it was released.
For me I am doing a lot my writing with it becuase my writing is worse than mediocre.
As long as you don't expect it to know things off the bat, it has a decent memory, though seems to be unaware of what it knows e.g. https://news.ycombinator.com/item?id=34370057)
So you can just feed it the good info, get it to update things with more context and it will spit out pretty decent prose (though you have to ask to for it to be well written).
There was a similar dynamic with translation market and Google Translate. It didn't matter that human translation was superior. What mattered was that Google Translate (or similar) brought down prices so much that it effectively destroyed the market for everyday-type of translations. Why? Because customers said that, well, just use Google Translate and then improve it a bit.
as a result, it's a lot harder to make a living as a freelance translator these days - there are fewer jobs and what jobs exist are often proofreading machine translations, which command lower rates because 'the machine did most of the work already,' even though they often require full rewrites.
at the same time, human translation quality has gone down too, since a lot of people will pass off machine translation as their own work, and when rates are too low, you can't spend too much time on any particular job.
That is, I don't know if I would say getting frustrated by an automated voice system is the same thing as a hammer breaking for your given Dasein, but one could make the argument I bet.
Does it, really? I find that most of the time that while people can spot the difference between facts and a sales pitch, repeating the pitch enough still changes the common discourse.
People can say things like "technology x may be more expensive but saves development time" (where X can be anything from the latest frontend framework to microservices to voice recognition) while in fact there is no data that it saves any development time at all.
Is it even true that voice recognition reduces the amount of customer service compared to a simple menu system? If it isn't, the premise doesn't hold.
1. Creativity can be achieved at faster speeds than humans can consume. 2. We have no evidence to say AI and the techniques behind it wont keep improving (there are already so many low hanging fruits, alot of it is infra problems and the other half that we can't even imagine can probably be solved by AI better than humans can)
By the way, I asked ChatGPT to fix/improve this previous comment and this is the result:
I agree with the article, but perhaps you are using ChatGPT in the wrong way. As a non-native English speaker (as many of you have likely noticed due to my grammar mistakes), ChatGPT has been extremely helpful in rewriting my texts, correcting grammar, and making them more polished. It has been an invaluable tool in that sense. I now feel more confident when sending emails and writing important messages, knowing that the grammar and tone have been reviewed and approved by this AI.
Proofreading shouldn't change the meaning or intention of a passage.
"trying to do the wrong thing with ChatGPT" vs "using ChatGPT in the wrong way" "have been "approved" by this AI" vs "have been reviewed and approved by this AI."
I prefer the original phrasing.
At some point, the boundary between simply "polishing things up" and actively guiding the interactions became blurry, and the AI eventually becomes the go between between everyone and starts taking over.
We know this response has not been "enhanced" by the AI if my second and third sentences have not been filtered out.
The "ChatGPT improved" result is certainly a little more polished, but given ChatGPT's ability to confidently misinterpret / hallucinate, I would be worried that it could subtley change meanings, especially for more technical content.
Of course this problem isn't just limited to AIs, I've known (native speaker) Project Managers and client liaisons who have tweaked text I have written to ask technical questions and in doing, totally changed the meaning (and then done the same in the opposite direction, so the response was especially bewildering!)
I can understand that it's improving your confidence (and by the look of your 1st paragraph your English is good enough that you could easily identify if it changed the meaning in your message -- and that combined with a confidence boost is maybe worth the extra proof-reading you need to do) but that wariness of risk is just my $0.02.
One of the heaviest consequences of AI proliferation is that the value of understanding will continue to plummet, while the value of asking for help and delegating will rise.
This is all fine as long as there are plenty of willing subordinates, man or machine, to do the dirty work of actually knowing things for you, but what happens when they wise up to the fact that they're getting the short end of the stick?
Imagine if restaurants all started using AI to cook their food. It would destroy the taste of food because no matter where you go the food will always taste the same. The great thing about restaurants is you could go to two different pizza places on the same street and order the same type of pizza and they would both taste different. Now imagine if AI got a hold of the recipe and made the pizza at both places. They would turn out same and taste the same.
How is that better?
The original was perfectly clear, and had your own unique "voice" to it. The ChatGPT version sounded like every other piece of ChatGPT, and honestly, I've become somewhat allergic to them now.
The thing is... the only time that grammatical mistakes actually matter is when they make meaning ambiguous. And since ChatGPT doesn't know what you meant, I would be more worried that it would further distort any mistakes in that area.
It is probably wrong too. ChatGPT is a native English speaker as far as I can tell.
“Maybe you are trying to do the wrong thing with ChatGPT” vs “Perhaps you are using ChatGPT in the wrong way” are subtly but substantially different statements.
One implies the incorrect usage of the software, the other implies the software being used for the incorrect things.
Oh look another apologetic non native English speaker using perfect English
- By outsourcing your thinking to ChatGPT, you are not investing into improving your own skills.
- ChatGPT helps paint a picture of you that is not true to who you are
- This may lead to situatons that may make you very uncomfortable, for example speaking in person with people who built a ChatGPT-based image of you
It is probably a good idea to not rely on ChatGPT for your text any more than you are relying on Photoshop for your pictures.
I've been using ChatGPT (as a non-native speaker) as well and found it tremendously useful—not only to catch grammar mistakes or provide some better vocabulary, but to understand what might have been off and provide more alternatives. I often ask it to rephrase my sentences in different styles and then pick and choose what I like the best ("in the style of the new yorker", "in the style of ayn rand", "in the style of ezra klein", "in the style of CNN").
Anybody that cares enough to have ChatGPT edit their sentences and cares about what they are intent to express, will I think benefit tremendously from such a tool.
In fact, it is what ChatGPT is really good at: a search engine for vibes, tailored to what you feed it. It is I think a much richer, much more enticing tool that following some rules out of Strunk & White or some other bland business writing handbook.
I am not worried about whether chatgpt would hallucinate as I’m sure you read the responses and I know it’s easier for me to read and listen in a foreign language than compose.
By the way, I think you’re doing an amazing thing here taking the technology and making a profoundly useful tool out of it. By reading its revision you can get corrections and pointers. I use chatgpt similarly but in other domains. It’s not always “right,” but it’s close enough to point me in directions I wasn’t aware of before.
For example, it took your perfectly good:
>I am not a native English speaker (as many of you already noticed due to my grammar mistakes), ...
And replaced it with
> As a non-native English speaker (as many of you have likely noticed due to my grammar mistakes), ...
While a lot of native speakers would write something like this, it's awkward and incorrect. You didn't mean to say "As you may have noticed, as a non native speaker, blah blah", you meant to say "As you may have noticed, I am not a native speaker."
If chatbots become popular, the ad space would go from search engine to product placement in chatbots replies.
But chatGPT would actually write useful texts, most of them more useful than the average blog post. Over a few years they might be just as good as the best human articles.
Normal people could write a draft version and have chatGPT reword it, that would control the contents of the article. But I expect spammers would just clone other articles, they don't have the time.
One thing AI could do is to scale up validation. Search every topic and note the answers - do they concur? do they disagree? what is the distribution. Maybe there is no answer?
This would be a reference for AI to stop hallucinating and making factoid mistakes. The model should know when it doesn't know, or when it is stepping on a landmine by forgetting to mention something.
In the end I think the internet is going to be full of spam, AI generated or not. And we will need to use more AI to extract the signal from the noise. This time it should be local AI under user control.
If web sites start being written by A.I. and ChatGPT starts getting trained on A.I. output, it's going to start spewing garbage results.
It's only useful if trained on human input, A.I. have no model of the world and no A.I. can function if trained by A.I. written texts.
There’s also a weird assumption that humans don’t write a bunch of garbage as is.
As to point 2 , the argument is that, like with google search results the dataset will be corrupted and get worse, not that it won't contain some human written prose, so that's irrelevant when you say humans will still write things.
Claim 3 seems to be an argument that we can solve the problem of machine learning being based on bad bots by throwing more bots at it, which seems to be a very Rube Goldberg proposal which I'm not sure will work.
Finally is the claim many humans are dumb, so the more chat bots that impersonate humans, the better I guess?
Yes, you can solve dataset problems with models. Look at ConstitutionalAI from Anthropic. They only write a set of rules (the so called constitution) and the AI will derive a large dataset of demonstrations all by itself. This can do away with the RLHF stage where OpenAI has the best results so far, but costs money to label and is harder to update.
The model "Claude" from Anthropic is being discussed recently on Twitter
Google is overall a pretty bad benchmark IMO. There's a lot of quality content that they seem to struggle to find. Not that it doesn't exist, you just won't find it for a host of complex reasons.
With Google you are the product
It used to bother me immensely. On behalf of the employer and for their own sake. I can still find that sentiment if I look for it but today, my perspective is that at least in many cases, it's ok. If the company does well and the employee doesn't hate their life because of unfulfilling work, we're all better off than if they went unemployed.
Yes, ideally everyone should follow their passion and pursue the life of their dreams but some people don't think of work as the source of that. However, that doesn't mean that they would rather not work at all -- and I don't mean just because of the loss of income. I know that UBI is all the rage in here and I think it's an interesting thought experiment but I just don't see a world where only the 1% of performers in any field are required and able to work is a promising thing.
Not only do some people not think of it that way, I doubt enough necessary jobs are the dream jobs of enough people for it to work out. I bet we need the vast majority of people to be working in jobs that aren't any part of any dream of theirs.
Besides, maybe their dream wouldn't pay, so they fund it with their bullshit day job. Then it are pursuing their dreams. Hell, maybe that dream's just "raise and provide for a family"!
> I know that UBI is all the rage in here and I think it's an interesting thought experiment but I just don't see a world where only the 1% of performers in any field are required and able to work is a promising thing.
I entirely do not follow how this would be the result of UBI.
I think we're agreeing there.
>I entirely do not follow how this would be the result of UBI.
I didn't mean to imply that it would. I've just seen people argue that UBI is a solution to the problem of people not having work.
Ah, I took you as meaning that UBI would cause only the top 1% of talent to work. Sorry, guess I misread that.
I'd say automation would free all those folks up to find meaning in other pursuits, like, say, the arts... but, uh, unless there's a society-wide rejection of AI art, that's looking rather iffy now. Like they could maybe still do it, the same way people enjoy doing Sudoku puzzles even though computers can solve them better, but they'll have to be content with the finished work having the same value to others as a solved Sudoku puzzle, i.e. none whatsoever.
Just to rant about that a bit, we acknowledge and ignore that a large proportion of people barely scraping by would probably not do what they were doing if not for the threat of homelessness and starvation. It's true that the way that things are built would have to change with UBI because most of what we benefit from in the current system is based on the exploitation of the most desperate. We should note though that a large proportion of the benefit of that exploitation is not going to bettering society, but into the pockets of the most wealthy.
At least with UBI we would remove exploitation of one's living situation as a tool to extract labour, which removes a lot of power from the existing monopolies.
I've seen what happens when an individual human enter a delusional feedback loop and diverge from reality, that pivotal moment when sensemaking becomes self-referential. They get a psychotic break. I think we're going to get front-row seats at watching a whole, planet-spanning civilization enter a psychotic break.
Arguably, we are already there. It's just that, I think deploying ChatGPT will greatly accelerate this. This won't just be fringe or extreme subcultures, but rather, the mainstream cultures, as our sensemaking turns inwards.
I for one welcome our new chatgpt overlords.
However, when will they be able to do work that is worth doing? GPT4? GPT5? There will be a point where we have to grapple with that.
Enjoy them doing something productive (I hope)
Of course, this often does not happen for one of two reasons. First, society doesn't always choose the civilized answer; it often chooses to just marginalize the folks who just lost their jobs. Second, society often really wants to pick the thing they're teaching these people to do -- this is how you get people trying to teach coal miners how to code, despite coal miners being by and large uninterested in coding.
These are pretty real problems that we're already grappling with, but we're going to need to get very serious about it very soon. LLMs aren't the only reason why.
The rest of them, and their families, ended up dying in utter destitution. That's what "we" will probably do again.
[1] https://en.wikipedia.org/wiki/Luddite#Government_response
The state made machine breaking a capital crime. That is, if you break or destroy a machine, you can be executed.
Many who were tried and convicted of machine breaking were accused or guilty by association, or literally broke machines, and were executed. States responded with death and contempt, not merely force.
Funny that bullshitting was the easiest job to automate to perfection. But now everyone can spin up their AI bullshit operation. Anti-bullshit AI would be very valuable in this situation.
Probably: Let them eat cake
If you meant, "when can AI perform all writing tasks such that it would be a waste of time for human beings to develop the ability to express themselves in writing", well, fine question, likely answered far too soon.
"Production of content" can probably go a lot further than you think, for a model that has internet/API access.
Would love to see more warrants on the claims you are making - feel like you haven't fleshed it out enough in this comment.
Ex: practicing scales on my guitar is work worth doing, because of its effect on me, not because it produces anything of value in itself or of interest to anyone. It is meaningless that my computer can do scales vastly better than I can, because the production of scales is not what makes the work worth doing.
If the work leads to the production of something worthwhile net of any other losses from the work, the work is "worth doing."
So "work worth producing" is, at minimum, a subset of "work worth doing"
A swimming pool can have a certain amount of urine in it without anyone noticing, and that’s the way it’s always been. It’s not a problem as long as the amount is small and you add cleaning agents. The problem is that filling the web with content created by ChatGPT means adding more and more urine to the pool until someone notices.
It will be fine, for a long while, until it suddenly reaches a threshold where it’s noticeable, when all of the highly ranked hits you get for a question are subtly incorrect to the point that you can’t trust any of them, when more and more Wikipedia contributors are discovered to just copy-paste text from an AI, and then it’s no longer fine, but it’s too late because you’re still in the Internet swimming pool and now you’re swimming in piss.
Has anyone tested what happens when the output of a LLM is fed back as the training data?
Why was this possible? The game itself acted as an anti-bullshit filter. So you can train on your own generated data if you filter it.
Like this one: Large Language Models Can Self Improve
1. It presents information really well that is completely false. I don't know exactly how it works under the hood, but sometimes it spits out paragraphs that read like the fever dreams of a flat earther.
2. It will cause our communication skills to further atrophy. Maybe in an innocuous way (no one memorizes epic poems anymore), but with GPT I think it will make people less likely to have the ability to be persuasive.
(spellcheck means we no longer have to worry about getting distracted by orthography; if gpt means we no longer have to worry about getting distracted by grammar, I should hope we could then move on to helping people express actual ideas)
Last week the dude next to me in the train asked chatgpt to generate a "good morning message" for his gf and sent it verbatim to her.
At the same time he was working on what looked like a work related slide presentation, 90% of it was filled with chatgpt answers
As a test I asked for the world record speedrun for pokemon yellow, it made up a time associated with a runner who didn't exist yet the time was still rough 20 minute ballpark of what the any% time was in 2021. Bizarre stuff.
I expect it will raise up a lot of writing to low bar of competent but voiceless. Which would actually be an improvement on many SEO spam blogs.
I don't really see a problem here. Us, meatbags have been at it for a few millennia, bullshitting for fun and profit, and that train is moving ever faster with the improvement of communication technologies.
From what I've seen, this new rival isn't even exactly persuasive, it's merely super-confident (and we're giving it a lot of credit because we talk to it as a novelty machine). It has no malign (or benign) intent, no idea about where and how and where to put the pressure, and training for that (as opposed to training for a nice-looking texts that keep the topic and have a correct grammar) is something that must be very non-trivial. It doesn't even have any solid agenda besides a few silly filters and canned responses when it hits those. Though, of course, it can be trained for any certain bias.
And maybe I'm too optimistic or naive, but I'm also guessing that if a future AI actually manages to becomes persuasive, then trained and unleashed onto conspiracy communities it very well could lead to their demise, outcrazing the craze to the extent it collapses into its own weight. Just let's not forget a kill-switch, lol.
"Sure, ChatGPT can write some basic code, but it can't come up with creative solutions like I can." Yeah, that's true. But how did I get to the point of being able to make creative solutions? Years of working in the trenches, learning all the ins and outs of how to code and not over-engineer that college never even had a hope of teaching me.
Yes, nobody wants a life-long career developing formulaic CRUD apps, or writing catalog product descriptions, or making logos for every tom-dick-or-harry consulting shop with delusions of grandeur. But if you want to become a skilled developer, writer, or designer, you kinda need to grind through a lot of this bullshit work to grow a sense of taste to be any good at the higher-level work.
So what happens when we cut off that pipeline of junior jobs? It's going to happen, and it's going to happen from the bottom, up. We're not going to see big corporations take risks on trying AI software developers when they have thousands of resumes of the best undergrads to sift through every month. We're going to see it from the companies on the fringe, who couldn't afford to hire even a junior developer. They're going to take a low-risk chance because the alternative is just not do the work at all. And they'll prove out the model and 20 years later the only people who understand the basics anymore will be more interested in management positions than writing yet another Django app, and 20 more years later they'll be retiring with nobody else coming in behind them.
If you can't get paid to produce art as a journeyman that supports your basic needs you won't make it to being a master artist except in your free time, maybe, and you won't have the same work references for work on projects during your development. It will be somewhat like in the writing world, where many or most of the major writers are independently wealthy because nobody else can afford to take the time that writing requires off of real work that feeds and houses them.
Imagine replacing all the engineers in training with ML and wondering why there are no engineers later. Thats where we're headed.
The ML experts who are making these systems aren't professional writers or professional artists. Industry is always convinced by automation and lower costs, and boards who make these decisions are senior professional management not creatives. Neither group ever cared about the ladder behind them in other industries or their own companies.
The question is one of what we want to accept as societal norms, or push for legislation of regulation. Personally I think the idea of banning such systems won't work, but all writing done by ChatGPT or other 'creative' work done by other ML models should be regulated such that it must be marked with a discloser that it is the product of a 'creative' machine learning model. Human authorship is an important distinction that needs to be clearly distinguished.
To reconcile this statement with his admission that ChatGPT reliably turns out passable results, we must assume the average student simply rearranges word patterns as well. So developmentally, this must be a mile marker on the road to competence.
"[Writing] is an embodied process that connects me to my own humanity, by putting me in touch with my mind"
ChatGPT's experiences with its own "mind" are fleeting and lost forever with each reset of prompt and zeroing of prompt history. Prompt compression and embedding architectures, combined with the steady growth of hardware memory capacity, should allow new models that continuously generate and process "thought" prompts. This should allow the primitive emergence of self-narratives and make a leap, I feel, towards the generation of true meaning.
> To reconcile this statement with his admission that ChatGPT reliably turns out passable results, we must assume the average student simply rearranges word patterns as well.
Alternative explanation: the “passable results” test does not distinguish between ”rearrange word patterns” and whatever the students might otherwise be up to.
> So developmentally, this must be a mile marker on the road to competence.
Alternative: the students with passable results are not on the way to competence
Other alternative: it is a mile marker for students but not for the AI (just as learning to walk mile be a mile marker for human babies on the way to baseball playing, but not for foals, who learn fast to walk and never to play baseball).
The fact that students and many people cannot write is because their thoughts are a disorganized mess, only held together by the self deception they are not a mess. Yet, attempting to write down their thoughts, the disorganization is easily seen and then they organize their thoughts through the process of writing legible prose. Writing is how many organize their thinking, and if they never wrote they'd never organize their thoughts.
What if on day one you showed up to high school English class, and there were 2 terms on the board: 1. COMMUNICATION 2. PLAY
The teacher then said: "The most successful people in the world are the best communicators. This is true in every field - business, science, technology, medicine, sports, the arts and politics. The most successful people also have a love of play and get the chance to practice it during their work day. Whether its in business, the arts, or intellectual pursuits. Would you like to learn to be successful like them? Have I got your attention? Alright, let's go."
Although this would be a non-egalitarian and elitist curriculum , it would be highly appealing for me. I would want to attend a course like this, and I would forgo the author's offer to not attend and still get an A. I would attend, and get the A.
Surely we can improve course design so high school students aka humans 'studying life' can learn success, and it just so happens that writing is part of becoming your best self and reaching your potential.
Teachers could benefit from some 'marketing style' re-positioning and re-framing of the curriculum to what is relevant and resonant for students.
There is a major relevancy problem in the high school curriculum.
What if the gym teacher said: "today we're going to talk about Jordan. Best fadeaway jumper in the business. How did he do it? You're about to learn about one of the greatest players in history."
Or if the English teacher said: "Obama. Jobs. Musk. Why are they so persuasive from the podium? Listen up, because you're about the find out how they craft presentations that move millions of people and billions of dollars in product."
My high school English teacher let me write an essay analyzing lyrics in songs by the Doors as poetry. It was awesome, and I got an A.
-Dan
There were times later on when I made most of my income from writing, so I'm sure she was wrong.
Or you were just really bad at everything else you could do.
I got a perfect score on english sat, math sat, physics sat, etc. doesn't mean I'm looking at all of those things as jobs.
When he is writing a story that largely consists of internal dialog and a character mostly in isolation his writing is brilliant.
His style is painful in the Passenger and an English teacher would be write to criticize him. That book has long passages of dialog and his style is a detriment in that book. It just causes the reader to have to work extremely hard to decipher the conversations. I have not read Stella Maris yet, hopefully it has less back and forth dialog where his style suffers.
1. What did you do yesterday?
2. What will you do today?
3. Any blockers?
1. titles of tickets moved within the past 24 hours
2. title of ticket in progress
3. No blockers right now
1. What we discussed I was going to work on yesterday.
2. Continue to do my job.
3. Of course not. If I did, I would have made it known long before this.
Product Manager-aaS
> in general
Research, composition, and communication are a big chunk of what ChatGPT is good at assisting. I don't know enough to claim that it /is/ the general case, but it's very likely not a minority.
And if that's the case, and these activities are not /essentially/ bullshit, then I'm still wondering how AI disruption can be claimed to be a reliable signal of bullshit.
A pump, a motor and a timer
There is a secondary definition of AI that says basically an AI is anything that we haven't yet figured out how to do. Once we figure out how to do it, it ceases to be AI.
Indeed, terms can and do have multiple meanings, but I'm not sure that secondary definition fits the topic of discussion.
That’s not the way anyone uses the term.
No. In that case the intelligence is applied by the user. A can opener could be a component in a larger AI system. An AI needs to be aware (in the artificial sense) of its surroundings.
Edit: I forgot that AI can openers do exist. The can opener I own is of the dumb machine kind and I hadn't fully considered what else might use that term. Yes, those are an AI.
> That’s not the way anyone uses the term.
Except when they do. Indeed, the "an AI is anything we haven't figured out yet" definition is about as common, but doesn't seem to be the definition that fits the topic at hand.
A third definition where AI is defined as an ANN seems to be emerging in popular culture, but again that doesn't really fit the topic at hand as a lot of the job-replacing intelligences are not based on ANNs.
Perhaps you are actually thinking of what is more commonly known as AGI (artificial general intelligence)? None of these machines of which we have spoken of are AGIs. That I can agree with.
If, for whatever reason, consumers of content are not immediately attuned to the difference between vacuous GPT-nonsense and something that would bring them value if they just thought hard enough, it might lead to an overall reduction of returns to thoughtful, high-quality content, even if that content is only momentarily over the head of the reader, but wouldn't be in the long run.
This obviously to the detriment of both the consumers and producers of content.
These tools are being used in the name of productivity, which isn't a bad thing in itself, but when it's used to make a small number of people more wealthy and put a large number of people out of work, then it's an issue.
I would venture to say the obsession with standardized tests sucking the fun out of school stems from the same problem. Namely, the commodification of our education system in service of our larger economic system.
Everything is centralised right now. So disclosures, trust, transparency don’t mean anything. Atleast not on a web that we don’t own.
Also, since when have humans being transparent. We’re all data collectors as our brain is aggregating thoughts of others and those of our own and then sharing them with technologies.
Now that technology showed us the cons of a centralised system, it spooked plenty.
The future is reminiscent of Westworld, Dark, Matrix. Atleast on internet — where everything is code.
But as it improves, and it will improve, it will create so much nonsense content online, kill off many starter writing jobs, etc etc to the point where its effects will definitely kill off things that in the long run may be worth preserving.
To me the way to look at all of this is looking forward, not what happened last week. Just because a major illness hasn't killed you yet doesn't mean it won't in the future. Important to look at the trajectory and slope of the line of change.
Well, the chess kind [0] at least. Basically, you combine an AI and a human to get really amazing chess playing.
Same thing, but with all the image and chat AIs.
Example: You're a film student and are trying to work on costume design. You got no budget, but a lot of time and friends with some sewing machines and leftover Burning Man parts. You take a look at something like Jodorowsky’s Tron [1] and say 'Holy crap, that head gear is super wild and I know how to make something like that!' So you pay $10 for MidJourney, get Jodorowsky’s Tron going, put in the supplies you got, and you get some stuff you can actually make. Not only can you get the outlines of the hats and head gear, but you also get an idea of the lighting, the lenses, the farming, etc. You get the cool aesthetic you're looking for and how to actually film it too. You get the A, natch.
Sure, you can work it out with pencils and felt beforehand, just like they were doing in 2010. But now? You've shortened the time, you can play with the lighting and lenses and framing, you can change the actors out, etc. You'd have gotten there by yourself eventually, but with the AI you've shortened the time and exhaustion.
I'm super pumped for the creativity that AIs will unleash. Let's use the tools and FLY!
These were things that were extremely common in STEM classes when I was in school and were rare in English classes.
This author/professor has a very solid point that English instructors being less lazy and grading on the process instead of the final result would go a long way in terms of combatting cheating and producing better writers.
Another issue I remember was humanities professors at the college level expecting papers to be padded out to X pages even when you didn't need that length to make a strong argument or prove a thesis. The professor wants extra pages of BS to get the good grade, why not let ChatGPT take your 10 page paper that includes all the correct information and pad it out to 25 pages of redundancy and repetition to meet the weird requirement? I think in some cases it might have been hard to resist if these programs had been around, because I always resented having to write that extra nonsense in a class I was taking as a required elective.
It can create a believable made up process too. If someone wants to cheat themselves out of homework the only thing that could stop them is openai increasing the price after mass adoption.
However, it is important to note that ChatGPT is not a replacement for human creativity and originality. The creative and unique aspects of writing are still the domain of the writer, while the tool is simply a tool to assist in the more mundane tasks. Nevertheless, as technology continues to advance, it is likely that we will see more and more writers turning to tools such as ChatGPT to improve their productivity and efficiency.
*GPT helped me write this too, in the style of a Wall Street Journal article.
I just asked it to write an article about how things are different for CIOs today. It's a bit shallow and formulaic although I'm not sure it wouldn't pass muster as a column in one of the trade rags. However, I could probably take it, rewrite it a bit, flesh things out, maybe add a quote or 2, link to a survey, etc. while maintaining the basic structure and it would be very publishable in a lot of places. And I bet I could do that in an hour or two.
What it will do to management? Can ChatGPT pass an "inbox test" yet? That's a test for managers, where you get a background document and a queue of incoming messages. You have to reply to each message, in writing, in order. Someone who uses inbox tests needs to try that.
In my day to day life, I can’t think of a single thing that ChatGPT does better than a search engine and I can better judge the trustworthiness of the sources.
On the other hand, it does a great job writing working Python code based on plain language requirements. It does at least as well as a junior dev for run of the mill scripting.
No most development is not reversing a binary tree while juggling two bowling balls and riding a unicycle on a tightrope.
"... the kind of tweaking you see me doing above, tweaking which I can do because I possess a relatively sophisticated understanding of narrative craft and know what to tell the AI to do"
This will likely entail teaching them how to do the job without the use of the AI (as was always the case). (The same applies to software development [which is just a different form of writing]).
Sounds pretty much like everyone I know.
ChatGPT output is often so absurdly verbose it becomes tiring to try and make sense of the text (when the text makes sense in the first place). The effort we're willing to put into big swaths of text is already low, and I fear ChatGPT will only make it worse.
Supposedly you can summon Anubis with the right prompt as well.
But nobody does.
However, I can see a media outlet incorporating a set of "checklist" prompts in their publication process. Any GPT-N generated article will go through them before the article is pushed out. One of these checklist prompts can be about reducing verbosity.
I'm not familiar with the US education system. Is that last point true, or sarcasm?
I'm not familiar with this exam but if I said "Flowers represent fertility (because they are pollinated)" and a classmate says "Flowers represent gravity (since they fall)" it shouldn't really be evaluated for accuracy, just how well we made our argument since we're not really discussing facts.
It's about SWARMS of AI chatbots amassing karma points, likes, and engagement, and at any given moment they can overwhelm any community and gang up on people whose viewpoint differs from the one they are programmed to promote.
I want my warm, sunny hill where I can discuss philosophy if I so choose with some bread wine, after which I'll play old arcade games and then work on a classic car, after which I'll make love to my wife.
Can you ask chatGPT if it can arrange for that, please?
* wakes up *