a thoughtful web.
Good ideas and conversation. No ads, no tracking.   Login or Take a Tour!
comment by veen
veen  ·  405 days ago  ·  link  ·    ·  parent  ·  post: From Bing to Sydney

Aight, let’s talk use cases then.

an incomplete list of things I have used gpt3 for that I find useful

- automate menial text related tasks in a time efficient manner. “Change these 200 rows of code so that they all match this format”

- find bugs in my code that I’m too dumb/chase stared at too long to notice

- explain this complex text in a way that I understand, e.g. an academic paper, or difficult code. Answer follow up questions I have about it too.

- write the excel formula to do X Y and Z, a combination of things that is hard to google

- take my unstructured mess of a meeting notes, summarize the main three points and write an email to my colleagues sharing that information.

- find the sentiment of this wall of text I received from someone

and, perhaps most often used by me:

- write the shitty first draft of (code / email / report / …) so that I can use my creative energy only for the important stuff

Any of these tasks can be done by someone. It’s not the second coming of Christ by a long shot: I think of it more like 3D printing. But the ability to automate simple human knowledge work taps a well that is in my opinion very deep and very wide. It genuinely is a timesaver and I am already throwing a few bucks at it per month. None of the things on my list are interesting, but they are a part of the work that need to happen and I’ll happily offload that and more to an uninspiring but useful bot.

To make a point by asking a question: ChatGPT is the fastest new adopted thing ever in terms of DAU. It is constantly at capacity despite having Microsoft behind it. If it was just a fad, why are so many people still using it so much you think?





kleinbl00  ·  405 days ago  ·  link  ·  

I think thee and me have slightly different definitions of "tool." You have a list of seven things there that basically boil down to "reformat." reformat 200 rows of code, reformat errant code, reformat complex language, reformat an excel formula, reformat your meeting notes, reformat bad writing into not-bad writing.

That's useful! Don't get me wrong! "Hey Siri code me a piece of middleware that overlays the AVS-B traffic from my Pi-Aware into something my Geochron can speak" is the sort of thing I look forward to doing. Could I do it now? theoretically. Pretty sure there's some deal-breakers in there. If I march through it like an expert I'll find them. If I "hey Siri" my way through it I'm going to get a dozen flavors of "I'm sorry I can't do that Dave."

3d printing is an interesting example. Let's run with it.

I spent two days, off and on, dealing with this piece-of-shit model. I did it because my wife dropped the trimmer and broke that lever. "Hey no problem this is exactly the sort of bullshit some twat wasted too much time on for internet points," I said to myself and two minutes later that thing was spooling out in PETG.

- only for me to discover that it's a spindly piece of shit just aching to break.

- So I spent five hits in Fusion (Autodesk: "we'll give you a free CAD package and it will only make you hate us more") iterating a little bit of strength into it.

- Only to evaluate that the only places you can add thickness interfere with its function.

- By which time the glue had cured on the broken part, which isn't made out of laminated cheez-whiz.

It's still cheez whiz! That part is injection-molded polystyrene, what the cool kidz call 'ABS' when they order it in 5lb rolls. PET is stronger than ABS. Yet because of the manufacturing process, a broken, glued part works better.

3d printers are tools. They have utility. But when all you have is a hammer the world is made of nails.

Another analogy - all your friends MiG weld. Why? It's stupid-simple. It requires no skill at all. And the tools are cheap. And if you're just gobbing aluminum together it'll probably do. But it's not your best bet for steel, not by a long shot. And it's not your best bet for strength, not by a long shot. And it's not your best bet for anything with any thickness, not by a long shot. I've mig-welded once; it's like using a hot-melt glue gun. Me? I learned how to do aluminum with a buzz-box made-it-ourselves TiG welder with no pedal. I learned how to do angle iron with baling wire. I learned how to do boilerplate with 3/32 TerrorRod back before welding masks were clear until their batteries told them not to. Most people didn't. Most people look at this and think "WIZARDRY"

...but in my family you didn't get respect until you could do that with gum wrappers.

So your use cases, with a couple exceptions, boil down to "getting around the inefficiency of the shitty UI modern tech companies have saddled us with." Looking movie times up in the paper is inefficient and asking Siri is faster (unless Siri is confused about what year it is and gets combative). Will it allow a segment of humanity to do things they couldn't have done before? Incrementally, yes. Why does OpenAI have more "daily active users" than any other web company? Because it's Bitmoji 2023. Joanna Stern really wanted to write a "holy shit search has changed forever" article but what she ended up writing was a "huh if you ask it anything complicated it lags for 30 seconds and then gives you marginally useful data. "

let's talk about why that matters

I see a tool, you see a revolution. To me it's a tool that marginally improves the UI of whatever it's matched to. To you it's an efficiency booster for writing. What the tech companies see, though, is

Hey Siri be my friend

And don't for a minute think they're considering anything else.

As exemplified in the article above, chat engines drive more affinity than search engines and affinity is the only thing keeping you from bailing for another webpage the minute you're not happy with their results. They'll put it on phones soon and suggest you leave the mic on all the time so they can model you better -for your own good, of course. And hey - it'll probably give you better results than if you didn't.

But it won't ever give you better results than if you know what you're doing.

It's really telling, in my opinion, that none of the chatbots have given a shit about accuracy so far. It doesn't matter. It's not within the business model. What matters is engagement ("the fastest new adopted things ever in terms of DAU").

Ask Facebook - does anything drive engagement better than an argument? Mark my words - there's code in OpenAI that stochastically fucks up just to make it less creepy. We just haven't found it yet.

kleinbl00  ·  405 days ago  ·  link  ·  

This, more than anything, is probably why I'm less impressed by ChatGPT than anyone else - I'm pretty good at wordz. Faking any register of speech just requires a good ear and mine do the job.

I'm a shitty fuckin' pianist, tho. 'sokay, so's Brian Eno. He's famous for saying he prefers to work with people who don't know how to play because they're more improvisationally free. I don't know that I agree but I can see his point: there are things you do with instruments because everyone does that with them, everyone teaches you to do that with them. It depends on what you want to do.

Korg came out with a synth called the Karma some ten, fifteen years ago. It has an improvisation engine on it. You play a chord in a tempo and it'll give you an improvised bassline to go with whatever your right hand is doing. It's kinda cool? but it also makes everything sound like you're playing the Holiday Inn Lounge on Turnpike 80. Which, let's be honest, is what you're doing. You're not cutting new hits with your Korg Karma you're in a '70s cover band that plays for beer money and now your chops sound like you minored in music (and majored in pipefitting) rather than giving up on piano lessons in 10th grade.

    - write the shitty first draft of (code / email / report / …) so that I can use my creative energy only for the important stuff

The reason language breaks up into registers is to exclude segments of society. You don't write in legalese because it's fun, you write in it so that everyone knows you're a lawyer talking to lawyers and judges. It's protective coloration for the payload of what matters, which is why lawyers kept Lotus Word Pro alive long past the point where everyone else switched to Microsoft: it's got a macro function that lets you dump big stupid blocks of legalese into whatever you need whenever you need it. Other professions didn't peacock feather their language nearly so hard as lawyers so they just adapted to whatever's easiest which is usually what comes in the box.

The easier it is to mad-lib this stuff, the more professions with the need to protect their expertise behind register will resort to complication and metaphor. If AI is what makes it easy and accessible, over time the register will shift to be increasingly hostile to AI.

veen  ·  404 days ago  ·  link  ·  

I agree, it makes sense for Microsoft to aim for the search big bucks and if it means making Confidently Incorrect ClippyTayGPT they'll do it. My guess is that we'll look back in a few years and say "remember when they introduced LLMs to the world with Sydney? LOL".

But I'd also wager that these short-term affinity hunts won't matter; the cat is out of the bag, Bing/Sydney feels to me like the Napster of tomorrow, and I'm more interested in what the Spotify of LLMs looks like. It might be better to discuss around an extended version of what I wrote earlier:

    Any of these tasks can be done by someone who knows what they're doing

...as I think that summarizes where we agree. I hadn't thought of these AIs as the erosion of skill yet but that is indeed what it is: you don't need to know the details, here's a mediocre version that can pass, maybe. It draws a line in the sand where on one side of the line you have people enough in the know to see the difference, and on the other hand you have the clueless who gaze upon the MiG weld in astonishment. Given time and text to churn through, LLMs will inevitably improve and push that line a bit further. Or a lot further. Hard to tell now.

I'm wondering, though, how many people's work will end up at the wrong side of the line. How much white collar work boils down to 'refactor information'? A cousin of mine attempted a job at an investment firm. She'd spent all her day reading patents and applications, and her job was to summarize that (technical, jargon laded) information for the investors to judge easily. Her PhD was needed to apply, yet that job is already on the chopping block with the current models because ChatGPT is almost as good a UI to academically based patents than a PhD-carrying human is.

I'm inclined to draw a parallel that feels trite, so shoot it down if you see the flaw with it: isn't the march of progress almost inevitably towards higher abstractions that can be used by more people? Chris Sawyer famously wrote Transport Tycoon and Rollercoaster Tycoon in Assembly. Today you whip up a generic Game World™ in Unity 5 and plop down assets from a library. It's easier, it's much more accessible, and it almost certainly performs for the worse. My excitement for Felt comes from its fantastic UI for GIS. It makes a set of abilities accessible to a broader audience. At the same time, my old uni has whittled geomatics down from a full department to a small graduate degree years ago because who the fuck needs to know how satellites work when you want to draw five X's on a map?

Fast forward a few years. Oh, shoot, nobody knows how it actually works anymore. Everything is mediocre now. There are more people who can create, but there are fewer people who can do so at the highest level of skill. You need subtitles because nobody does proper audio anymore, your online map doesn't quite put the X where you intended it, your Unity game isn't that fun but it's passable, your AI generated text has some weird phrasing but it does summarize 240 pages of technical patent jargon in less time than it takes you to get out your wallet.

kleinbl00  ·  404 days ago  ·  link  ·  

    I'm wondering, though, how many people's work will end up at the wrong side of the line.

LOTS

Once upon a time there were typing pools. Executives didn't write anything, they dictated and girls (always girls) typed it up. Those girls didn't make a living wage, not really, they were in it until they found husbands who were often among the executives they typed for. It was a social construct more than anything and reflected that careers were things that women held temporarily or ephemerally based on the employment of their husbands (Elizabeth Warren and her daughter wrote a whole book on this).

I'd have a bookkeeper except we have Quickbooks. I'd have two or three medical asisstants except we've got EHRs. I'd have an IT guy except I can roll my own. etc etc etc. I helped a buddy's firm not hire another machinist because a robot allows them to run lights-out over the weekend - they have enough of an assembly line that they can spend an hour loading pallets on Friday and come in Monday morning to 150 parts. The machinist would be prolly $80k salary plus 50% night premium times 50% benefits is $180k a year. The robot is $150k amortized over three years.

The VCs want us to call this "disruption" but Schumpeter called it "creative destruction" and Marx & Engels just called it "capitalism:"

    Modern bourgeois society, with its relations of production, of exchange and of property, a society that has conjured up such gigantic means of production and of exchange, is like the sorcerer who is no longer able to control the powers of the nether world whom he has called up by his spells. ... It is enough to mention the commercial crises that by their periodical return put the existence of the whole of bourgeois society on trial, each time more threateningly.

And I mean, Engels saw capitalism at its worst - Manchester workhouses in Victorian England. He was unimpressed, as any thinking human would be. Automation and mechanization destroyed the middle class wholesale and England did fuckall about it for the better part of 80 years. Reason number whatever as to why Steampunk sucks: it celebrates a truly wretched period of fabrication where everything was of grindingly poor quality. Reason number whatever as to why bullshit cosplay millennials love Steampunk: the aesthetic is "bodge a bunch of bullshit together with utter disregard for design and then weather it as if you don't fucking care about anything in your life."

You wanna see where this ends up?

Design up to the '90s was intentional. Design in the '90s was My Very First Desktop Publishing Program. At the time? Nobody gave a fuck who came up with that nothingburger of a pattern. Solo has no fucking idea. But Millennials, who grew up in an era of fuck-all-designers, fixated on shitty fucking design and now there's a 6000 word article speculating as to who came up with this dialtone of a pattern. That's Steampunk: Contemporary Edition and it's what happens when expertise gets driven out in favor of mash-the-keys-make-a-riff.

A personal anecdote: Once upon a time there were post sound departments. You'd shoot your movie or TV show, you'd edit it, and then the sound department would make it sound good. Then editing became a fast'n'reckless process because everyone shot digital which meant you went from a 5:1 or 3:1 shooting ratio to a 500:1 or 1500:1 (or worse: I boomed three minutes of commercials once for which we rolled on 72 hours of footage) shooting ratio and the edit became more of a winnowing process than a creative one. I paid for my Pro Tools PC (and my crypto!) with proceeds from my resale of my bought-at-foreclosure portion of Todd-Soundelux, 500 people who largely never worked again. About that time Izotope came out with Total Mix, a software package for editors (they hid its very existence from their sound betatesters) that would ostensibly make the garbage editors crank out sound marginally acceptable.

It failed miserably, Izotope withdrew it from the market and promptly scrubbed any mention of it from the Internet but us mixers remember. We're the only ones, and there are only a handful of us left because left with the choice between slowing the fuck down and hiring professionals or continuing to pump out garbage and subtitle it so that the audience can understand their shitty work product, production companies went "subtitle the shit out of it we'll convince the world they like it."

And here's the thing - there were a few designers in the '90s. There are sound mixers who survive. Not everything in the Victorian era was a shitty bodge, just most of it. All that really happens is everyone gets marginally better access to something and the general quality gets profoundly worse. This isn't a new thing, this isn't a victorian thing, this is "creative destruction" or "capitalism" because lowest price will always win and the overwhelming majority of consumers don't care enough to reward craft so you either pump out mediocre garbage or you hope you can make a living as an "artisan."

This, by the way, is the fundamental argument of Fully Automated Luxury Capitalism. Societies have a choice: widespread catastophe for the formerly employed or a vastly-expanded social safety net to stave off unrest. There's a reason the USA rebuilt Europe with vastly more social reforms than we had ourselves - the place was ready to tip into communism or barbarism at the drop of a hat so you'd best make sure that nobody had a reason to riot in the streets.

We're there. We're there again. And yes. LLMs are going to surprise a lot of people with how replaceable they are. And if there's a choice between "but I do it great" and "but the robot does it just barely well enough to bother but it does it for nearly free", the robot is taking your job. When quality is expensive and mediocre is cheap, worse-than-mediocre becomes the baseline.