|
I like this quote. We used some ML at work. Without giving too many details, we were trying to stand up a parallel pipeline entirely from scratch. Part of that process involved curating 1-2 million entries. The manual pipeline had gotten to where it was after years and years of manual data entry, corrections, etc. We got the automated pipeline up and running in 1-2 years with a 98% automation rate thanks to ML, we basically automated dozens of data entry people doing that work for years, and ended up with a much cleaner data set. We then started using the same techniques to clean up data from the manual pipeline. So yeah, those manual data entry people probably should be worried about their jobs. But we'll need more and more developers to build stuff like this and glue all the tech together in a way that makes sense, and AI seems nowhere near being able to understand the context of an individual corporation.
|
#
?
Dec 6, 2022 04:22
|
|
|
|
| # ? May 27, 2024 09:26 |
|
|
Used chat gpt to rewrite and design the CSS for my lovely blog. A little back and forth, some minor edits, and I'm pretty happy with it. I'm not worried about gpt coming for my job, which at this point is primarily around making sure other people don't do incredibly dumb things. I can easily see how it could reduce time and number of people needed for product development. But I don't see it driving that number to zero. Or handling any of the meaty problems that are actually interesting to work on.
|
|
#
?
Dec 6, 2022 04:33
|
|
|
just finally logged into ChatGPT to experience the madness myself and it seems to be falling over making me wonder if my particular AI-generated Sonic the Hedgehog movie scripts were simply too compute-intenstive for the cloud
|
|
#
?
Dec 6, 2022 04:44
|
|
|
luchadornado posted:So yeah, those manual data entry people probably should be worried about their jobs. The manual data entry work seems to me to be extremely similar to the labeling work required to create ML or to spot-check/QA while maintaining it. The more successful ML/AI models become, the more domains they can possibly be applied towards and still succeed, so the reduction in human effort per task will be substantially offset by the increase in the overall domain. Even unsupervised and active learning techniques seem to just leverage proportionately fewer bits of human work per task but it doesn't eliminate that human work entirely. This can significantly reduce the cost per model (because far less human input required), but this just seems to me that it would result in a lot more areas would find applying ML tools to be cost effective. I don't think the tech alone is able to determine whether more or less of this sort of work gets done overall (though it will determine how much work is required per project.) Rather I think the overall economic cycles are likely to determine employment trends-- but that's not exactly any sort of revelation as it's basically always been true. To bring it back to the top, yeah, they probably should be worried about their jobs *right now* because there's all sorts of noise about a possible coming recession, but I'm not sure they should be extra-worried just because of technological progress.
|
|
#
?
Dec 6, 2022 05:05
|
|
|
Mega Comrade posted:
Just as long as the picture of the multicultural board room doesn't show people's hands. None of the image generators can create a hand that looks even vaguely realistic. Like, there are always too few or too many fingers (sometimes crazy numbers of fingers) or other anatomical grotesqueries. I've played around with ChatGPT. One thing I fiddled with was creating image prompts for internal decorating, since that was suggested as an application. And it kinda gives you the same thing even if you vary the phrasing of the prompt by quite a lot. I even tried it out for something work-related today -- helping a student find a particular type of genomic data to answer a particular type of question arising from some analyses they'd done. ChatGPT gave some answers that were more-or-less correct, and pretty close to the off-the-cuff answer I gave the student. "Consortium X and Consortium Y are generating this kind of data. You can also look at Big Unstructured Data Repository Z". But I found a way more appropriate specialized database that aggregates all of that data in an easily-searchable way, and it was just the top hit in the right Google search. I was kind-of impressed, and it helped jog my memory a little, but the thing essentially seems to be a kind of Stack Exchange paraphraser.
|
|
#
?
Dec 6, 2022 08:28
|
|
|
oliveoil posted:Year 1: better than the interns Again I'll bring up Tesla. Getting Tesla to be able to drive like an idiot was fast, getting it to drive actually safely and reliably still hasn't happened, 6 years after it was promised despite billions of dollars poured into it. When will it drive like a professional? Just because dalle 1 to 2 was a big jump does not guarantee 2-3 will be. With every iteration, the next one is greatly more difficult to achieve. And id still argue, most interns would know 41 is not an even number, so we aren't even at year 1 yet. Going back to your original question though, what do we do, in the worst case scenario where these AI programs actually improve at the ridiculous speed you fear? How do we keep ourselves relevant? We don't, nothing you can do, if they can take programmer jobs like you worry, they can take 90% of jobs in society, at which point society has either switched to universal basic income or collapsed under majority unemployment and poverty. Lead out in cuffs posted:I was kind-of impressed, and it helped jog my memory a little, but the thing essentially seems to be a kind of Stack Exchange paraphraser. If it's answers are being posted to stack overflow and it's using stack overflow as a source of truth...then how long until it's feeding it's own answers to itself? Mega Comrade fucked around with this message at 10:09 on Dec 6, 2022 |
|
#
?
Dec 6, 2022 08:55
|
|
|
|
|
#
?
Dec 6, 2022 14:33
|
|
|
Mega Comrade posted:If it's answers are being posted to stack overflow and it's using stack overflow as a source of truth...then how long until it's feeding it's own answers to itself? There was a post linked to on Hacker News a few days that I can't seem to find, that stated that this self-feedback loop was already present and a drawback. While there is crystal ball gazing going on, this whole conversation reminds me a bit of Stephenson's The Diamond Age. Society at large was fine with genericized, cheap output of any type, but there was an entire society of well-off people (the Vickies) who based their cultural identity around the "realness" of organic output. IIRC, one of the points the book was trying to make was that truly significant societal change was benefited by "real" influences.
|
|
#
?
Dec 6, 2022 14:40
|
|
|
luchadornado posted:There was a post linked to on Hacker News a few days that I can't seem to find, that stated that this self-feedback loop was already present and a drawback. Stephenson needs to get it though his head that the message he thinks people will walk away with is not the only message people could walk away with. He's a dystopian prophet; as remarkable as it is sad to me that people herald the coming of the dystopian parts of his work as progress.
|
|
#
?
Dec 6, 2022 14:46
|
|
|
https://twitter.com/ostwilkens/status/1598458146187628544 https://twitter.com/ostwilkens/status/1599026699999404033
|
|
#
?
Dec 6, 2022 16:47
|
|
|
redleader posted:unlike all the blub programmers out there, i (as a 10x engineer) have nothing to worry about show me an ai that can debug code and then maybe i'll get worried.
|
|
#
?
Dec 6, 2022 18:00
|
|
|
Bruegels Fuckbooks posted:show me an ai that can debug code and then maybe i'll get worried. https://youtu.be/m0b_D2JgZgY
|
|
#
?
Dec 6, 2022 18:45
|
|
|
Jabor posted:Hmm. Maybe these developments are concerning to the sort of people that have one year of programming experience ten times over and haven't progressed beyond that. Which definitely doesn’t describe me. 
|
|
#
?
Dec 6, 2022 19:36
|
|
|
Doesn't describe me either I'm only around 9 years right now.
|
|
#
?
Dec 6, 2022 20:09
|
|
|
oliveoil posted:Year 1: better than the interns
|
|
#
?
Dec 6, 2022 22:21
|
|
|
First the AI replaced the Interns, and I did not speak out - because I was not an Intern. Then it replaced the Junior Engineers, and I did not speak out - because I was not a Junior Engineer. Then it replaced the Senior Engineers, and I did not speak out - because I was not a Senior Engineer. Then it replaced me - the 10x Engineer - and there was no one left to speak for me.
|
|
#
?
Dec 7, 2022 04:22
|
|
|
If you're a real 10x engineer then how you can't just negotiate for 9x the comp by threatening to quit?
|
|
#
?
Dec 7, 2022 04:47
|
|
|
Bruegels Fuckbooks posted:show me an ai that can debug code and then maybe i'll get worried. it won't need to, since an ai will write the code properly in the first place
|
|
#
?
Dec 7, 2022 06:16
|
|
|
redleader posted:it won't need to, since an ai will write the code properly in the first place All errors are caused by the users, never the code.
|
|
#
?
Dec 7, 2022 13:16
|
|
|
JawnV6 posted:if you believe any horseshit someone fits an exponential plot to, or in this case wholly invented yourself without assistance, how'd you escape the Kurzweil cult Also I don't base my concern on the naive exponential kurzweil cultist thought process. I can personally think of several different algorithms that would likely allow chatgpt to provide the outputs of smarter people. Doubt any one of them would take more than a few months of work for somebody experienced with model design and development. And if I can think of a few then someone at openai or deepmind probably already has a much longer prioritized list of candidate algorithms to try.
|
|
#
?
Dec 7, 2022 15:58
|
|
|
oliveoil posted:Also I don't base my concern on the naive exponential kurzweil cultist thought process. I can personally think of several different algorithms that would likely allow chatgpt to provide the outputs of smarter people. Doubt any one of them would take more than a few months of work for somebody experienced with model design and development. And if I can think of a few then someone at openai or deepmind probably already has a much longer prioritized list of candidate algorithms to try. https://twitter.com/EigenGender/status/1600099152804777984
|
|
#
?
Dec 7, 2022 16:46
|
|
|
oliveoil posted:Also I don't base my concern on the naive exponential kurzweil cultist thought process. I can personally think of several different algorithms that would likely allow chatgpt to provide the outputs of smarter people. Doubt any one of them would take more than a few months of work for somebody experienced with model design and development. And if I can think of a few then someone at openai or deepmind probably already has a much longer prioritized list of candidate algorithms to try. Oh sure. Be sure to phrase your algorithm in terms of a probability graph to possible next words, because that’s the entire gadget. I’m sure we can back-fit facts into the thing, nbd
|
|
#
?
Dec 7, 2022 17:04
|
|
|
It's not rocket science. It's black boxes of correlation. How hard can it be to feed one into another once you've identified the key factors that need to be correlated? E: quote:a probability graph to possible next words, because that’s the entire gadget Okay fine that kinda destroys all my ideas but it did give me another one. oliveoil fucked around with this message at 17:21 on Dec 7, 2022 |
|
#
?
Dec 7, 2022 17:12
|
|
|
oliveoil posted:It's not rocket science. That's correct. Rocket science is simple enough that we can teach the results to high school students and derive them for sophomore physics majors. This is actually complicated.
|
|
#
?
Dec 7, 2022 17:21
|
|
|
oliveoil posted:It's not rocket science. It's black boxes of correlation. How hard can it be to feed one into another once you've identified the key factors that need to be correlated? oliveoil, after a brief moment approaching clarity and introspection, is back at it again. 
|
|
#
?
Dec 7, 2022 17:22
|
|
|
(Garfield walking past the "No Garfields" sign) "I wonder who that tweet is describing."
|
|
#
?
Dec 7, 2022 17:31
|
|
|
Dear Diary, today I got to poo poo on a guy in front of his boss because he stopped skimming my nerd essay halfway through and jumped to the opposite conclusions. That's my story thanks for reading
|
|
#
?
Dec 7, 2022 18:07
|
|
|
I realize that must feel good, but tbh I assume no-one reads anything and conclusions should be put in the first line. "TLDR" / "BLUF" (bottom line up front) is like a super-power when convincing people to do anything.
|
|
#
?
Dec 7, 2022 18:12
|
|
|
minato posted:I realize that must feel good, but tbh I assume no-one reads anything and conclusions should be put in the first line. "TLDR" / "BLUF" (bottom line up front) is like a super-power when convincing people to do anything. stop trying to make BLUF a thing - it's not happening
|
|
#
?
Dec 7, 2022 18:15
|
|
|
The more theories I see from confident people about this, the less I worry about my job prospects.
|
|
#
?
Dec 7, 2022 18:41
|
|
|
luchadornado posted:stop trying to make BLUF a thing - it's not happening It's standard in the military and will spread from there.
|
|
#
?
Dec 7, 2022 19:59
|
|
|
luchadornado posted:stop trying to make BLUF a thing - it's not happening
|
|
#
?
Dec 7, 2022 20:02
|
|
|
Literally remoting into a presentation to DoD and BLUF was on the 2nd slide. It had like 400 words on it and I did not read. But yes, the good way to write an email as someone who loves technical details is: - Write and sort out all the awesome technical poo poo - Go back to the top line and write your conclusion
|
|
#
?
Dec 7, 2022 20:18
|
|
|
you could assume that random lingo from the military would seep into techland in the 70s when half this poo poo was defense contracting. certainly you can still expect it if you work at lockmart, grumman, whatever. i dont think it can be expected in 2022 in generic tech companies
|
|
#
?
Dec 7, 2022 20:21
|
|
|
I still say 'foils' half the time
|
|
#
?
Dec 7, 2022 20:24
|
|
|
Answer First Communication
|
|
#
?
Dec 7, 2022 20:30
|
|
|
StumblyWumbly posted:Literally remoting into a presentation to DoD and BLUF was on the 2nd slide. I've been doing this for years in emails and get a lot of positive reinforcement from my bosses on this. Usually I'm reading through my wall-of-text email, get to the bottom, realize I've accidentally summarized the topic in the last line using the specific terminology sprinkled throughout the email, then cut->paste to the top Never heard of BLUF until moments ago though
|
|
#
?
Dec 7, 2022 20:38
|
|
|
JawnV6 posted:I still say 'foils' half the time stop multiplying binomials then
|
|
#
?
Dec 7, 2022 20:54
|
|
|
oliveoil posted:It's not rocket science. It's black boxes of correlation. How hard can it be to feed one into another once you've identified the key factors that need to be correlated? Funnily enough I have an idea! If it's not rocket science and not hard to do, then go do it! Welcome to programming.
|
|
#
?
Dec 8, 2022 05:37
|
|
|
|
| # ? May 27, 2024 09:26 |
|
|
downout posted:Funnily enough I have an idea! If it's not rocket science and not hard to do, then go do it! Welcome to programming. Funnily enough, I no longer have to: https://www.science.org/doi/10.1126/science.abq1158 The top 50% of competitive coders are probably way better at passing coding interviews than than the top 50% of successful candidates. AI can now code at a level close to good enough to get a junior job. No telling how close that is to being able to actually do a junior job but it ain't a good sign for juniors. oliveoil fucked around with this message at 16:36 on Dec 9, 2022 |
|
#
?
Dec 9, 2022 16:33
|
|