Fun new paper. This kind of thing really has come along over the past year or so.
Re: Thing: Artificial Intelligence
321. sounds better than steely dan.
2. is the narrator an AI bot too? just talk normal.
2. is the narrator an AI bot too? just talk normal.
Re: Thing: Artificial Intelligence
33I'm deeply ambivalent about it. As stated upthread, combined with the flavour of libertarian capitalism that's popular in the circles of the people funding it it has nightmarish potential. I can't help but be impressed though that you can verbally describe a hardware design to it and it will produce almost fully functional firmware based on your specification:
Pretty incredible, and certainly not just an "overgrown spell checker". I can't remember if it's mentioned in that video but probably one of the highly skilled jobs for humans in the coming while is likely to be "prompt engineer" - I.e. knowing how to ask the correct questions in the correct way to get AI to do boring work faster than a human would do it.
Pretty incredible, and certainly not just an "overgrown spell checker". I can't remember if it's mentioned in that video but probably one of the highly skilled jobs for humans in the coming while is likely to be "prompt engineer" - I.e. knowing how to ask the correct questions in the correct way to get AI to do boring work faster than a human would do it.
FM formerly known as tarandfeathers
https://github.com/DunningtonAudio - open source projects
https://github.com/DunningtonAudio - open source projects
Re: Thing: Artificial Intelligence
34All this means is that humans put thousands of hours into creating the prompts and the coded the ability for the "AI" (in quotes because it is not intelligent) to follow them. It's just not impressive, nor incredible.Daniel wrote: Tue Mar 21, 2023 7:00 pm I'm deeply ambivalent about it. As stated upthread, combined with the flavour of libertarian capitalism that's popular in the circles of the people funding it it has nightmarish potential. I can't help but be impressed though that you can verbally describe a hardware design to it and it will produce almost fully functional firmware based on your specification:
Pretty incredible, and certainly not just an "overgrown spell checker". I can't remember if it's mentioned in that video but probably one of the highly skilled jobs for humans in the coming while is likely to be "prompt engineer" - I.e. knowing how to ask the correct questions in the correct way to get AI to do boring work faster than a human would do it.
Re: Thing: Artificial Intelligence
35AI?
A broad subject, with implications that go well beyond the domain of faddy apps.
Suffice it to say, I'll try to keep a dog around the house, as they're the only ones who seem to be able to suss out terminators from real people.
*bark* *bark* *BARK-BARK-BARK*
Don't say eighties cinema didn't try to warn you.
A broad subject, with implications that go well beyond the domain of faddy apps.
Suffice it to say, I'll try to keep a dog around the house, as they're the only ones who seem to be able to suss out terminators from real people.
*bark* *bark* *BARK-BARK-BARK*
Don't say eighties cinema didn't try to warn you.
Re: Thing: Artificial Intelligence
36It's just number crunching, but it's the next level of number crunching.
Who would have thought that there was infinite fractal geometry hiding in f(z)=z^2+c? We needed computers before we could see it.
Now we have computers crunching huge databases and revealing more patterns. It's all very interesting. It's a different, potentially powerful lens you can view the data through.
It only resembles humans to the extent that it reflects human-made training data, or to the extent that humans resemble pattern-matching robots, which may be more than we like. But humans anthropomorphisise EVERYTHING, and project meaning onto everything, so it doesn't even have to be all that good to spook people.
Who would have thought that there was infinite fractal geometry hiding in f(z)=z^2+c? We needed computers before we could see it.
Now we have computers crunching huge databases and revealing more patterns. It's all very interesting. It's a different, potentially powerful lens you can view the data through.
It only resembles humans to the extent that it reflects human-made training data, or to the extent that humans resemble pattern-matching robots, which may be more than we like. But humans anthropomorphisise EVERYTHING, and project meaning onto everything, so it doesn't even have to be all that good to spook people.
That's not how neural networks work. The whole reason they are interesting and not just a regular computer program is because they are self-learning. You don't program them, you train them.All this means is that humans put thousands of hours into creating the prompts
Re: Thing: Artificial Intelligence
37I think it's fairly incredible to have a tool that can parse billions of lines of code from github, and reassemble them into something functional based only on plain English prompts, then to iterate revisions to fine tune the behaviour. As someone who has to write code occasionally but has no real interest in being a programmer I think that's a pretty exciting prospect and potentially a massive timesaver. I can well see that if one knew how to drive it well it could eventually turn a job of several hours into one of several minutes, leaving a human several free hours to, for example, fuck around debating AI on an Internet forum.
I don't think anyone is saying it's currently equal to or better than human intelligence, and it's worth bearing in mind that emulating human intelligence would be a pretty pointless end goal for an artificial intelligence project. We already have human intelligence, it will be more interesting to have something different.
I don't think anyone is saying it's currently equal to or better than human intelligence, and it's worth bearing in mind that emulating human intelligence would be a pretty pointless end goal for an artificial intelligence project. We already have human intelligence, it will be more interesting to have something different.
FM formerly known as tarandfeathers
https://github.com/DunningtonAudio - open source projects
https://github.com/DunningtonAudio - open source projects
Re: Thing: Artificial Intelligence
38This.Daniel wrote: Tue Mar 21, 2023 10:35 pm I don't think anyone is saying it's currently equal to or better than human intelligence, and it's worth bearing in mind that emulating human intelligence would be a pretty pointless end goal for an artificial intelligence project. We already have human intelligence, it will be more interesting to have something different.
at war with bellends
Re: Thing: Artificial Intelligence
39FWIW this is what one of the coders on our team had to say:
It'll be a while (maybe a short one) before you can ask it to fix delicate bugs or improve the handling of something - anything a bit "squishy" in concept I reckon. I suspect there'll be a period where some of the work available will be to take the output from some AI thing and tweak it but, you know, anyone offering that can politely go fuck themselves
I tried using GPT-3 last week for a code thing (some projectile function) and it gave me broken answers 5 times in a row, even as I was trying to help it. The most right answer did a bunch of trig before converting it back to a vector. I said it could work directly with vectors and it agreed, did almost the right thing and then didn't use it in the value it returned. I don't truly understand what's going on but it definitely feels like its fed on a diet of stackoverflow and reddit and shit in = shit out.
It'll be a while (maybe a short one) before you can ask it to fix delicate bugs or improve the handling of something - anything a bit "squishy" in concept I reckon. I suspect there'll be a period where some of the work available will be to take the output from some AI thing and tweak it but, you know, anyone offering that can politely go fuck themselves
I tried using GPT-3 last week for a code thing (some projectile function) and it gave me broken answers 5 times in a row, even as I was trying to help it. The most right answer did a bunch of trig before converting it back to a vector. I said it could work directly with vectors and it agreed, did almost the right thing and then didn't use it in the value it returned. I don't truly understand what's going on but it definitely feels like its fed on a diet of stackoverflow and reddit and shit in = shit out.
Re: Thing: Artificial Intelligence
40I do not code by trade, but I know enough to write simple automation scripts that help me at my job and ChatGPT-3 has helped me get through a few hurdles in implementing and expanding some of those. Some of it has been near perfect, but some of it I've had to correct (so you still have to know what to correct). If you know how to prompt it and understand its limitations, it can be useful. As I am looking for a new job at the moment, I've also played around with it writing cover letters based on job descriptions. That feels a bit wrong, but it's interesting. It overpromises for sure (I'm not fluent in multiple languages, for one).
It doesn't super excite me or worry me yet in terms of disruption, but I see why people are starting to make a fuss about it.
It doesn't super excite me or worry me yet in terms of disruption, but I see why people are starting to make a fuss about it.