And i don’t mean stuff like deepfakes/sora/palantir/anything like that, im talking about why the anti-genai crowd isn’t providing an alternative where you can get instant feedback when you’re journaling
And i don’t mean stuff like deepfakes/sora/palantir/anything like that, im talking about why the anti-genai crowd isn’t providing an alternative where you can get instant feedback when you’re journaling
What I don’t like is that they’re selling a toy as a tool, and arguably as the One And Only Tool.
You’re given a black box and told to just keep prompting it to get lucky. That’s fine for toys like “give me a fresh low-quality wallpaper every morning.” or “pretend you’re Monkey D. Luffy and write a song from his perspective.”
But it’s not appropriate for high-stakes work. Professional tools have documented rules, behaviours, and limits. They can be learned and steered reliably because they’re deterministic to a fault. They treat the user with respect and prioritixe correctness. Emacs didn’t wrap it in breathess sycopantic language when the code didn’t compile. Lotus 1-2-3 didn’t decide to replace half the “7’s” in your spreadsheet with some random katakana becsuse it was close enough. AutoCAD didn’t add a spar in the middle of your apartment building because it was statistically probable after looking at airplane wings all day.
I mean software glitches all the time, some widespread software has long-standing bugs in it that its developers or even auditors can’t figure out and people just learn to work around the bug. Photoshop is made on 20 year old legacy code and also uses non-deterministic algorithms that predate AI (the spot healing brush for example which you often have to redo several times to get a different result). I agree that there’s a big black box aspect to LLMs and GenAI, can’t say for all AI, but I don’t think it’s necessarily inherent to the tech or means it shouldn’t be developed more.
Actually image AI is severely simple in its methods. Provide it with the exact same inputs (including the seed number) and it will output the exact same image every time, with only very minor variations. Should it have no variations? Depends; image gen AI isn’t an engineering tool and doesn’t profess to have a 0.1mm margin of error like other machines might need to.
Back in 2023 already China used an AI (they didn’t say what type exactly) to blueprint the electrical cabling on a new ship model, and it did it with 100% accuracy. It used to take a team of engineers one year to do this and an AI did it in 24 hours. There’s a lot of toy aspects to LLMs but this is also a trap of capitalism as this is what tech companies in startup mode are banking on. It’s not all neural models are capable of doing.
You might be interested that the Iranian government has recently published guidelines on AI in academia. Unfortunately I don’t have a source as this comes from an Iranian compsci student I know, they say that you can use LLMs in university but need to note the specific model used, time of usage, and can prove you understand the topic then that’s 100% clean for Iranian academic standards.
Iran is investing a lot in tech under heavy sanctions, and making everything locally (it is estimated 40-50% of all uni degrees in Iran are science degrees). To them AI is a potential way to improve their conditions under this context, and that’s what they’re exploring.
Do you have a link to the story? I ask because AI is a broad umbrella that many different technologies fall under, so it isn’t necessarily synonymous with generative AI/machine learning (even if that’s how the term has been used the past few years). Hell, machine learning isn’t even synonymous with neural networks.
Circling back to the Chinese ship, one type of AI I could plausibly see being used is a solver for a constraint satisfaction problem. The techniques I had to learn for these in college don’t even involve machine learning, let alone generative AI.
I sent the story on perplexity and looked at its sources :P (people often ask me how I find sources, I just ask perplexity and then look at its links and find one that fits)
https://asiatimes.com/2023/03/ai-warship-designer-accelerating-chinas-naval-lead/ they report here that a paper was published in a science journal, though Chinese-language.
I did find this paper: https://www.sciencedirect.com/science/article/abs/pii/S004579492400049X but it’s not from the same team and seems to be about a different problem, though still in ship design (hull specifically) and mentions neural networks.
This is sort of the issue with “AI” often just meaning “good software” rather than any specific technique.
From a quick read the first one seems to refer to a knowledge-base or auto-CAD solution which is fundamentally different from any methods related to LLMs.
The second one is some actually really impressive feature engineering used to solve an optimization problem with Machine Learning tools, which is actually much closer to a statistician using linear regressions and data mining than somebody using an LLM or a GAN.
Importantly, neither method is as computationally intensive as LLMs, and the second one at least is a very involved process requiring a lot of domain knowledge, which is exactly the opposite of how GenAI markets itself.
yeah my dad can kill a dozen people if something goes wrong at work. Yet they use windows and proprietary shit.
If software isn’t secured it shouldn’t be used.
We can make software less prone to errors with proper guidelines and procedures to follow, as with anything. Just to add that it’s not solely on software devs to make it failproof.
I would make the full switch to Linux but I need Windows for photoshop and premiere lol. And I never got Wine to work on Mint, but if I could I would ditch windows today. I think helping people get acquainted with linux is something AI can really help with, and may help more people make the switch.
apologies if this is annoying, but have you tried Lutris?
it’s designed for games, but i use it for everything that needs wine because it makes it easy to manage prefixes etc. with a nice gui
No worries, I haven’t tried it but I also don’t have my Mint install anymore lol (Windows likes to delete the dual boot file when it updates and I never bothered to get it working again). I might give it another try down the line but I’m not ready to ditch Adobe yet. I’ll keep it in mind for if I make the switch in the future.
yes. It’s a tool that can (and must) be seized and re-appropriated imo. But it’s not magic. Main issue is that capitalists are selling it as some kind of genius in a bottle.