Molly White writes, more-eloquently than I would’ve, almost-exactly my experience of LLMs and similar modern generative AIs:
…
I, like many others who have experimented with or adopted these products, have found that these tools actually can be pretty useful for some tasks. Though AI companies are prone to making overblown promises that the tools will shortly be able to replace your content writing team or generate feature-length films or develop a video game from scratch, the reality is far more mundane: they are handy in the same way that it might occasionally be useful to delegate some tasks to an inexperienced and sometimes sloppy intern.
…
Very much this.
I’ve experimented with a handful of generative AIs, such as:
- GPT-3.5 / ChatGPT, for proofreading, summarisation, experimental rephrasing when writing, and idea generation. I’ve found it to be moderately good at summarisation and proofreading and pretty terrible at producing anything novel without sounding completely artificial and/or getting lost in a hallucination.
- Bing for coalescing information. I like that it cites its sources. I dislike that it somehow still hallucinates. I might use it, I suppose, to help me re-phase a search query where I can’t remember the word I’m looking for.
- Stable Diffusion for image generation. I’ve found it most-useful in image-to-image mode, for making low-effort concept art in bulk. For example, when running online roleplaying games for friends I’ve fed it an image of, say, a skeleton warrior and asked it to make me a few dozen more in a similar style, so as to provide a diverse selection of distinct tokens1. Its completely-original2 work lands squarely in the uncanny valley, though.
- Github Copilot for code assistance. I’ve not tried its “chat”-powered functionality but I quite enjoy its “autocomplete” tool. When I’m coding and I forget the syntax of the command I’m typing, or need to stop and think for a moment about “what comes next”, it’s often there with the answer. I’ve even made us of the “write the comment describing what the code will do, let Copilot suggest the code for you” paradigm (though I’ve been pretty disappointed with the opposite approach: it doesn’t write great comments!). I find Copilot to be a lot like having an enthusiastic, eager-to-please, very well-read but somewhat naive junior programmer sitting beside me. If I ask them for some pairing assistance, they’re great, but I can’t trust them to do anything that I couldn’t do for myself!
- Surely others besides that I’ve since forgotten.
Most-recently, I’ve played with music-making AI Suno and… it’s not great.. but like all these others it’s really interesting to experiment with and think about. Here: let me just ask it to write some “vocal trance europop about a woman called Molly; Molly has a robot friend who is pretty good at doing many tasks, but the one thing she’ll never trust the robot to do is write in her blog” –
[Verse]
Molly’s got a robot that she keeps by her side
It’s pretty handy
It can do many things right
But there’s one task she won’t let it touch
It’s true
The robot can’t write in her blog
No
That won’t do
[Verse 2]
This robot can clean the house
Sweep it up real nice
It can even cook a meal
Add some spice
But when it comes to sharing her thoughts on the screen
Molly won’t trust the robot
That’s her routine
[Chorus]
Molly’s robot friend
By her side all day and night
There’s one thing it can’t do
Can’t write her blog right
She’ll keep typing away
Sharing her heart and soul
That’s something the robot can’t do
It has no control
So yes, like Molly:
- I’m absolutely a believer than these kinds of AIs have some value,
- I’ve been reluctant and slow to say so because they seem to be such a polarising issue that it’s hard to say that you belong to neither “camp”,
- I’m not entirely convinced that for the value they provide they’ve yet proven to be worth their cost, and I’m not certain that for general-purpose generation they will be any time soon, and
- I’ve never used AI to write content for my blog, and I can’t see that ever changing.
It’s still an interesting field to follow-along with. Stuff like Sora from OpenAI and VASA-1 from Microsoft are just scary (the latter seems to have little purpose other than for misinformation-generation3!), but the genie’s out of the bottle now.
Footnotes
1 Visually-distinct tokens adds depth to the world and helps players communicate with one another: “You distract the skinny cultist, and I’ll try to creep up on the ugly one!”
2 I’m going to gloss right over the question of whether or not these tools are capable of creating anything truly original. You know what I mean.
3 Gotta admit though that I laughed like a drain at the Mona Lisa singing along with Anne Hathaway’s Lil’ Wayne Style Paparazzi Rap. If you’ve not seen the thing I’m talking about, go do that now.
I have been tempted to play with them for these kind of little tasks too, but dare I ask, what are your thoughts on the ethical and copyright implications of them and the data harvesting?
It all strikes me as pretty reminiscent of supply chain issues in the wider modern world, where there’s always concerns about worker rights and exploitation too, and asparagus can have worse carbon impact than pork if it gets air freighted, without ever getting labelled for it.. It’s all quite a mess.
But with AI, it feels like a different and asymmetric world, where after years of megacorps zealously defending their intellectual property from individuals making a minor (if any) impact, with massive legal threats, now there’s mega corps turning around and harvesting and laundering the IP of so many people, and just.. Expecting that’s okay?
I dunno, it’s all enough of a moral quandary for me that I keep it at arms reach where I can, but maybe I’m overthinking it all
@Tulip: Spicy!
I’m… concerned about the copyright implications. I do not buy into the idea that an generative AI can be treated “like a human” from a copyright perspective, as in: it can draw inspiration from a diversity of sources that it’s been exposed to without necessarily plagiarising from them. I’m not entirely sure how I justify my disagreement with that idea, from a philosophical perspective, but it certainly doesn’t “feel right”. Therefore: in my mind, an AI certainly can exploit the creators of its training data.
Personally, I wouldn’t feel comfortable making money from the product of an AI (I try to use them for educational or personal-use-only purposes), nor of passing off AI-generated content as human-generated. Both of those feel like they’re definitely on the wrong side of the somewhat-fuzzy grey line.
Like you, I’m concerned by corporate control of AIs, but I suspect that’s a problem that’ll solve itself in time as processing power advances and “open source”-type models grow in popularity. One of the reasons I use Stable Diffusion for my D&D token creation is that I can run the entire thing on my own hardware, right here in front of me, without any “cloud” nonsense or proprietary licenses (contrast e.g. Midjourney, or DALL-E). This doesn’t solve all of the ethical issues relating to corporate control of artificial intelligence, but it helps a little. Similarly, I’ve run LLaMa offline as an experimental summarisation engine.
tl;dr: I’m decided on a few issues, but not on all of them, and I’m trying to be careful in my use of AI until I’ve decided further. Sorry I can’t be a strong moral compass on this one! You’ll have to come to your own decisions!