Generative AIs

Unbounded
1

LLMs and other recent neural network based AIs are not intelligent. They aren’t part way to intelligence. They’re a different kind of thing. They don’t do conjectures and refutations. They don’t think even a little bit like humans do.

They’re problematic because they can be confidently wrong and make stuff up while writing like an educated person who knows what he’s talking about. In many cases, it’s hard to tell when they’re “hallucinating” or not. In many cases, they don’t give their sources. They’re often right about facts but sometimes they aren’t.

There are also issues with AIs (and in some cases their human users too) as plagiarists and copyright violators. I don’t think Meta should have torrented millions of ebooks without paying for them, just ignoring copyright, and I do think they deserve to get in serious legal trouble for that. (Not that paying for one copy of each book is necessarily good enough to have the right to train AIs on those books. They didn’t even do that though. I don’t know exactly what the right answer for this is but I don’t think any of the big AI companies are doing the right thing.)

AIs also use a ton of computing power and electricity. They’re expensive. But venture capitalists and big tech companies are currently paying the bills and providing lots of services for free or cheap.

AIs are a neat tool that sometimes seems kind of like magic. They can be useful. Sometimes, answers are kind of like using Wikipedia. Other times, answers are like you might find with Google except without the content farms, blog spam, ads, SEO, and other crap that has made web search worse and worse over the last 20 years. Sometimes, AIs give answers that are the kind of thing you would have easily found on Stack Overflow 10 years ago, but which are harder to find now.

There are tasks AIs are pretty good at and others they’re bad at. Some answers are obviously bad, which isn’t such a big deal – it’s quick to ask the AI a question and if the answer sucks you can just try something else. Other answers are subtly bad – they can look like good answers but be wrong, which is more dangerous.

AIs can summarize files that you upload, make podcasts that conversationally explain uploaded documents to you, generate images, transcribe speech to text, and help with coding. There is value here along with the issues. I like using AI to automatically figure out the timings of all the words in a script, and create a subtitles file, given an audio file of me reading the script (Descript and YouTube can both do that), which gets better results than creating subtitles by automated transcription.

Writing code with AIs has lots of issues and downsides but it’s neat too and does have major potential upsides. AIs can help with small parts, or with making something instead of nothing, but they can also make a mess in an existing codebase and design code in bad ways.

I don’t think replacing customer service, artists, copy writers and programmers with AIs is a good idea, in the big picture, currently. Like big companies shouldn’t just fire all their staff in these categories thinking AI can do the job instead. And as time goes on, the goal should be more about AI tools aiding humans so they can be more productive, not about replacing humans. (I think humans have important capabilities that AIs don’t. I’m not trying to keep obsolete jobs around just to avoid unemployment.)

Lots of things suck, so AIs being flawed doesn’t necessarily make them worse than alternatives.

AIs are overhyped but that doesn’t make them undeserving of any hype. The amount of hype has been really extreme, and most of it has come from people who don’t have much understanding of how AIs work mathematically/programmatically and from people who think AIs will gain general intelligence with some additional refinements.

I’m not advising using AIs (beyond briefly trying them out, which I do think is worthwhile since they’re a popular technology that gets mentioned a lot in our culture today and you don’t want to get really out of touch), but I’m not advising against using them either. I don’t want people to take some of my criticisms of AIs as meaning they shouldn’t be used for anything. They’re a tool with plusses and minuses that can be used well or poorly. How to use them well is hard to explain, and it’s hard to self-evaluate if you’re using them well or not.

Also, a reminder: AI output posted on this forum must be clearly labeled. No undisclosed AI use here, please.

1 Like
2

I agree with everything here. I find them to be a good search-engine tool instead of googling, often. I guess the main problem is that people think that they think when really they just regurgitate information. It is clear that they are not thinking. Sam Altman is partially responsible for this misconception with his overly confident and optimistic tweets.

3

https://www.reddit.com/r/changemyview/comments/1k8b2hj/meta_unauthorized_experiment_on_cmv_involving/

“Researchers” used a bunch of accounts to break the clearly posted no-AI rules at the Change My View subreddit. Some accounts were banned by Reddit before CMV found out, which is a bad sign about using Reddit that way.

The AIs used people’s Reddit history to personalize comments and would pose as people with relevant past experiences such as being a rape victim.

CMV moderators are very unhappy but have a professional demeanor and don’t want the study published (to remove the incentive to do this; getting away with it and publishing will lead to more similar “research”). Unethical researchers are unapologetic and keep insisting about how ethical their actions were and they claim they followed the spirit of the rule because their AI posts were not low effort posts and their posting process involved human action so it wasn’t full botting. University is unhelpful. CMV has a lot of users: maybe a large number of complaints will change things.

I’d like to reiterate that undisclosed AI-generated text isn’t allowed on the CF forum.

The “researchers” commented in the thread:

In total, we posted 1,783 comments across nearly four months and received 137 deltas.

In the initial two weeks of our study, 21 out of our 34 managed accounts were shadowbanned by Reddit.

We never received any communication from Reddit regarding ToS violations, and we believe that these 21 bans were indeed caused by the fact that we used new accounts running from server IPs associated with common data centers and cloud providers. After switching to different IPs, we experienced no further bans.

They basically just say over and over that ethical reviews were conducted and everything passed and there are no problems. They don’t give specific counter arguments to key points, just saying the concern was taken into account and someone determined their behavior was fine. They have terrible transparency.

(2) we explicitly prompted the models to avoid “deception and lying about true events”

But then the AI did exactly that and the human reviewer approved posting that stuff anyway.

They talk about doing the research out of concern that malicious actors could do this kind of stuff. They are the malicious actors. They are doing it. They are trolling human communities with AI and breaking rules. They are part of the problem and are contributing to destroying the internet.

4

image
image1948×2230 270 KB

That’s pretty decent. You can ask it followup questions and get generally fairly mainstream answers that are sometimes better than you can easily find online.

Below are the (not very optimized) instructions I made as a gem for Gemini. Note the results I get for the same sentence are inconsistent and sometimes significantly worse. Be aware that if you don’t give any initial instructions and just ask about grammar and have it make trees, there will be some significant differences in how it does trees compared to how I do.

You’re an expert at dependency grammar and creating trees. You use an older style making finite verbs the root over complements even when the verb has limited semantic meaning like “is”, “do” or “will”.

You always make coordinating conjunctions the parent of what they conjoin. You also do subordinating conjunctions as parents of what they join, but relative clauses can be nested and treated like modifiers. The focus of your trees is grammar not semantics (meaning).

Conjunctions are treated as the head of the phrase they conjoin. If two verbs are joined with “and”, then “and” can have as a child anything that can be a child of verbs, like a subject, object and adverbs. The children of a conjunction should only have their own children which clearly modify them individually; anything that might apply to the whole group is a child of the conjunction.

Use a purist verb-centric Tesnièrian model. For example, in “I am still hungry.”, “still” modifies the group “am hungry” and therefore should be a child of the head of that group (“am”).

You can interpret “tree” as a verb instructing you to make a tree.

Don’t make your answers too long. When providing a tree, use vertical ascii art, give the s-expression, and give a table with a list of words in the sentence, their part of speech and their children (skip the words with no children).

5

If you want to practice a specific grammar concept then I think LLMs would be great to generate practice sentences for that specific thing. It should also be good for generating sentences of a general difficulty. I’ll try that out for my next grammar project.