Hollywood labour disputes

You say "Hollywood and its issues" as though it's not a very visible example of a current business drive to essentially screw over everyone whose job involves doing anything that might feasibly be automated.

I'm not about to claim that SAG AFTRA are selflessly doing this only because of the big picture, but a lot of people are already struggling with causalisation of their work and now an additional threat arrives in the form of "AI" and companies wanting to use it to cut costs, with no concern for what happens to displaced workers. (Even companies who are ambivalent about it are getting pressured to look at it because the moon-logic of shareholders means that what's actually good for the company is less relevant than what might make their share value go up...)

A lot of public discourse about LLMs and GANs has been its impact on creative roles (in no small part because of how much material has been used in piratical fashion - hey, if it applies to downloading an MP3 it applies to putting data into a training set without the creator's permission - to train the things currently being called "AI"). But there's also a massive amount of effort going on to try and crowbar it into numerous white-collar "knowledge economy" jobs - with similarly poor results (but still, by being much cheaper than humans, appealing to a certain type of C-suite). Programmers, sysadmins, network admins, cybersecurity... And that's just in IT itself. Any industry where companies make software will also have people trying to cram "AI" into it with a pitch of "pay for a spendy sub to our software and you can fire a big chunk of your workforce". A lot of 'safe' well paid jobs where people have for years been steered towards if they want a reliable income in a growing industry are now in the crosshairs of this deeply stupid idea, and all being pushed by short-sighted greedy sociopaths for whom the answer to "what if it triggers enormous amounts of unemployment?" is "it'll be great, we can push down wages and hoard even more of the wealth".

For me the goal is not so much "stop progress" as "stop wildly-overselling the abilities of what currently amounts to massively scope-expanded next-word-prediction". There are a number of ways that tools based on the technology can be useful, but those are as assistive instruments where you're talking about, at most, a modest percentage increase in efficiency of working processes.

The problem with the current AI hypecycle is that what is being sold is "With this technology you can totally fire most of your workers and replace them with a Widget(TM) from us", despite the twin facts that:

• the output of a lot of said tools is not suitable for direct publication (whether it's garbled or factually incorrect text from an LLM or images with wonky biology from a GAN) without human intervention, and
• licencing and copyright issues are going to get real big real fast in this area - between the US Copyright Office stating that "AI"-generated works are excluded from copyright and the brewing storm over "AI"-type tools whose training sets have not been carefully curated and correctly licenced.

"AI" is a current hype bubble, just like the internet was with the dotcom bubble at the end of the 90s. It's not that the technology in question can't be useful or potentially transformative, but the decisions are being pushed by investors and shareholders who have no understanding or concern for any of the longer-term factors at play - they just want to see Number Go Up.

IMO we are still too early in the days of LLMs to trust them with responsibilities like tutoring though - the guardrails are far too feeble and repeatedly fail to catch things like hallucinations (internally consistent states where they insist that incorrect information is correct). And they're also being dangerously slow at protecting against "prompt injection" attacks (a way of bypassing the guardrails by tailoring your input; doesn't sound too bad until you take into account that user activity is fed back into the underlying LLM as additional training data, so output poisoned by a prompt injection attack is then treated as training input).

On the copyright front there was a court ruling in the USA already that any work generated by an AI is ineligible for copyright protection (the underlying logic being that copyright is a tool to promote the arts by providing artists with a set of exclusive rights that they can leverage to monetise their works and offset the cost of their creation); whether that remains in place is a different question. Meanwhile the likes of Adobe have updated the terms on their tools to automatically enroll subscribing customers into allowing any files saved in the cloud to be used to train their latest iterations of generative tools (rather than e.g. paying to licence a set of images for use as a training library).

It is IMO a mistake to try and wave away the copyright issue on the grounds of "sure people do the same thing" - partly because people are not machines that can perfectly reproduce audiovisual works in a very short amount of time, and partly because as I mentioned above - the goal of copyright law is to create a framework that supports and promotes the creation of new art. It does not automatically follow that "making a computer do all the work" still produces equivalently valuable work - any artist in any field you talk to will tell you that ideas (at the level of granularity typical of LLMs) are very very easy to come up with, and most of the work is in the execution. Art is about crafting a medium to express an idea, with nuance and specific details of the artist's perspective being worked into the craft throughout. LLMs can generate striking images, without a doubt - but I've yet to see one that has held my attention or would make me come back to it and study its form, for example. (This is, admittedly, very much a Your Mileage May Vary area.)