The explosion in AI this last year is a lot. The comparisons of Midjourney v1-5 all in the last year are incredible and amazing. And this week with FB’s Llama being leaked, the Stanford paper behind it spelling out how to use it (and also just an interesting read: 2048 GPUs took 21 days to compile it), and it being retooled for home use as Alpaca with C and Rust kits, and being able to run a LLM on a home PC with CPU only is, well, wild. I have a lot of conflicting thoughts, but a few that are breaking through are:

Distribution

This looks like the opening up of the decentralization of AI tools that we’ve sorely needed. I’ve long believe Sci-fi misprepared us for the “future” of the 2010s. We were raised on Star Trek etc that showed us talking to AI assistants and there was this niggling underlying assumption of some locality? It was never spelled out, but when Siri and Alexa and friends started arriving, they were not under the hood what we’d been expecting, but packaged like we expected so many started using them. They were however missing the utter lack of locality, and that these tools operated as a centralized cthulian monstrosity with tentacles into every users home all feeding data back to a corporate owned behemoth. It was horrifying.

These new developments, even in the last week, are putting these tools in users’ hands and more importantly devices, including a Pixel 6 Android phone. I think it’s important this isn’t lost in the noise. There are a lot of questions and thoughts about the ethics of these tools, but we’ve increasingly been having them pushed on us regardless by corporations with out a lot of consent or choice, and as I said, they are all cthulian nightmares. This leak turn tool is a massive shift in power (the point of decentralization is to decentralize power) and in the frame a huge win. Weather you want to use it is obviously still up to you, but I think it’s very interesting that we at least have the option now. And some people will use it just like most corporations are. The future is arriving very fast.

Sci-Fi Singularity Half Measures: Tulpa’s and “Beta Level Simulations”

I was a singularity brain wormed kid of the 90s and early 2000s. I’ve since chilled out a bunch and dropped out of that cult of religion, but some of the impulses and desires remain, like reaching for forms of immortality, which is I think is a not uncomon human impulse. So more grounded seeming measures like Alastair Reynolds’ Beta Level Simulation stuck in my mind longer. And popped up again in a fantasy book I’m reading now with the recurrence of the idea of Tulpas, this time just as detailed mental models of others in one’s own mind.

All that is to say, These new developments of people getting their hands on LLMs and even adjusting them has put me back in the mind of these ideas as something we might be able to start pursuing. Dump in all the writings of yourself (handy if you’ve been keeping a journal) and possibly a corpus of all the media you find influential and important (books, articles, TV and Movies) and voila, a shoddy sim of yourself to leave in perpetuity. I find the idea intriguing for obviously narcissistic, self aggrandizing, wanting to leave things behind, and fear of death reasons.

Disruption and Accelerationism

We aren’t ready for these tools. Especially the text to image and, nowish to soon, video. Just in the news was Trump announcing he’d be arrested as some stunt that didn’t happen, but to tide folks over, someone generated some photos of the event anyway. Meanwhile Clarksworld has closed submissions after being drowned in AI generated short stories. Our brains and social structures aren’t ready for the amount of fake media we’re able to generate now. Also concerning is that this tech is now being used to generate deepfake porn of unconcenting people, now spreading to twitch streamers.

I have no idea how we handle any of this, but if I had to guess wildly it seems like curation may play a role. It always has, we’ve just outsourced much of it to news orgs historically, and in more recent decades, Bayes algorithms of search engines etc. But as people are increasingly aware of how craven, pathetic and biased traditional news is, and “traditional” search tools are also grappling with LLMs and how to “harness” them and presumably filter out outputs of others, it seems relying on social systems is important. The quote “You either build your own information processing pipelines or you become subject to someone else’s” keeps increasing in relevancy. Talk to your friends. Find people you can trust and may be experts in relevant ares to follow.

Finally it is of concern that these AI tools, still mostly in corporate hands, are having their take off in power in this last year, and that those corporate owners would be quite eager to put everyone out of work, and just own all the production for themselves. Hope we all have enough money saved up with in the next few years to “retire” and live the rest of our lives on…

And things are speeding up very rapidly, this last week alone has been ridiculous. Makes it very hard to predict much of anything. The event horizon of this tech is pulled in real tight.

Ethics

I just watched DAIR’s Stochastic Parrot’s Day event and it had some great discussions on the deep ethical problems underlying these tools. Basic shit that I learned in the launch Course classes on ML, like, curation of data is important, being out the window as most of the teams behind these models hover up all the data they can, with examples of some of the models being coaxed into spitting out near verbatim copyrighted work they were trained on, and lawsuits being launched. Of small concolsation is some copyreight departments flat out refusing to grant any copyright to AI generated works for the time being.

Then we got into ethics issues that have also been plaguing social media for decades, the very real and awful practice and cost of out sourcing moderation and censoring and tuning of these models to underpaid, overworked, under supported people in vulnerable positions.

There are a lot of legal and ethical problems underpinning these tools that are only just beginning to be grappled with. But despite all that development and deployment of them is racing ahead, and the distribution and decentralization of them is only going to speed that up further. I have no good answers here, mostly just acknowledgement.