#perpetually screaming

You're right: the conclusion that we should spend lots of resources making AI less harmful as opposed to using less AI in the first place is a strange takeaway! Especially if one of your first (or only) impressions of my views on AI was through that tag message.

I agree with you that we should absolutely be using less AI. The marketing behind current AI platforms as a way to "get things done" and "simplify your life" or as a way to "cut labor costs" (on work that's better done by humans anyway) is incredibly harmful and unnecessary. AI should not be used for the vast majority of stuff it's being pushed for. Because a lot of the hype around these well-known AI platforms (ChatGPT, Claude, Grok, Sora, Midjourney, etc.) is a thinly veiled excuse for continuing to exploit the masses in service of the super rich and their interests (resource extraction, profit, and information control). And that's not to mention the vast amounts of human exploitation already happening with training dataset labeling or content moderation of these platforms.

What I was referring to in my original tags was making AI more sustainable in situations where it can be useful.

An example of this is semantic search, i.e. being able to search digital texts (webpages, PDFs, etc.) by word "meaning" instead of only matching strings of words verbatim. A practical example of this would be a tool in which the act of searching a biology textbook PDF for "taxonomy" not only highlights every single instance of the word "taxonomy" but also highlights things like "phylogenetic nomenclature" or sections of text about how taxons are described. These topics are still about taxonomy, even if they don't directly contain the word "taxonomy." And if you're someone who needs to find a quick reference for making a study guide or for writing an introductory section on a report, a tool like this can be incredibly useful.

But isn't that what an index is for?

Yes! You could use the index of that biology textbook PDF to find this same information, if you wanted. But how often would you actually go into that index as opposed to using the CTRL-F or CMD-F shortcut to find things in a PDF? Or on a website, which often doesn't have an index. Maybe you do enjoy the process of searching through text yourself --- that's ok! But I hope you can still recognize how useful a search tool like the one I've described here can be, and how useful semantic search itself can be.

Back to my search application example. In order for semantic search to identify sections of text in a PDF that most humans would agree are related to the original search query, we need some way to tell the application which words are related to each other. The best technology we have for doing this right now is a large language model (LLM), which is a highly complex statistical model of how words or groups of words are related to each other. Applications like ChatGPT use this type of statistical model to generate "relevant" text in response to a user's prompt, by repeatedly running the conversation through this model and predicting what should be the next word. The LLM is the relational information, not the act of generating text. On a technical level, my hypothetical search application would be taking the relational information inside the search query and matching it to similar relational information inside the search text, returning those results to the user.

Unfortunately, good LLMs these days (even the smaller ones!) still require shitloads of electricity and drinking water in order for them to learn these word relationships with reasonable accuracy. Good LLMs may also require shitloads of electricity and drinking water in order to run, especially if I host my application on the cloud as opposed to making it a download that each user can run directly on their own device.

We absolutely need less AI. Less AI is integral to environmental sustainability and human well-being. But we also need better systems for the situations in which AI can be useful. The lessons "we should use less AI" and "we should spend a bunch of time and money making it less harmful" are not mutually exclusive to me.

That's why I want to study it, and continue to talk about it.

Also: the problem with data centers being pollution farms has existed long before generative AI became as widespread as it is. Google, Netflix, Amazon, etc. already needed fucktons of powerful computers in data centers so that they can handle requests from millions of daily users, and so that they can store and transfer the vast amounts of information required to make their platforms work. Environmentalists have been sounding alarms about the climate and water crisis worsening as a result of these data centers for years. The reason why it's become more talked about now is because generative AI --- and its associated "race to AGI" (artificial general intelligence) --- that these corporate fuckers are deluding themselves over has made these environmental problems exponentially worse. What we're seeing now, in rapid, real-time change, is the cost of Silicon Valley's prevailing attitude of "move fast and break things" combined with tech CEOs' refusal to be held accountable for their companies' so-called "progress."

(It's not progress. We're being exploited.)

@iceandrain - I hope you got something out of what I've written here, because I would really like to continue this conversation with you! I'm also not entirely sure how familiar you are with the topics I've covered in this post (especially the more technical stuff), so if there's anything here you need me to clarify or re-explain, please let me know and I will happily do so.