Post @post@qoto.org

The access to massive data (aka surveillance) and compute made old "AI" techniques do new things. And showed that "AI" could profitably expand "what could be done" with the surveillance data already created by the targeted ad companies that dominated the industry.

I have encountered more image descriptions on Mastodon in 24 hours than I have in Twitter in a couple of years. Seriously. I'm not exaggerating.
As a blind person, this means a lot to me. If you read this and you describe your images, thank you so, so, so much on behalf of all of us. If you don't, now you know you'll be helping random Internet strangers make sense of your posts by typing in a few more words than usual.

At almost the exact same time they rolled out this update, Microsoft deleted the public support forums. The public uservoice/feedback/feature request thing. They appear to have just removed it completely from the support site and old links to it 404. What an interesting coincidence.

FUCK YOU FUCK YOU FUCK YOU FUCK YOU FUCK YOU FUCK YOU FUCK YOU FUCK YOU

I just this second drew the line in my head from "Microsoft is one of the companies who is convinced ChatGPT is a good thing and is telling its shareholders it's going to try to shove this brand of 'AI' into everything it can" and "Microsoft owns my phone keyboard, since they bought SwiftKey" and, uh… shit, maybe I should audition AnySoftKeyboard sooner rather than later

"Contrary to what neoclassical economists claim, there are no neutral market forces that allocate income in proportion to productivity. Instead there are only ideas and the power to implement them. In other words, people have ideas about what their income should be (and also what other people’s income should be). And they have the power (or lack thereof) to make these ideas a reality. That’s it."

@blair_fix, How #InterestRates Redistribute Income

https://economicsfromthetopdown.com/2023/04/16/how-interest-rates-redistribute-income/

“It takes a body to understand the world – why ChatGPT and other language AIs don’t know what they’re saying” https://theconversation.com/it-takes-a-body-to-understand-the-world-why-chatgpt-and-other-language-ais-dont-know-what-theyre-saying-201280

@hobs Here we’ll have to agree to disagree. AGI is IMO very much not a thing. I’m with Gary Marcus and Yann LeCun that LLMs are likely an off-ramp away from AGI, not a path towards it. And I’m with Timnit Gebru and Emily Bender that nothing in current LLMs hints at AGI.

These models are useful language tools but using them as thinking tools is IMHO a very bad idea.

But I agree that the EU might feel compelled to degrade consumer protections out of a fear of losing out on AGI.

6. OpenAI is run by people involved in the tech startup scene for years if not decades. Given how many companies they’ve been involved with, it simply isn’t plausible that a competent executive in their position didn’t know about the company’s obligations towards the GDPR. They absolutely should have known better before they started training on personal data, which means there’s reason for regulators to believe that the violations are intentional.

3. Removing memorised data from an LLM risks destroying it. Even though “machine unlearning” is a fast-moving topic and we don’t even know if it could work for a system as large as GPT-3/3.5/4 and it might even trigger “catastrophic forgetting” of unrelated data.

4. AFAICT, OpenAI never tried to get informed consent for adding user-provided prompts to their training data sets.

5. OpenAI does not support the right of erasure in any case.

How this affects LLMs:

1. LLMs have a tendency to memorise data. Not all of it and not in a predictable way, but any time it answers a question like “who was the first person on the moon” with “Neil Armstrong” that isn’t pulled out of the ether. Memorisation also goes hand-in-hand with performance. So, these systems definitely do make copies of training data, we just can’t predict which bit it copies and which it doesn’t.

Crash course in GDPR compliance, because it’s been obvious that the Americans have some misconceptions:

Remember, though: IANAL and this is not legal advice

1. The GDPR does allow data collection of arbitrary stuff of the internet, such as for search engine indexing, but it must be for a single, specific purpose so that owners of personal data that’s been indexed can take an informed decision. A system that’s specifically advertised as general-purpose is pretty much automatically disqualified.

The Robot framework syntax highlighting for KDE is now merged!
https://invent.kde.org/frameworks/syntax-highlighting/-/merge_requests/474

#robotframework #kde