AI's next leap requires intimate access to your digital life

I think what you train should stay on your own box and not be lured into sharing for some profits, as the profits will be collected in a few large funnels to pay large salaries for CEOs and what not.

The only future where this happens is if little, non- infrastructure / specialized hardware models maintain functional parity with "throw more hardware at the problem" big models.

That seems... unlikely.

And then we'll see the same evolution that built Google.

Someone with the best algorithm leverages that into a revenue stream, then uses that revenue to buy hardware, then monetizes its userbase via advertisement, then uses that additional revenue to buy market share.

I'm sure it'll make someone a ton of money, but it's not a future I'd like to live in (again).

Diversity is strength.

> On the other hand I have started using llama 3.3 and it is approaching usefulnes

cute phrasing! so what you're saying is that you're using it, but it's not useful. You can imagine it being useful, and it produces output similar to useful output, but it isn't useful.

That perfectly matches my experience with language models, going back to when I was running GPT2 locally.

Has there been any evidence that the improvement is not asymptotic? For ten years language models have been approaching usefulness.

When do we get to usefulness?

For me it's slightly nuanced. Personally, I'm less concerned about yet another attack surface for my personal information from a 3rd party I didn't expect, it's probably all out there anyway. If I gave all my data to an AI company and some hacks in and steals it, for me anyway, I'm less concerned (I realize very many people find that strange/insane). I am considerably more concerned about being psychologically manipulated by an AI tool. I would say however, to your point, it would be shit if someone hacked in, changed or inserted data, and that was use to manipulate the AI, however my sense is that might be difficult/unlikely?

Sure, most technology replaces human labor. The printing press put thousands of book copiers out of work. The word processor destroyed the entire job description of “typist”.

The promise of AI is that it will free human labor to focus on higher value areas rather than copying books or typing up the boss’ scribbled notes.

I could see the endgame of AI as surveillance and manipulation of information to achieve political goals. In the old days, expensive intelligence agencies and security services had to work hard to figure out targets' affiliations and beliefs, to propagandize, and to subvert opponents (or one's own society). Social media then persuaded users that it was a good idea to list their family, friends, and associates in a giant database along with their political beliefs, hopes and fears, and pictures of their lunch, and its algorithms surfaced whatever the makers wanted (cf. TikTok). Now, AI is poised not only to comb through what e-mails and files hadn't been available before, but also to surface or even generate information, giving it a prime spot to censor or manipulate what we see.

> I wrote my PhD thesis exactly around this topic (~8 years ago).

For anyone interested, from checking the HN profile the thesis seems to be “The Power of Robot Groups with a Focus on Persuasion and Linguistic Cues”.

https://ir.canterbury.ac.nz/bitstreams/cac255b7-43fe-4e08-94...

Imagine trying to perform a “right to forget” request on data that was trained into the AI?

Techbro: "Yes, we are working on robots and that kind of stuff, but it is a more difficult field. To get the accuracy and versatility of .."

> Are dark patterns harder than captchas?

Yes, yes they are. A captcha is a simple system which returns a binary pass/fail. A dark pattern plays with your understanding of reality to make you choose what is beneficial to the company at your expense.

AI systems will be gamed to follow dark patterns too.

> Why would it not avoid dark patterns?

What do you think is going to happen if an AI system which avoids current dark patterns starts being widespread? Bad actors aren’t going to shrug their shoulders and stop, they’re just going to make systems which trick those AIs. Adapting a dark pattern is significantly faster and more cost effective than training a new AI to bypass those new approaches.

I presume Google Mariner is using the digitized behavior of googlers searching for products, analytics from shopping websites (I would use it), even ad interactions (why not?). If it works, boom, lots of unproductive stale data turned into a potential product (if someone builds this, I want 5%).

The problem is: that data has a lot of users under the influence of dark patterns. People shopping in stupid ways, falling for tricks. They probably know this and improved search against some of these patterns in the past. But it's not perfect, right? Search isn't.

Large models fall for tricks all the time, you have to assume the other way around. They will fail, and you can reduce that failure rate, but never zero it.