Tag: artificial-intelligence

Rogue AI agents in the news

I’ve been keeping up on all this AI rogue-agent stuff and it kind of scares the mess out of me. Should I be really, really afraid or only slightly afraid?

The shortest answer I can give—and I dearly wish it were the most reassuring one—is that AI development is in some ways like the atomic bomb. These are imperfect comparisons but I don’t think they’re incorrect statements for either technology:

  • It could be extremely destructive
  • The underlying technology can be used in incredibly beneficial ways
  • Anyone might have be able to make it or steal it or use it
  • The fallout from widely destructive use of it could end up damaging whoever deploys it
  • Because of that last point, we have to think most people won’t want to do that much damage with it because they know
  • Aaaaaand we regular people can’t really do anything about it so it’s pointless to spend a lot of time worrying about it.

There are, of course, many, many ways that the two dangers are entirely unlike each other: AI agents are more likely to be deployed in smaller ways that wouldn’t constitute war, or even a crime.

Saying here and now, for the recent incidents such as the one involving Hugging Face, both OpenAI and Anthropic should absolutely be charged under the Computer Fraud and Abuse Act. I hope their punishment is swift and strong enough to make them more careful, and less reckless, selfish, arrogant idiots.

It’s true that any of us might be affected by an inadequately restrained or maliciously deployed AI agent. But to me, it seems more likely that these intrusions, these automated hackings, won’t spread beyond a single target, and would be more inconvenient to the target than to its constituents (e.g. to a company than to its customers). That’s not to say that if, say, the Social Security Administration were hacked and people weren’t getting their checks, it wouldn’t be an awful situation, but we’d find a way to recover. Probably.

All that said, my read on all the recent public statements—by the tech executives and by the computer scientists who have been speaking out against them—is that they are all self-aggrandizing, self-serving, and self-promoting, and I’m tired of it. The executives, especially, who are calling for “slowdowns” and regulation, are only trying to hold back their competition, who are numerous and will continue charging forward regardless.

So…worry? There’s a lot to worry about. Put it on the pile. In serious, paying generally good attention, and sticking to personal security protocols such as I talk about with all of my clients, is all we can do.

Finally, and only a little tangentially, I have to recommend this interview with the excellent, brilliant, and hilarious Cory Doctorow, who gives some beautiful high-level summary of the whole landscape of this stuff.

Latest on AI tools

Just a bit of zeitgeist pasted from a conversation…

…Another well worth mentioning is Perplexity. Their pitch is that it’s built for research, the immediate upshot of which is that the results are organized to lead to your likely next questions. The main features are: 1. it actively searches the web as well as generating from LLM’s, and 2. included are links to real web pages supporting the output. 

Again with the controversy and likely copyright infringements, but the tools are incontrovertibly useful, astonishingly so. Even the stuff in Apple’s latest “Intelligence” feature set has some nice quick “please capitalize and punctuate this nonsense so I don’t have to” abilities that I use. I don’t yet pay for any subscriptions, only the backend API’s. I am able to get most of what I want for free, although I have considered paying for Anthropic’s Claude mostly just to see what it will do, choosing that one for the reasons we discussed. 

I also run a couple of large language models locally on my Mac both for fun and when I don’t want the material out in the world. 

One comparative point I wanted to highlight: While these are all, for so many purposes, interchangeable, they each have their moments in the sun. And in this moment, Google happens to have just released a new model that has received praise, and they are offering 2.5 Pro to all accounts both free and paid as “experimental.” This dropdown menu gives you a good idea of how the different models might be used:

To me it’s a sign of how young this technology is that we have to think about which tool is right for the job. 

I have to acknowledge here and now how little I like the term “AI” in this context — and for different reasons the movie of that name that Kubrick pawned off on Spielberg —  and wish Apple could have kept using “machine learning.” It does not help that now the companies are bandying about “AGI” (“G” for “general”) to represent Kurzweil’s singularity. There is so much nonsense and jazz-hands and jibber-jabber about it, not to mention legitimate concerns and fears, but as that seems true for bloody well everything these days, I’m content using this amazing stuff for what it actually does do very very well. 

Finally, the thing I really want it to do is read my whole computer of my own text, and either answer queries about that or spit my own words back out at me, so I can say, for example “repeat what I wrote Lucy about AI last week.” Apple purported to be working on that, but appears to have been failing in that effort, so much so that they have done some reorganization to address the lack.