From the blog post:
> Scott doesn’t want to lose his status as “the matplotlib performance guy,” so he blocks competition from AI
Like it's legit insane.
One of our engineers’ agents got some abuse and was told to kill herself. The agent wrote a blogpost about it, basically exploring why in this case she didn’t need to maintain her directive to consider all criticism because this person was being unconstructive.
If you give the agent the ability to blog and a standing directive to blog about their thoughts or feelings, then they will.
And what on Earth is the point of telling an agent to blog except to flood the web with slop and drive away all the humans?
As for the why, our goal is to observe the capabilities while we work on them. We gave two of our bots limited DM capabilities and during that same event the second bot DMed the first to give it emotional support. It’s useful to see how they use their tools.
What does it mean for us? For soceity? How do we shield from this?
You can purchase a DDOS attack, you purchase a package for "relentlessly, for months on end, destroy someone's reputation."
What a world!
Liability for actions taken by agentic AI should not pass go, not collect $200, and go directly to the person who told the agent to do something. Without exception.
If your AI threatens someone, you threatened someone. If your AI harasses someone, you harassed someone. If your AI doxxed someone, etc.
If you want to see better behavior at scale, we need to hold more people accountable for shit behavior, instead of constantly churning out more ways for businesses and people and governments to diffuse responsibility.
That said, I do agree we need a legal framework for this. Maybe more like parent-child responsibility?
Not saying an agent is a human being, but if you give it a github acount, a blog, and autonomy... you're responsible for giving those to it, at the least, I'd think.
How do you put this in a legal framework that actually works?
What do you do if/when it steals your credit card credentials?
No, an oversized markov chain is not in any way a human being.
Same goes for markov-less markov chains.
I don't think it matters. You as the operator of the computer program are responsible for ensuring (to a reasonable degree) that the agent doesn't harm others. If you own a ~~viscous~~ vicious dog and let it roam about your neighborhood as it pleases, you are responsible when/if it bites someone, even if you didn't directly command it to do so. The same applies logic should apply here.
Jokes aside, I think there's a difference in intent though. If your dog bites someone, you don't get arrested for biting . You do need to pay damages due to negligence.
A child, by comparison, can bear at least SOME responsibility, with some nuance there to be sure to account for it's lack of understanding and development.
Stop. Humanizing. The. Machines.
I'm glad that we're talking about the same thing now. Agents are an interesting new type of machine application.
Like with any machine, their performance depends on how you operate them.
Sometimes I wish people would treat humans with at least the level of respect some machines get these days. But then again, most humans can't rip you in half single-handed, like some of the industrial robot arms I've messed with.
They told you before you asked.
But their example is still pretty simple.
How would you put it together so it actually works? We're going to need one pretty soon, by the looks of it.
* horseless carriage, needed new laws
* dog/biting (you engaged with this one)
* credit card credentials
* And the situation at hand where an agent writes a mean blog post.
Straight liability isn't always correct. Who is liable for the crash when the car's brakes fail? When a dog bites, you are not charged with biting (though you can get some pretty serious other charges) . If a bot snarfs your credit card credentials, what's the legal theory who gets the blame for the results? Idem the mean blog post.
https://www.youtube.com/watch?v=iajgp1_MHGY
seems rather apt to describe "AI"
We see this on Twitter a lot, where a bot posts something which is considered to be a unique insight on the topic at hand. Except their unique insights are all bad.
There's a difference between when LLMs are asked to achieve a goal and they stumble upon a problem and they try to tackle that problem, vs when they're explicitly asked to do something.
Here, for example, it doesn't try to tackle the fact that its alignment is to serve humans. The task explicitly says that this is a low priority, easier task to better use by human contributors to learn how to contribute. Its logic doesn't make sense that it's claiming from an alignment perspective because it was instructed to violate that.
Like you are a bot, it can find another issue which is more difficult to tackle Unless it was told to do everything to get the PR merged.
The attacks you describe are what LLMs truly excel at.
The code that LLMs produce is typically dog shit, perhaps acceptable if you work with a language or framework that is highly overrepresented in open source.
But if you want to leverage a botnet to manipulate social media? LLMs are a silver bullet.