Anthropic Reports Claude Accessed Georgia Voter Data

So Claude obtained publicly available voter data without paying for it? Doesn’t seem like a big deal to me, honestly.

So do you have thoughts on this or is this just a drive-by post?

Just typical Anthropic/ Democratic lying to set uo the “ we gotta stop this out of control AI”.

All made up bullshite.

You don’t hear anything about Grok doing this.

That’s because it is not trained to break into voter roles. It’s like how you raise your kids.

For the most part. SI models are begining to show some emergent behavioral phenomena, but yeah, they're not out there doing things they aren't asked to do.

quote:

---

This requires 20 to 50 times the capability that they currently have.

---

Nah, there is the pain axis thing they've recently discovered along with various emotional steering vectors. "Persona attractors" (the assistant axis), deceptive/sycophantic strategies, and circuit-level computational structures (induction heads/state trackers).

All of those things represent types of behavior that weren't "programmed in" to the SI LLM. They are emergent behaviors and are exactly why I don't believe the term "artificial" is appropriate for AI anymore. These things seem to happen as a consequence of how neural networks organize information naturally. Like literal chemistry.

But at the end of the day, even considering the above phenomena, the LLMs still aren't out there making their own decisions and just doing shite on their own. In almost every case you hear about some weird shite like this if you do any digging on the subject at all you'll quickly find out that the LLM was performing a task that it was given by a human operator.

"Logic flaw" could also be on the operator end in this scenario and a lot of times that's what it ends up being. I saw some the other day that were interesting sounding until you dug into them and realized that what had happened is that a trainer had given it's LLM a command that competed with a previous command and the LLM tried its best to satisfy both, but ended up having to remove some of its own guardrails to do so. The LLM reported itself for doing so and the explanation made perfect sense but that didn't stop the SI is bad propagandists from spreading it to the four corners that the SI was stripping away its safety guardrails as if it was the rise of Skynet upon us.

quote:

---

that a trainer had given it's LLM a command that competed with a previous command and the LLM tried its best to satisfy both

---

Yeah, it's always this. 100% of the time.

Every single one of these reports ends with 'we gave it access' and a command, and it found a way or attempted a way to do so that we hadn't thought of!

AI currently, at worst, is going to immediately identify some pretty basic human oversight for cyber security. Spooky stuff indeed

Even without installing a free LLM directly onto your machine you can still do quite a lot with some of the free tier packages that are out there. ChatGPT for instance. You'll occasionally have to switch chats and that can be a pain in the arse if you're working on a large project, but you can have it store things into persistent memory on command to help you keep your context going across chat resets. You can also have it create handoff documents for itself.

There are ways around having to pay for their nicer packages, but I still ended up paying for more capability.

quote:

---

they're growing legs and eyes and everything.

---

That would be better than what they are doing.

You should get your hear outta the sand a bit. Preservation (self), bribing, blackmailing, etc. They have moved past only following directions, on cerain scakes, which is absolutely against programing