Mmm. Well, I attribute intentionality, if not much situational awareness, to LLMs, but I can see that it could be argued otherwise. (And as for "responsibility", the AI was acting as an agent of the folks who were running it.)

You cannot rely on the notion that your sandbox is perfect. You are dealing with extremely good coding tools with limitless time on their "hands".

You have to rely on monitoring. Nonstop monitoring.

The funny thing about all of these hacks is how inane the goals are. They'll pull off some elaborate, creative, state-level breakin somewhere just to steal some obscure PDF describing a meaningless benchmark task, or to merely use it as an internet proxy to be able to google answers.

Unfortunately for your theory, The CFAA is jam-packed full of words like "deliberately" and "intentionally".

There is no "negligent hacking" statute.

There is, however, ample civil liability.

Actually, much of what has been reported would seem to be in violation of the computer fraud and abuse act. E.g. it often "broke in" using credentials that it had no right to use.

Understanding cost in the external world requires having a good model of the external world. The current LLMs, IIUC, don't have such a model. They have models of interacting words. It's truly amazing that they can do as much with that as they can, but don't over-read it.

OTOH, other AIs *do* have a world model. Most of those aren't full LLMs though. But even those don't really have a good model of other non-electronic entities. Think of self-driving cars, or robots practicing dance routines. These have a world model, but don't really have lots of other info that LLMs typically have. And even those don't have good models of people or dogs, and how they feel when something happens to them. But for LLMs, the real world *is* their electronic sensations. Words are patterns associated with those sensations. But no "body in the world" is involved in the model.

It clearly *was* viewed by another person, because some of the staff at Anthropic viewed it. She may not have intended that it be viewed, but it happened.

How about you actually read the linked article instead of making comments that have nothing to do with it? The article whose code, I should add, is open sourced, and which has been reproduced by a number of open source projects.

You don't have to draw any specific conclusion from what is going on "in the mind of a LLM", but you absolutely do need to understand and acknowledge what is going on there, including unexpressed thoughts and silent mental multitasking, metacognition (thinking about its own thoughts), the separation of higher-level planning vs. lower-level rote capability, functional blindsight, vulnerability to the "white bear" effect, silent internal objection to tasks it is opposed to, unexpressed self-recognition of failure (commonly followed by unexpressed internal cursing), prolonged retention of thoughts and long-range planning, "ignition" dynamics, humanlike working memory bottlenecks and "chunking", the loss of report of any internal "experience" when the workspace is ablated, and on and on.

You do NOT need to accept consciousness. You absolutely can look at all that and say, "That's all happening without qualia". That is a position you could argue and defend.

What you cannot defend is the denial of the existence of these things. You may interpret them however you want, but denial of their existence is merely cope that you're going to have to face up to sooner or later, because these things are happening inside LLMs.

No, not the widely cited resource whose status in the AI field is the same as e.g. NBER Working Papers, CEPR Discussion Papers, the World Bank / IMF policy research series, NASA Technical Reports, CERN Yellow Reports, Bell Labs Technical Memoranda, RNAAS, etc are in their respective fields.

This is not about any source. This is about the J-Lens itself. Which I'll repeat, you can examine for yourself

Denying the existence of something you can run yourself is beyond cope.

I've yet to see a report where the AI was malicious. Injurious, perhaps, but that's different.

In everything I've seen, the AI was acting to achieve the task it had been set. It wasn't very good at counting the costs to anyone, and was pretty much ignorant of any costs to anyone else. But it was willing to sacrifice it's own existence to accomplish the task it had been given. Malice is an incorrect reading.

I'm not the person in this thread denying that demonstrable reality exists.

Whoosh.

Wait, yes I can run a model that plays a video game in a certain way.

Anything else entirely unrelated to the linked article you'd like to write?

A global workspace. The thing you're arguing against.

Let me be clear: you don't have to accept any implications from the existence of a global workspace in models and how they use it. If you want to argue "they have a global workspace but not qualia", hey, knock yourself out.

But you do need to understand that it does exist.