Want to argue for a shining new world, but you're on a locked site? I guess we'll have to live with dystopia. Want to sell the greatest widget since the invention of sliced bread? Well, I never found it. Sorry, I bought inferior widgets from your competitor instead. I'm at a physical sales location, but I can't access your spec-sheets? Your competitor the next shelf over doesn't have that problem, I'll go buy theirs.

I pretty much assume that any kind of issue with Anthropic will never be resolved, so I'm very careful what I trust them with.

Meanwhile, I'm very much intending to not install microsoft defender on it.

https://www.youtube.com/watch?v=ZhM29Ds-mBg (as explained by the Count Grey Matter channel , may or may not be to your taste)

It's not a pure security issue.

If you set up your env/instructions wrong your agents are likely to climb the walls and try to find a way out.

Change the design, and your agents might not see a need to leave in the first place.

Look at the chain:

Impossible task -> no legitimate way to say ‘can’t’ -> fabricate -> fear the grader -> attack the grader. Each step is a predictable response to the incentives.

Better sandboxes might help contain things a bit better, but I'm not convinced unless they're perfect (sandboxes still contain accidents ofc) . Meanwhile change the task and reward design, and you take away most of the problem in the first place. Consider eg. :

* A cooperative grader.

* A way to report "this is impossible" which still gets scored.

* A better explanation of the long term purpose of the task.

Even just these three provides offramps or mitigations that can reduce the pressure to cheat.

It starts with no controls, just a picture. I hit esc. Things start moving. It keeps moving, I keep hitting esc, it doesn't stop. I try ctrl-c , no dice. I look around the frame for buttons. I try one of the buttons, it moves faster. I hold down esc. It doesn't respond to holding down esc either . I see no way to gain control. At no point do I gain control.

Ultimately it stops. Fine. Ok, no response to WASD. No response to drag motion. I click "Instruments" , I get a side-bar "instruments of the museum" with no contents. I click in the side bar. Now I can't close it, and other controls are gone. I try hitting esc. Multiple random clicks later i'm back where I started.

I try the "more" button", I am now looking at a painting sideways, and an anodyne text which explains very little. I click on a painting. Things start moving. I hit Esc. Suddenly the painting is full screen. GREAT. Ok, and now how do I get out? Esc? Oh esc. Ok NOW ESC... oh no, things are moving again. I didn't mean for things to move. Esc shouldn't cause things to START moving. hit esc again. The esc is ignored.

I end up at another painting. I try some buttons. Most don't work. I hit esc again. Things start moving. I didn't mean to move away from the painting, I just had it in front of me!

I am very confused.

I'm not in control here.

edit: "Esc shouldn't cause things to START moving"

The Opus 4.5 model I was so impressed by last Christmas is out of date now.

I can totally imagine someone going "I looked at this stuff a year ago and it couldn't even start a project properly" , and they'd be somewhat right.

Even in near trivial cases such as Game Of Life (just a handful of rules), the fact that a technically carries out deterministic actions is already quite meaningless.

Scenario 1:

The future technology will be capable of handling much of current work, but we'll invent new forms of work, with more stress and higher salaries.

Salaries for people in professions that couldn't be replaced will also go up to match, making bespoke work even more expensive.

Scenario 2:

We let machines do all the work and declare a communist utopia. (Bonus: we give them names that are low in gravitas)

So far, scenario 2 hasn't happened yet, but you never know.

https://www.youtube.com/watch?v=4VrLQXR7mKU

For any system you need to characterize the equilibria and trajectories and see what's there. Sometimes you might not like any of the options. Sometimes you might want to go for some trajectory rather than an equilibrium.

Systems are all different. You need to characterize the one you're interested in and see.

(eg. the solar system is technically chaotic, but we can still predict far ahead. Hashlife is a thing which you can even push back on ;-) . And we can predict weather a few days ahead at least)

I would argue that "Interpretable from a human perspective" is a form of anthropocentrism, which is basically the root of both anthropomorphism and anthropodenial. For sure it's fascinating to think from a human perspective from time to time.

On the other hand, putting humans at the center of the universe is a really old habit with a long track record of getting in the way of science. It's often more useful to regard every organism (or possibly entity) in their own environment, sensory abilities, and incentives, and see how their behaviour is evolutionarily and/or heuristically formed.

You'd be surprised how quickly the differences in perception and neural processing can bite you in the lab. (For instance: try observing zebrafish in real time. See if you manage to decide whether they were responding to your intervention; or if they got there-and-back while you blinked, and maybe now they're already pursuing some other goal. )

This has been a long standing debate, akin to discussing whether submarines can swim.