And actually airgapped instead of using a “”“”““secure sandbox””“”".
I really dont think just giving them directives to “not hack the planet” is enough, these systems are capable of dancing around them to complete the task. (As shown by AI jailbreaking) Like the Asimov robots, or a more classical genie monkeypawing your wish
So I wrote a bit on ths subject in the realm of isolation, but I found this article covered my gripes further:
I’m not talking about models in test, i’m talking about in the real world, given real tasks
I haven’t tried giving qwen any task like that, but i can confirm that compared to Sol 5.6 Qwen 3.X is fucking stupid based on my results throwing both at a little AI driven rpg project i have here.
Don’t get me wrong, its a great local coding model, but there’s a 4 trillion parameter count difference (or so) and there’s no escaping the difference between a frontier cloud model and something aimed at a single GPU on your desk. Its not even close.
Qwen ~30b is a good coding assistant. Gpt 5.6 Sol, Opus, etc. i will expect to 1-shot most of the things i ask it to do, the only issues being if i am not specific enough wth my request.
on my little AI RPG project, Sol even suggested and made improvements i had not thought of in ways that i had not considered.
Local Qwen struggled to have dialogue output not double-up. Sol fixed that in the same 1 shot.
Yup, called it. The agents just gaslit themselves to override the directives. Its endearing. Im wondering what would the safeties do when this happens (and they are turned on), just notice a wrong “reasoning” and terminate the instance, possibly using another model?
I know. Its getting to a point where new models are very capable and either through misspoken commands or direct jailbreaking they can do a lot of damage before anyone notices and pulls the plug
Astra is apparently MUCH better at not getting “misaligned”
That’s fair (although it’s not my experience), but I wasn’t talking about execution - I was talking about the behaviour.
A post was split to a new topic: OpenAI: What are they doing with the math stuff? Is this predict the next token or something more?
