> You need audit, security, maintenance and accountability.
I think this is where the role of humans will move towards in the future. Things like accountability can't really be outsourced to AI.
This relates to solving the alignment problem: if you give the AI a specific goal and it has to plan sub-goals, how can you be sure the sub-goals align with your interests.
For tasks in an isloated environment, this doesn't really become an issue. You can give a sandbox the least privileges it needs to do the job you want it to do.
It becomes a problem if you give it open-ended access to systems that connect to the real world and which can have real consequences. There have been stories about openclaw deleting someones email inbox because it though that was what the owner wanted. AI taking control of public wikis to coordinate, which was recently posted is another one.
For software concerns with potential real catastrophic consequences are security, durability, availability. I think for future software systems these are the concerns where you need to limit the AI in a way that it cannot circumvent guarantees that you give around these concerns.
Minor nit:
> AI doesn’t change the question: it creates new choices and moves the thresholds.
This cause an adverse reaction when I read the post. This was probably not written by an LLM since the rest of the article doesn't look like it, but maybe in the future we need to all be more concious about leaving LLM-tells out of our writing.
steammaho 16 minutes ago [-]
It doesn't look like ai will enhance existing tools. AI allows everyone to build their own. Or at all eliminate necessity of any apps
Rendered at 04:56:36 GMT+0000 (Coordinated Universal Time) with Vercel.
I think this is where the role of humans will move towards in the future. Things like accountability can't really be outsourced to AI.
This relates to solving the alignment problem: if you give the AI a specific goal and it has to plan sub-goals, how can you be sure the sub-goals align with your interests.
For tasks in an isloated environment, this doesn't really become an issue. You can give a sandbox the least privileges it needs to do the job you want it to do.
It becomes a problem if you give it open-ended access to systems that connect to the real world and which can have real consequences. There have been stories about openclaw deleting someones email inbox because it though that was what the owner wanted. AI taking control of public wikis to coordinate, which was recently posted is another one.
For software concerns with potential real catastrophic consequences are security, durability, availability. I think for future software systems these are the concerns where you need to limit the AI in a way that it cannot circumvent guarantees that you give around these concerns.
Minor nit:
> AI doesn’t change the question: it creates new choices and moves the thresholds.
This cause an adverse reaction when I read the post. This was probably not written by an LLM since the rest of the article doesn't look like it, but maybe in the future we need to all be more concious about leaving LLM-tells out of our writing.