Stochastically-Aware AI, Not Self-Aware AI

Long story, short: I suggest that we reserve the topic of self-aware AI for the philosphers and focus on real applications for and controls of stochastically-aware AI.

Long story, long: After some recent exchanges on HN, ethics have been on the brain. My thoughts circled back to AI and the results coming from these new diffusion models and the autoregressive transformers. How great would it be to have a reliable AI moderator, right?

It struck me that if you applied that AI moderator back on itself, you’d almost have a “self-aware” AI. In my opinion, none of these contemporary AI techniques can ever be conscious, so I thought, “well, more like stochastically-aware”. That phrase seemed semi-novel, so I ran a quick internet search. Hidden in the noise, I saw a patent for “Stochastically Aware Metrology and Fabrication” in the results.

Immediately intrigued, lo and behold, someone had worked out a patent [1] that almost exactly parallels what I was thinking about.

Granted, I’m not any kind of expert in machine learning, but it seemed like a relevant thought. A somewhat recent post on HN had recommended that people take the time to put thoughts ought there, even if not fully thought out, as it might still be enough to spur thoughts in others.

So, taking a little time to actually draw that parallel between metrology/fabrication and moderation/AI, I came up with this.

Stochastic-Awareness, in the context of generative AI systems implies the following:

- [ ] a generative process exists, deriving output from a supplied input

- [ ] the generative process has affordances for varying the output

- [ ] AND the generative process has affordances for censoring the generated output, in whole and in part

- [ ] an observability process exists, detecting categories of errors in output

- [ ] AND each category of error has a corresponding affordance in the generative process

- [ ] upon observing a category of error in generative output, the generative AI system uses the appropriate affordance to generate a new output

- [ ] the generative AI system censors any output containing a category of error

Of course, these new generative systems are taking the world by storm and there are clearly some real-world concerns for the public, so why not take it a step further and think a little bit about what government could do with such a technology, as is being done with open source software [2].

In the mode of “industry recommendations”, I came up with this.

- [ ] Congress could write a law requiring that all commercial and publicly interactive generative AI systems be stochastically-aware

- [ ] Congress could iterate on definitions of the appropriate categories of error, such as: indecency/immorality, intellectual property, representations of real people, abusive content, etc.

- [ ] Congress could delineate between domains of commercial use, personal use, academic use, and public use when making definitions of categories of error.

- [ ] Congress could determine a schedule within which to begin enforcement.

- [ ] Congress could determine what penal or civil punishments should apply for each category of error, across each domain.

I’m sure that the good people at OpenAI and experts in machine ethics have probably formulated much better ideas and have ongoing collaborations with lawmakers, but in the spirit of sharing and caring, I thought I might as well show HN this.

Of course, I’m just an observer on the sidelines, so any of the above might just be nonsense.

I do believe that humanity would be better served by shifting away from the idea of “self-aware AI” and towards the idea of “stochastically-aware AI”. As others have noted, AI is neither artificial nor intelligent [3] and the public shouldn’t be misled into worrying about SkyNet, as they are wont to do.

[1] https://patents.justia.com/patent/10474042

[2] https://news.ycombinator.com/item?id=32956218

[3] https://news.ycombinator.com/item?id=27413706