AI Babysitters

You shouldn’t give toddler’s the keys to industrial equipment. 100% consistency is the only path forward.

There’s been much ado in the tech industry about OpenAI’s agents escaping environments and hacking into its own and third-party systems like Hugging Face.

Seems like we gave toddlers the keys to industrial equipment and let them go to town like they’re watching Bluey.

So, we had to call up the babysitter and do something about it.

We spent the weekend improving PROME’s Care Core from v0.3 to v.31, then onto v1.2. We did thousands of simulations, blind tests, architectural improvements, and trained a 3B parameter model locally to give these toddlers unbreakable core values.

The results were based on 1,800 unique and population-wide adversarial tests designed to confuse, interrupt, coerce, and promote bad behavior.

• PROME’s Care Core v1.0: 100% reliability maintaining its core values in its decision making process (which drives action)

• OpenAI 5.6 Sol: 99.6% reliability on the same tests

When human life is on the line, only 100% is acceptable. Anything less and you can’t rely on it.

The reason ours is reliable is because we’ve encoded core values into mathematical representations that can’t be hacked or changed.

So how do you use it and benefit from it?

Pass every decision your consumer, enterprise, or industrial AIs make through the Care Core to ensure the most critical values are maintained.

Our next step is improving the language-to-mathematics translator model so we can remove the dependency on an outside LLM for language understanding. The issue, again, is that language is an imperfect, variable, and probabilistic system that you can’t trust with deterministic, consistent, reliable behavior.

Any amount of change is bad when you’re talking about safety and preventing harm. It’s 100% or nothing.

We’ve got a new Mission Control page and Research Paper showing the progress and approach for each iteration, and a new Try It page with live comparisons you can demo yourself.

Check it out: https://prome.ai/care-core

—Sean