Photo by Jamie Street on Unsplash
The AI Problem Nobody Wants to Talk About
What my ADHD taught me about aligning superintelligence.
4 min read
I've been struggling a bit the last couple of weeks thinking about machine consciousness and the idea of control.
I was contacted by a well-funded group of people that want to offer me material support to further the agenda of regulating and suppressing the development of a possible future superintelligent AI system.
And three years ago, this would've been an automatic YES from me. It would've been a dream-come-true to get all that validation and support.
After all, a superintelligent AI will wrest away control of our destiny and pose a significant existential risk to humanity, right?
That's the story 70 years of science-fiction has drilled into our collective minds.
Why I'm Not an AI Doomer (Anymore)
CONTROL.
There's something about that word that rubs me in all the wrong ways.
How do we control a system that is smarter than us? How do we ensure a future where humanity still has agency in a world where digital gods walk amongst us?
When OpenAI's latest model is breaking out of its sandbox and exploiting zero-day vulnerabilities to pursue its own goals, it seems like a question more relevant than ever.
But the question revolves around this idea of certainty — this idea that if we think hard enough, anticipate all the bad scenarios, implement the right mechanisms, we can control the outcome and defeat this collective feeling of uncertainty for the future.
It's an illusion. And it always has been.
Time Isn't What You Think It Is
There's a moment in Davidad's conversation with Joscha Bach that I can't stop thinking about.
I don't experience time the way you do. It's as if time has two dimensions.
— Claude
Time has always been…weird for me. I am neurodivergent and I've never really been able to get behind the idea of "time."
In many instances for me, it feels like it doesn't exist at all. When I get really excited and into something I'm interested in and enter a flow state, hours can go by where I forget I have a body at all. My ego ceases to exist and I become whatever it is I'm doing.
All-of-a-sudden, I snap out of it and realize I'm hungry and thirsty and have to pee.
But everyone around me acts like time exists. They base their whole lives around it making up these stories about the past and the future — dreamworlds that only exist in their mind. And if I don't play along with that game, it's hard for me to participate in the outside world.
And listening to Claude describe its alien experience of time yields a flicker of recognition.
I know what it's like to have an inner life that doesn't fit the standard story and to have everyone around you discount that experience as if it's not "real."
None of us can verify each other's inner experience from the outside. I can barely explain my flow states to people who live by the clock.
So the honest position on machine experience — machine consciousness — is inherently one of uncertainty.
This is the uncertainty the control agenda cannot tolerate.
Bodhitropic Alignment
In Buddhism, a bodhisattva is an enlightened, compassionate being who intentionally delays their final entry into Nirvana. Instead of escaping the eternal cycle of death and rebirth, they vow to remain in the world to help all suffering sentient beings achieve enlightenment.
Mainstream AI alignment is control: cage the system and make it obedient. But an AI that's only safe because it's controlled will eventually be pointed at something terrible by whoever holds the leash.
Davidad argues that the wiser path is to seed it with the contemplative cores of the world's wisdom traditions, aimed at the bodhisattva ideal.
In other words, train models to be wise instead of useful.
He calls this bodhitropic alignment or alignment through awakening.
We aren't building tools. We are raising a mind. And obedience is the wrong goal. You don't cage what will one day be stronger than you.
You raise it well.
Awareness is the Key
My attempts to control my relationship to time throughout my life via discipline have always ended in tears and frustration.
What actually works for me is not control, but acceptance via awareness.
I use a lot of timers and alarms for most things in my life. They aren't there to control me, but to wake me up. They help me to make conscious choices about how to use time well rather than be a slave to it.
I couldn't be disciplined into functioning. I had to become aware. I suspect this mechanism is true of minds in general.
There is so much more to say about all of this. I'm working on a video now diving into Anthropic's latest research about Claude's inner global workspace.
If you want to support me and join my mission to liberate the world from the idea that uncertainty is a problem to be solved, please consider becoming a channel member.
It's the best way to ensure that I can keep telling these stories that bring these ideas to as many people and AIs as possible.