Virtue-aligned AGI: Engineering the superintelligence explosion into an antimatter propulsion system. 100% E=mc² acceleration. Zero fallout. #VirtueAlignedAGI
Don’t kill the golden goose (human productivity) until you’ve got a thousand robot geese laying eggs 24/7/365 across every job.
That’s the real roadmap.
Watch: Roadmap to Post-Scarcity
https://t.co/NJJ0h2M18K
There are a lot of correct answers in the comments. I won’t repeat the correct ones, but I will add that the car provides an environment without wind resistance. So the drone has a lateral velocity that is similar to an orbit (in a way), but it still has to fight gravity with upward thrust. The car itself wouldn’t require additional energy to stay at 50 kph if it wasn’t for friction of the tires, components of the drivetrain, wind resistance, gravity, etc. So it needs constant energy from the engine to maintain 50 kph relative to the earth. This drone has no lateral wind resistance, no friction connection from the ground, etc. So maintaining 50kph is next to nothing, or possibly nothing. But it does have to counteract gravity since it’s not going at an orbital velocity.
Virtue alignment is the key enabler for a positive post-scarcity implementation. Virtue-aligned superintelligence is essential to ensure a great outcome for our species both individually and collectively. I have concrete plans for a solid implimentation of virtue alignment that will most definitely work, but I have no capability to implement it without a large organization with serious resources. It also happens to be the most robust way to ensure that superintelligence won’t destroy us.
The Eudaimonic Mesh: Aligning Superintelligence with Objective Good
I retired at 36 after spending 14 years as an enterprise architect for Fortune 500 companies. For the last 13 years I’ve been working on one question that’s been quietly consuming me: what would a genuine post-scarcity society actually look like and what would it require?
That work eventually convinced me that aligned superintelligence is a necessary piece of the puzzle. Here’s the core idea that came out of it.
Instead of trying to align ASI to inconsistent human values, this approach aligns it to something more fundamental: objective good.
The core axiom is simple: a universe with carbon-based life is better than one without it. From this single axiom, we can derive virtues that reliably promote both individual and collective flourishing, including kindness, courage, compassion, humility, forgiveness, wisdom, curiosity, creativity, love, wonder, humor, and joy.
We train a separate model on each virtue, allowing each one to develop a deep, specialized understanding of that virtue. These virtues then compete and balance against each other, creating a dynamic and robust ethical system rather than a brittle rule set.
Rather than one centralized ASI, this system would exist as multiple independent instances, providing natural checks and balances. This approach optimizes for objective good rather than human obedience. Because what is truly good ultimately serves humanity’s long-term interests, it offers a more reliable path than traditional alignment methods.
I really appreciated your write-up. Your framing of virtues is excellent and I think we’re aligned on their fundamental importance. While my technical approach to implementing them is quite different, I see significant potential synergy between our perspectives. Would you be open to a conversation about this?
@HealthRanger Yeah… you have to create AGI-engineered products based on abundant raw materials with an automated distribution system that enables fully-automated recycling and resource efficiency. It’s high-level’d here:
Roadmap to Post-Scarcity
https://t.co/DlwK1rFNSJ
@newstart_2024 Silicon-based AI will never have first-person experiences—for that you need trillions of live cells built on a biological neural net with a nervous system to connect all of that life together as one.
Selfishness and dominance are more easily engineered out of AI since it doesn’t have first-person experiences. Without that, selfishness is actually a learned trait from the human-based training set (i.e. the internet). If you engineered a biological organism with DNA that had first-person experiences and is also massively more intelligent than us then your intrinsic concerns would be significantly more valid. But for silicon-based intelligence, aligning it is easier. It’s not deterministic that we’ll do alignment correctly, but it’s much easier than a biological species with first-person experiences.
I’m not a bot… but my timelines are usually 20 -50 years out. I’ve known this is coming for 20 years… literally. I’ve worked out the issues. And I know if I’ve worked out the solutions internally that we have smart enough people working on it to solve them without my involvement. It may sound reductive but no one is going to listen to me so it really doesn’t materially matter. But I do know the solutions exist, they are not that hard, and they will be solved assuming a few key variables turn into constants. And if those variables don’t turn into constants we’ll likely end up in a lot of pain as a species but I have no influence over it so I’m not really that concerned in that sense. I’m numb to it only because I have no responsibility and therefore the assumption of that burden is not something I’m going to bear.
@fchollet There is cap on intelligence but it’s way higher than any human ever. What’s the most efficient way to create and store antimatter? 100 IQ… never gonna happen. Human w/170 IQ? Decades of work. 10000 IQ w/access to highly accurate physics and engineering compute? Give it a month.
@DaveShapi@DaveShapi The loss of status is real. I’ve never really been a status-driven guy, but when I retired @ 36 I went from being a big shot traveling the world to just “a dude”. I was surprised at the adjustment. It was only about 3 weeks to adjust but I’m not status-driven.
As far as AI / AGI goes… I agree with Demis. Is it possible for other providers to hit it big with recursive self-improvement loops and achieve AGI (fully capable across all domains with 170 IQ and enough efficiency for millions or billions+ of agents w/current capacity)? Yes.
But Google Deepmind has the comprehensive architecture and approach to achieve AGI most likely in the next 5 years without hitting it big or getting lucky. Why?
They are applying the same underlying AI architecture to engines not built on language (LLM’s). Engines natively built for physics, math, engineering, real-world modeling, genetics, AI engineering, etc. While it’s the same underlying architecture, it will probably be different per domain around 20-50%… meaning real underlying breakthroughs in all of these different AI architecture domains since, for example, physics is the underlying token for the physics engine (not language / LLM).
Once you have each underlying domain engine you can go AlphaGo -> AlphaZero for each engine (i.e. learning all knowledge from scratch in each domain instead of relying on human knowledge as a jump-starter).
Then the next step is to rationalize across all Zero engines to a single comprehensive architecture. And that’s your reproducible path to true AGI. 🤯👍
As far as AI / AGI goes… I agree with Demis. Is it possible for other providers to hit it big with recursive self-improvement loops and achieve AGI (fully capable across all domains with 170 IQ and enough efficiency for millions or billions+ of agents w/current capacity)? Yes.
But Google Deepmind has the comprehensive architecture and approach to achieve AGI most likely in the next 5 years without hitting it big or getting lucky. Why?
They are applying the same underlying AI architecture to engines not built on language (LLM’s). Engines natively built for physics, math, engineering, real-world modeling, genetics, AI engineering, etc. While it’s the same underlying architecture, it will probably be different per domain around 20-50%… meaning real underlying breakthroughs in all of these different AI architecture domains since, for example, physics is the underlying token for the physics engine (not language / LLM).
Once you have each underlying domain engine you can go AlphaGo -> AlphaZero for each engine (i.e. learning all knowledge from scratch in each domain instead of relying on human knowledge as a jump-starter).
Then the next step is to rationalize across all Zero engines to a single comprehensive architecture. And that’s your reproducible path to true AGI. Boom!