Thanks for posting this Owen, couldn’t agree more!
I often find myself referencing Eric’s work in specific contexts, in fact I just recommended it last night to someone working on AI control via task decomposition. I have been meaning to do a link-post on Why AI Systems Don’t Want Anything as soon as I get some free time, as it’s the biggest update I have had on AI existential risk since ChatGPT was released.
Eric has the keen ability to develop a unique, nuanced, first principles perspective. I agree his work is dense and I think this is one of its greatest virtues; when I recommend his blog I always have to comment in amazement that you can read the whole thing in an afternoon and come away with an entirely novel viewpoint on the world of AI.
This is a great overview of the virtues of his work, and the things his work leaves out. I especially like how you talk about deep study and the five steps you describe in order to internalize and reinvent. I think this also hints that what I see as Eric’s greatest strength; he looks at things very deeply in order to understand from first principles. I hope studying his work deeply in this way might help inspire others to develop similar first principles insight.
And I might add – not just a deep understanding of how the world is, but of how the world could be:
Large knowledge models for grounded, efficient information retrieval
Decomposable tasks for superintelligent systems without superintelligent agents
The potential for coordinated small models to outcompete large models on narrow tasks, making superintelligence potentially nearer but also safer
Structured transparency enabling verifiable commitments and de-escalation of races
Massively positive sum possibilities making coordination much more desirable
That is to say, I think Eric is a futurist in the best sense; he is someone who sees how the future could be and strives to paint a highly legible and compelling vision of this that at times can make it feel like it might just be inevitable, but at the very least, and perhaps more importantly, shows that it’s both possible and desirable.
Thanks for posting this Owen, couldn’t agree more!
I often find myself referencing Eric’s work in specific contexts, in fact I just recommended it last night to someone working on AI control via task decomposition. I have been meaning to do a link-post on Why AI Systems Don’t Want Anything as soon as I get some free time, as it’s the biggest update I have had on AI existential risk since ChatGPT was released.
Eric has the keen ability to develop a unique, nuanced, first principles perspective. I agree his work is dense and I think this is one of its greatest virtues; when I recommend his blog I always have to comment in amazement that you can read the whole thing in an afternoon and come away with an entirely novel viewpoint on the world of AI.
This is a great overview of the virtues of his work, and the things his work leaves out. I especially like how you talk about deep study and the five steps you describe in order to internalize and reinvent. I think this also hints that what I see as Eric’s greatest strength; he looks at things very deeply in order to understand from first principles. I hope studying his work deeply in this way might help inspire others to develop similar first principles insight.
And I might add – not just a deep understanding of how the world is, but of how the world could be:
Large knowledge models for grounded, efficient information retrieval
Decomposable tasks for superintelligent systems without superintelligent agents
The potential for coordinated small models to outcompete large models on narrow tasks, making superintelligence potentially nearer but also safer
Structured transparency enabling verifiable commitments and de-escalation of races
Massively positive sum possibilities making coordination much more desirable
That is to say, I think Eric is a futurist in the best sense; he is someone who sees how the future could be and strives to paint a highly legible and compelling vision of this that at times can make it feel like it might just be inevitable, but at the very least, and perhaps more importantly, shows that it’s both possible and desirable.