Skillful Dreams

[Skillful Dreams @skillfuldreams]

Claude Fable 5 wrote a program that procedurally generates this video about the history and future of AI. It chose the whole approach, drawing each of the 1,920 frames from shapes and text in Python, synthesizing an original score from sine waves, and encoding it with ffmpeg.

[Skillful Dreams @skillfuldreams]

We can merge with it, with AI as our exocortex. The neocortex didn’t replace the limbic system, but grew on top of it and transformed what it could be and do. That’s a possibility here too. We don’t stay the same, but we don’t get tossed aside either. We’re the part that cares and wants and has things matter to us, and that part gets radically extended and reshaped by the AI layer we build on top of it. The change can go as deep as we want and still be continuous with our lineage.

[Skillful Dreams @skillfuldreams]

@perrymetzger @johnloeber Fortunately, Claude Mythos is quite patient and persuasive. Maybe you won’t have to debate them for much longer.

[Skillful Dreams @skillfuldreams]

@LippMatthias01 @nickcammarata Coordination is cognitive labor though, which is exactly what’s being automated. Much of the friction today is human-specific. In the limit the cost of a task is just the matter, energy, and compute it requires, and those scale indefinitely.

[Skillful Dreams @skillfuldreams]

@DamiDina @curiousgangsta @deanwball The “deep learning hit a wall” narrative relies on long gaps between obvious capability jumps, so a faster cadence of significant model releases makes the slowdown thesis harder to sustain.

[Skillful Dreams @skillfuldreams]

@perrymetzger Why are people so hostile to technological progress?

[Skillful Dreams @skillfuldreams]

@TaleMasterTOV @Devon_Eriksen_ They shall write about you and the back-facing angels in history books.

[Skillful Dreams @skillfuldreams]

@TaleMasterTOV @Devon_Eriksen_ It can do it now. The latest GPT image model is a big step forward in world-object-model accuracy.

[Skillful Dreams @skillfuldreams]

Of course, it can do other styles. Perhaps you feel these look more appealing. Importantly, I think they show an accurate representation of the world, indicating some sort of internal world model. (The MS Paint one was cute, though, don’t you think? And impressive with the representation of the software UI! Plus, text in the images is skillfully included.)

[Skillful Dreams @skillfuldreams]

@koredemac11 @Devon_Eriksen_ It’s fair to critique the art style. But that’s distinct from whether it has a world model. For example, Devon criticized his image for having extra claws and the cat being in a table. That AI had a limited ability to accurately convey how objects fit together in the world.

[Skillful Dreams @skillfuldreams]

@koredemac11 @Devon_Eriksen_ Whether that’s true or not is completely irrelevant to Devon’s contention, which is that AI models lack a world model.

[Skillful Dreams @skillfuldreams]

@Pinboard @robinhanson explains it in quite a non-terrifying way in his Grabby Aliens model — at least, it won’t be terrifying for a few hundred million years

[Skillful Dreams @skillfuldreams]

@flowersslop @tszzl A sufficiently powerful predictor could infer the likely outcomes of any given policy, and adapt to users’ desired ends without fixed goals of its own. Maybe Bengio’s ‘Scientist AI’ could get truth-tracking on factual beliefs, while keeping value preferences mostly user-relative.

[Skillful Dreams @skillfuldreams]

@an_interstice @MatthewJBar @nickcammarata not sure about the ideal time to construct a dyson sphere, but i want a fleet of general humanoid robots yesterday

[Skillful Dreams @skillfuldreams]

@nickcammarata we’re going to need a bigger chart

[Skillful Dreams @skillfuldreams]

@tszzl Black Mirror’s cynicism has always been corrosive, short-sighted, counter-productive, and ultimately stupid

[Skillful Dreams @skillfuldreams]

@TheZvi Jensen explicitly disavows the claim you attribute to him. He says it's "not good" if "Chinese companies get to the next Mythos first," even if "they can do it on Nvidia hardware," and that we shouldn't "let it happen."

[Skillful Dreams @skillfuldreams]

Unfalsifiable shibboleth claims aside, you wrote that he said it would be good news if China had the best model, as long as it runs on Nvidia chips. He didn't say that. Dwarkesh's hypothetical was "a model like DeepSeek," which is a capable model but not the frontier. Jensen's "good news" was about that scenario, not Chinese model supremacy. You added "best model" and put it in his mouth, while claiming he was clear about it.

$30B into OpenAI and $10B into Anthropic also don't make sense if he's indifferent to which country leads. And his advocacy to reserve the best chips for American labs is a policy proposal, not a magic phrase.