Can you give me a concrete example of something fitting this that isn’t niche? I think few are arguing that there won’t be an appetite for real human actors for example, and I’m sure the rich will still choose to pay for (human) hand built things like luxury cars, but that still comprises a tiny portion of the worlds labour.
@Dan_Jeffries1 @Sergey_lll I see tweets and comments like yours often, but I am yet to see someone articulate a concrete vision of what human labour will be in a post AGI world where machines are both more intelligent and more physically dexterous than humans.
From my experience, the easiest way to get best one shot results is to use a mixture of models.
First define plan via claude, then have ChatGPT provide feedback, then give the feedback back to claude until convergence i.e ChatGPT signs off on it.
Having claude provide feedback on its own plan in the same thread eats up additional context, so a separate thread is preferable.
This back and forth picks up ALOT of ambiguity that streamlines a big implementation plan into something really cohesive which really improves how well it’s executed off the bat.
If you are wondering how these guys are able to drop screenshots of formulas into their open claws or Claude terminals and one shot profitable trading bots while you get nowhere close, it’s because they don’t and are simply lying.
Polymarket doesn’t even use LMSR, it uses CLOB for its market making model.
Hot take: FDM-1's computer use demos are a red herring. Yes, it can use Blender and browse websites. But GUIs are scaffolding we built for human limitations; teaching AI to navigate them was never the endgame. The real breakthrough is underneath: a model that watches raw video and learns to act. Computer screens are just the ImageNet, the richest source of watch-then-do data on earth. The driving demo is the tell. Pretrained on screen recordings, <1hr of fine-tuning, and it's steering a car in SF.
https://t.co/cYfyvtsXKy