Cool! Does that mean that even open source models can be trained on this book to start self-improving without a full research lab support?

Interesting predictions! The examples are not internalizing the main conclusion though: if the companies will no longer exist, the job market won’t be flooded with CVs. There won’t be a company to order in bulk from either.. so no jobs, no money? What’s the next logical steps?

Tried renting gpus on vast.ai using Pinokio.co docker image to test some models for getting 3d models from images. It was.. not very encouraging. Has someone had success with Pinokio?

Well said. Soon, I think it will be so difficult for people to relate to a world view where we’re only present on earth

First publicly known AI trained off-world is a pretty neat thing

Making the worlds/levels so easy to implement must surely unlock a lot of creative games going forward?

Oh, this is neat

There's super-users and there's super-super-users. Good read and interesting to follow the progress over the year! For what it's worth - I agree that the top models has passed some inflection point where you just assume the code is good and start focusing more on the results.

I think the algorithm is quite good.. I like and comment GRPO fine tuning topics, and it shows me someone who’s testing the same thing as I am

I'd like to do the same, just to call my agent Jean-Claude, and then also for the model to be local so not actually a Claude..

Would be cool with a Colonization or Transport Tycoon game built on top of this mixed with the progress from the red mars book series. We could have tournaments, AIs training on how to beat the game and then just.. do it in real life

Maybe the prompt injection won’t be so bad after all?

Would be interesting to post an intellectually challenging task into this forum and see if they catch interest enough to model and solve it. Perhaps get them to discuss ARC-AGI tasks?

Unexpectedly, it seems pigs are flying now. We have to find something new to compare unlikely events to.