So, 80 iterations... wonder how well it does after 80 MILLION iterations... or 80 BILLION.... Takeoff approaches!

We could likely be there with what we have if we simply put EVERY G     See More planet working in sync to run code like this, and iterated endlessly, simply spitting out a model to run on some other new GPUs in a new data center periodically...    See Less

Jürgen Schmidhuber again! The guys knew something.     See Less

This is so incredible interesting to see that they have built something similar to my own project. The difference is that my system doesn't constrain it to be code knowledge only but rat     See More for any type of knowledge in a specific area of expertise to be self improved upon using external knowledge. I hope to make a paper about it soon :-)    See Less

Great Darwin awards
Self improve AI with no guard rails.     See Less

its genetic algorithm around prompt engineering/tools its not singularity its an automated prompt engineer calm down     See Less

This sounds indeed interesting, exciting and seems like another step in the direction of the intelligence explosion. I had thought that more self-improvement was already used (like DGMs) in     See More t FMs, but apparantly they don't. The final point made about measures becoming 'bad measures' when they become targets, is telling. Because the benchmarks really seem to be used as TARGETS these days, all FMs are highly incentivized to tweak the model towards the 'measures' . And if they will now start to use DGMs for this, there is a likelihood of depreciating value of these benchmarks...
So - solving for real life problems is key, and the real world is 'messy' as remarked in the video. Let's wait and see.    See Less

Remember these machines operate on different time frames. And massively parallel all sharing the same memory. They can iterate and simulate millions times faster. As soon as a evolutionary p     See More introduced it will become no longer predictable. Absolutely Crazy times. Demis is going to make AGI before 2030.    See Less

Yeah the concept isn't super complex I had this thought back in the early 2000s. For safety of processes part of the idea developed into "dream states" like hallucinations but     See More ty to have a sandbox with more capacities - while there would also be an operating mode that is reflective of actions in the real world. While they could intersect a lot of the virtualizations would enable things that may not be safe. Since the modelling was based on aparati that is self reference and object reference there were certain situations where self reference conditions might not be good if they weren't sandboxed. Or likewise there were some actions that might be necessary but not desirable until the full outcomes of a system were understood.    See Less

Great for starters like me to build space shipe recursive improvement and building things inreal time🎉🎉🎉🎉     See Less

Usually I would be quite skeptical about this but ai had been blowing right past benchmark after benchmark. I'm excited and terrified.     See Less