

Luckily China is slightly on the backfoot here and so it’s in their best interest to make all this stuff as public as possible
My hope is that (and I think somehat grounded based on engram and MOE) research will congeal towards fully modular networks that can run in a couple gigs of vram and connect to some neural memory store where all the knowledge is encoded…then the weights only have to learn algorithms to manipulate data, rather than learning the entirety of human knowledge in a single weight set…kinda like how computers went from running in warehouses to lounge rooms from the 70s to now


Anti ai being psyop and people using it to push agendas are not mutually exclusive
People are being paid large amounts of money to build these datasets, and the builders are incentivised for the training set to emulate real data, our OP is a clean example of this
For your point on my relativity comment, Llms are just as capable at producing referenced text as a human. And if the knowledge representation is air tight then hallucination is detectable. Eg llms solved fermats last theorem solved in lean.
Idk how ai winter fits here, did you check out the METR long horizon benchmark I mentioned. The bottleneck now is finding tasks it can’t solve.
If you build a ladder to the stars, you don’t have to achieve that goal to land on the moon.
Imagine yourself in 2020…what tasks would you have said was unachievable by a computer in that time? How many of those tasks are trivial in 2026? Fermats last theorem? proof against np hardness of 3 sat?
Anyway guess my point is, don’t kid yourself, these things are powerful, ignoring them and pretending their capabilities haven’t grown since 2022 is naive. And if we let oligarchs run this shit then the next century is gonna be fucking grim. Knowledge is power…arm up