>AI is still at its infancy>We are still using hardware meant for video games to power AI. Once we crack either one or all of organic/analogue/quantum computers, it's over>Companies are still focusing on making stuff that hits headlines like AGI. Once they move to actual profitable AI like bespoke specialised AI trained on private data, it's overIt's insane that even in its infancy, it's still so impressive. It's gonna get so much worse after the "bubble" pops.
>>17082448SNNs are already used in a lot of applications. they didn't yet cracked it for LLMs. pretty sure they're working on it. huge efficiency gains. like you wouldn't believe. like orders of magnitude. it's pretty nuts. but they didn't yet crack it for LLMs.look up SNN neuromorphic hardware, they're analog chips. they do have some digital for interfacing, but the "thinking" bit is fully analog. quite interesting hardware.
>>17082501Wow. You're so smart and technologically informed. But have you been taking your Thorazine as prescribed? Take it. You wouldn't believe the YUGE efficiency gains for your brain to cull spurious thoughts.
>>17082501>SpikeLLM..etc. and hybrid quantization can boost spiking mechanisms into the 7–70B parameter range>neuromorphic chips (loihi 2, FPGA MatMul) juices 1.5B-parameter models with ~10–20× more energy efficiency vs GPUsIt's coming alright, Jensen be shaking.
If there are any further speedups it just means a faster firehose of slop. This bubble is going to pop and all your retarded AI nigger faggots are going to kill yourselves.
>>17082658What actually causes this much denialism?
Google actually designed early neural-network chips optimized for the CNNs they were applying for Google Image Search , they were called Tensor Processing Units (TPUs). And a lot of modern chips, like the apple M chips, have designated NN architectures built into them in addition to graphical parts of the chip. The problem is that the NNs that current LLMs use vastly larger sets of weights than older NN applications, and GPUs were already designed to support very big matrices which are basically what 4k and 8k chips are designed for. And in fact, the modern GPUs are pushing the limit of the size of single-wafer layouts, which in turn push the limits of what you can fit on a reasonable 8U or 12U rack electronics unit.To move beyond this we may be looking at standard server racks needing to be taller and wider, meaning a complete replacement of the entire industry standard electronics specifications, and remaking every server rack and every board and every cooling system and every data center . Instead GPUs are basically saturating what the current infrastructure supports so they will probably be dominant at least until semiconductor lithography attains smaller than the 2-1 nm process, but already this pushes the physical limitations of silicon semiconductor circuits, especially with regards to the wavelengths of UEV optics
>>17082982sorry I meant EUV, not UEV.
>>17082524sarcasm is the lowest form of wit. focus on fixing that
