[Joel_Mckay]

Magnetic core memory final form were single large perforated plates with many conductors plated on the ferrite surface through-holes, and only the vertical stack of bus wires were threaded through the plates. This meant weaving was less of an issue, and higher >1kiB modules were feasible in a smaller area. The main draw back is it sill had destructive read-once access, so always had higher latency in addition to being slow.

The DDR market will adapt, as China grey market state fab smells the opportunity. They have been counterfeiting cmos chips for decades already, and dram is not as complex as people like to assume.

Neuromorphic computing will likely kick over the LLM sand pile at some point, and all that discounted hardware will need re-homed. We can wait for the bubble to run its course, and actual investors realize they were conned. =3

[UncleOxidant]

> Neuromorphic computing will likely kick over the LLM sand pile at some point, and all that discounted hardware will need re-homed.

I don't see a lot of work going on in neuromorphic - there was some work at Intel, IIRC. Not saying you're wrong, but just wondering where you think it's going to come from?

[Joel_Mckay]

> just wondering where you think it's going to come from?

Will likely evolve like any regular biological system, and consume translated LLM weight sets into its initial training condition 3D propagation structure.

The speed at which this occurs will likely initially be measured in weeks due to slower growth state writes, but once bootstrapped the GC is self-propagating.

I normally don't like to speculate, but the barrier to entry would actually be much lower than traditional silicon fabrication processes. It was an old idea from science fiction, that until recently was highly impractical. Asimov was likely wrong about the physical process, but not about how it is made.

https://en.wikipedia.org/wiki/Positronic_brain

Due to theoretical constant morphological changes under GC, one must acknowledge the inherent lack of safety such systems would pose. Have a nice day. =3

[fc417fc802]

I expect you'd carry out most such work in the form of simulations, only moving to hardware once you'd demonstrated an efficient algorithm. If I'm right about that then it would be easy for any corporate R&D on the topic to fly under the radar indefinitely.

On the academic side of things there's a steady drip of papers on things like spiking neutral networks so I'd say the general theme is being explored.

I figure if a breakthrough happens it will be overnight just like what happened with transformers.

[Tuna-Fish]

The next big shift will be HBF. All that DRAM holding essentially static weights that are read in nice, long linear reads in inference machines is wasted; if you had a proper interface to it you could replace it all with flash for a tenth of the cost.