|
|
|
|
|
|
by ggm
801 days ago
|
|
|
We'd still be using triple-DES to protect data, arguing that the NIST time to break it was still far out beyond. And hash functions would be like the CRC32 in TCP, not the modern stuff. CISC computers which did more in parallel per instruction would be common because they existed for concrete reasons: the settling time for things in a discrete logic system was high, you needed to try and do as much as possible inside the time. (thats a stretch argument. they were what they were, but I do think the DEC 5 operand instruction model in part reflected "god, what can we do while we're here" attitudes) -We'd probably have a lot more Cray-1 like parallelism where a high frequency clock drove simple logic to do things in parallel over matrices, so I guess thats GPU cards. |
|
|
Definitely sounds right that we'd get an earlier, heavier emphasis on parallelism and hardware acceleration. I'm guessing the slower speed of causality also applies to propagation delay and memory latencies, so there wouldn't be new motivation for particular architectural decisions beyond "God please make this fast enough for our real-time control systems or human interaction needs".
If we got deep learning years or decades earlier, that also seems scary for AI existential risk, as we are just barely starting to figure out how the big inscrutable matrices work, and that's with the benefit of more time people have had to sound the alarm bells and attract talent and funding for AI interpretability research.