Iteratively solves a bunch of PDEs to model how charge flows across neurons at a large scale. I read this as work on developing the computational tools to more effectively model the brain, understand certain phenomena.

I reckon electrical activity across neurons is only one piece of a large puzzle

Also Apple silicon is goated for being able to run these types of models crazy fast completely on device

Claude is amazing for things that are right at the frontier of your own ability. You don't know how to do it, but you know a good implementation when you see it. Trying to do things beyond that frontier quickly can devolve into misery.

My biggest hangup is the loss of art and culture as we know it. Why express when everyone perfectly understands each other? Why play when there's work to be done? Material conditions would absolutely be better, but it's the death of something fundamentally human. The hivemind is a new thing entirely

For experienced developers against codebases they are very familiar with. In newer codebases, coding can be a long tail of things you are not experienced with, which AI accelerates. A 50% productivity boost in this context I think is realistic. If you know everything already, AI inhibits you.

And Carol was forced to go through conversion therapy by her parents because they discovered she's lesbian. I feel like the show so far is about how trauma and ego separates us from others. I can't shake the feeling the happy ending here is Carol joining the Plurb willingly.

I would still definitely be pro joining the hivemind. And am super interested by people who are terrified of the idea. Being part of a universal consciousness is highly desirable in many spiritual practices. But I empathize with Carol because, it's so natural to cling to your individuality.

Not unlike running a decent gaming PC. Which seems right? Given the efficiency gains of an H100 versus consumer hardware, and batching over users.

Some back of the napkin math?
1 prompt takes around 1 second of H100 time at 70% which the authors state need ~1k watts
With batching maybe things are 5x slower (5s per prompt) so 200watts over 5 seconds.

So ballpark 200watts during inference? Maybe some baseline to extend to long running agents.

Thanks! Yeah fiber in reference to reducing power in transmission to the router.

2015 is before widespread fiber internet? I wonder if that has a significant effect on power usage for the wifi numbers if we were to measure things again today.