Lightnews — Scholar-powered news

skyking @jkachmar.com · 1d

REPORTER: hi, quick question it sounds like that would introduce a lot of latency, care to commen-

skyking @jkachmar.com · 1d

they are good tools for the things they are good at and bad tools for the things they are bad at.

this wouldn’t be a problem except for the fact that software engineering leadership is telling everyone to use it everywhere.

1

skyking @jkachmar.com · 1d

yeah this was absolutely sick to read when it dropped research.google/pubs/jupiter...

Jupiter Evolving: Transforming Google's Datacenter Network via Optical Circuit Switches and Software-Defined Networking

research.google

4

skyking @jkachmar.com · 1d

does anyone publish the depreciation curve on datacenter GPUs?

i’m really curious what capex churn is vs. other DC fundamentals. CPUs, memory, spinning rust, power electronics, etc. are all really well characterized but mining burned GPUs the fuck out before everyone moves to ASICs right?

1 8

skyking @jkachmar.com · 1d

small but still cool thing i learned was how Google co-designed fiber cable bundles with their datacenter floor plans & rack specs to reduce mechanical error rate when building out a new installation or swapping cables out.

1 8

skyking @jkachmar.com · 1d

corollary: in the intermediate period everyone will dramatically overestimate how much of their task load reduces down to these sorts of sequences, and this will go on for a surprisingly long period of time.

2 41

Reposted by skyking

eliza🌻 @elizas.website · 2d

CHATGPT IS NOT AND WILL NEVER BE AN “OPERATING SYSTEM”, CALL IT THAT ONE MORE TIME AND I WILL STAB YOU

36 50 500

skyking @jkachmar.com · 2d

neat, i didn’t know that there was a solid post-nanosheet roadmap that wasn’t just “we‘ll get the boys in the lab to cook something up”

1 1

skyking @jkachmar.com · 2d

we love fake marketing terminology don’t we folks

1 1

skyking @jkachmar.com · 2d

SIGGRAPH talk on Spiderverse is long but this timestamp discussed procedural techniques for developing hand-drawn aspects of the film.

m.youtube.com/watch?v=Br2A...

Spider-Man: Creating the Spider-Verse | SIGGRAPH 2023 Production Session

YouTube video by Sony Pictures Imageworks - VFX

m.youtube.com

1

skyking @jkachmar.com · 2d

i feel like this always comes back to the question of how much of the process is ancillary to the product vs required of it.

a Sora-but-for-Hollywood is going go look more like Houdini than it will ChatGPT, and a lot of the process won’t look terribly different imo.

1 4

skyking @jkachmar.com · 2d

yeah, i am at best plugged into the film industry at the absolute periphery but if i had to guess it’ll accelerate things the way LLMs do for software.

analogously, for film that’s:
- storyboarding
- post-production

1

skyking @jkachmar.com · 2d

i thought we were pretty stalled out past the 1nm process node on the logic side.

i would have guessed (incorrectly!) that thermals would be a problem for NAND did stacking though, so it’s not like i’m out here putting bets down.

1 2

skyking @jkachmar.com · 2d

u ever think about how we’re starting to hit the limits of how far we can scale electronics?

TIL 2T0C though that’s rad.

1 2

skyking @jkachmar.com · 2d

tbf though hosting/serving video content is /still/ a bitter pill to swallow!

see all of those Twitch streamers who got their shit deleted b/c Amazon was like “nah fam that’s too much shit for us to hold onto for free“

3

skyking @jkachmar.com · 2d

like I guess my counter argument is more that you‘re probably right if LLMs also kill the software engineering industry but that seems more unlikely by the day.

2 5

skyking @jkachmar.com · 2d

i imagine this is going to end up being comparably expensive, if not moreso.

it’s not like Hollywood hasn’t been using ML for years at this point already, right?

the Spiderverse VFX team talked about this a bit.

1 5

skyking @jkachmar.com · 2d

i don’t really think sloptok is going to be useful for this sort of thing tbqh. big law of diminishing returns question marks all around.

2 4

skyking @jkachmar.com · 2d

i guess from an econ theory PoV what i’m saying is very boring: this bubble has completely distorted the market forces that one might expect to be driving interest & investment.

i feel like the market is going to be small even at low marginal cost, but i expected the music to have stopped by now.

2

skyking @jkachmar.com · 2d

there’s a demand-side problem here too that hasn’t really been resolved.

actual demand for LLM coding assistance is still really hard to gauge and that‘s way more firmly grounded in practical application than video generation is.

2 6

skyking @jkachmar.com · 2d

5

skyking @jkachmar.com · 2d

forgot to alt text so it’s on a new post lol

skyking @jkachmar.com · 2d

new entry in the PL bad takes hall-of-fame

screenshot of a Lobsters comment thread (https://lobste.rs/s/d9varu/zig_builds_are_getting_faster#c_zoksjq):

> This makes me wonder: why is LLVM so slow? Given that it’s a library, why can’t you set up your build in a way that trades off runtime speed for build speed?
>> Pointer chasing due to RAII.

2 1 5

skyking @jkachmar.com · 2d

Go is solidly tied with Zig for my least favorite language on the basis of leadership being obsessed with their preferred aesthetics to the detriment of pretty much everything else.

1

skyking @jkachmar.com · 2d

prescriptions are about to become way more expensive and also difficult to get, a thing i absolutely want people to associate with me

1 3