Longhorn

[Longhorn @never_released]

The RTX Spark NPU software stack appears very broken right now

https://t.co/3ol6t86md5

Cc @anemll

[Yuan Yi Zhu @yuanyi_z]

“a convicted murderer... who will never pass the good character requirements to practice”

*Monkey's paw curls*

[Sylvain Catherine @sc_cath]

While Zucman ponders why on earth billionnaires would react to a one-time 5% wealth tax and consider leaving California, back in France, his mentor is advocating for a combined package of 800bn euros, including a 50% tax on all gains exceeding 100 millions over the past 10 years.

[Longhorn @never_released]

(oh, and this was clearly done by scourging GitHub as they contacted multiple emails...)

[Longhorn @never_released]

if you do unsolicited spam once, don't aggravate your case by doing it a second time as a reply

thank you

[Yan Q. @armsq17]

It's the same pretty much everywhere with rent controls. A story from Berlin:

[Longhorn @never_released]

@Sebasti66855537 Models designed for client deployment already are designed with that constraint in mind to not needlessly lose perf during inference so I wouldn't bet on it.

And for FP8, Hopper's widening to FP32 accumulate is actually FP22, not using all the bits either.

[Longhorn @never_released]

@Sebasti66855537 NV has that throttle because training with 16-bit acc is very hard so it affects training perf compared to RTX Pro GPUs.

Plenty of inference workloads though already use non-widening ops and won't see a benefit.

[Longhorn @never_released]

@Sebasti66855537 note that this is just about widening ops

ie brings FP16 with 32-bit accumulate into the same throughput as FP16 with 16-bit accumulate

[Tobias Schneider @tobiaschneider]

rent control as commonly implemented is straightforward incumbency protection which naturally advantages old people and those who will not have to upgrade housing when they get jobs or start families

[antirez @antirez]

The point about local AI.

[flawedworld | @[email protected] @flawedworlddev]

Its funny seeing how conversations about multi tenanted hypervisors not being able to truly isolate tenants due to hardware side channels are now becoming common due to AI , when anyone who has seriously been in that area has known this for years.

[Longhorn @never_released]

Slower than HDD speeds, tradeoff of QLC is really not worth it, Apple. Especially not for your customers that pay the most

https://t.co/kvryDCp7Lx

[Longhorn @never_released]

> Once that cache and a secondary TLC-like buffer were exhausted, raw QLC writes averaged 79.4 MB/s and dropped as low as 25.6 MB/s.With the drive 60% full, the SLC cache shrank from around 250GB to 58GB, and raw QLC writes averaged 45 MB/s with a floor of 1.1 MB/s

[Jared Holt @jaredlholt]

For fuck’s sake

[Longhorn @never_released]

frupidity (noun): frugality to the extent that it becomes stupidity

[Anthony @Catholicizm1]

Diabolical boomer parents strike again

[Samuel Benson @sambbenson]

"Years earlier, [Romney had] bumped into Zach Braff on a flight and the sitcom actor had given him a useful piece of advice: for most not-quite-A-list celebrities, a simple baseball cap and a pair of reading glasses were usually enough to avoid being spotted in public."

One of the countless gems from @mckaycoppins' ROMNEY: A RECKONING

[Scott Lincicome @scottlincicome]

Just such a fun chart here from @greg_ip. If only we could tease out some meaning. Ah, well, nevertheless. 🤣