WEBVTT

00:00:00.000 --> 00:00:04.000
Every company in Silicon Valley says they're making the world a better place.

00:00:04.000 --> 00:00:08.800
We're not gonna do that. We're just gonna tell you what those companies actually did this week,

00:00:08.800 --> 00:00:14.200
which is usually much worse. Okay, tech news time. Google Research published a blog post this week

00:00:14.200 --> 00:00:17.280
on TurboQuant, a compression algorithm

00:00:17.280 --> 00:00:20.800
that shrinks an AI model's working memory by six times

00:00:20.800 --> 00:00:24.240
with zero accuracy loss, leading people to ask.

00:00:27.240 --> 00:00:30.480
Is the memory crisis over? Can we buy RAM yet?

00:00:30.480 --> 00:00:35.120
Well, we're here to give you a definitive, maybe?

00:00:35.120 --> 00:00:39.000
You see, TurboQuant targets the key value cache.

00:00:39.000 --> 00:00:43.120
Basically, when you chat with an AI, it keeps a running record of your conversation

00:00:43.120 --> 00:00:48.000
so it doesn't lose context. As a result, that cache eats more GPU memory

00:00:48.000 --> 00:00:51.400
the longer you talk, and you talk for a long time,

00:00:51.400 --> 00:00:55.080
making it one of the biggest bottlenecks in running these models at scale.

00:00:55.080 --> 00:00:58.280
Most existing compression methods carry memory overhead,

00:00:58.280 --> 00:01:05.040
which decreases their efficiency. But TurboQuant eliminates that through circle math.

00:01:05.040 --> 00:01:09.000
Is this real? Yeah, you said to simplify it, right?

00:01:09.000 --> 00:01:12.440
They use circle math, leading to up to an eight-time speedup

00:01:12.440 --> 00:01:17.040
on NVIDIA H100 GPUs. The underlying research is actually about a year old,

00:01:17.040 --> 00:01:22.400
but it's getting attention now ahead of its presentation at ICLR 2026.

00:01:22.400 --> 00:01:27.760
The news sent finance bros into a tailspin, and memory stocks like Micron and Sandisk

00:01:27.760 --> 00:01:31.600
tumbled in response. However, analysts say this is likely

00:01:31.600 --> 00:01:36.560
another deep-seek moment, where the release of a cheaper AI model back in 2025

00:01:36.560 --> 00:01:41.960
caused a trillion-dollar market panic before everyone realized that cheaper AI

00:01:41.960 --> 00:01:45.000
just means more people using AI.

00:01:45.000 --> 00:01:48.160
Sandisk's CFO told Bank of America analysts

00:01:48.160 --> 00:01:53.840
that he expects the improved efficiency will boost demand by making AI deployments cheaper.

00:01:53.840 --> 00:01:57.840
So if you were holding your breath, hoping that this killed the memory crunch,

00:01:57.840 --> 00:02:03.080
maybe hold your breath a little less hard, stop, stop holding your breath, you're dead, you're gone.

00:02:03.080 --> 00:02:06.880
Breathe, please live. Apple has officially confirmed

00:02:06.880 --> 00:02:11.680
they are discontinuing the Mac Pro, with no plans for future hardware.

00:02:11.680 --> 00:02:15.080
The company quietly pulled the former king of Macs

00:02:15.080 --> 00:02:20.360
from its website yesterday, where the buy page now just redirects to the Mac homepage.

00:02:20.360 --> 00:02:24.840
Kind of like how your landlord changes the locks while you're at work all the time.

00:02:24.840 --> 00:02:28.920
The iconic cheese grater design debuted in 2019

00:02:28.920 --> 00:02:32.200
after the thermal disaster of the Trashcan era.

00:02:32.200 --> 00:02:37.000
Well, I liked it. But Apple only ever gave it one Silicon update,

00:02:37.000 --> 00:02:41.640
the M2 Ultra in 2023, before leaving it to collect dust

00:02:41.640 --> 00:02:46.400
at its $7,000 starting price, or collect trash.

00:02:46.400 --> 00:02:50.080
The Mac Studio, which already has the newer M3 Ultra

00:02:50.080 --> 00:02:54.320
for $5,000 less, is now Apple's top-end desktop,

00:02:54.320 --> 00:02:58.280
with an M5 Ultra refresh expected later this year.

00:02:58.280 --> 00:03:01.840
Also gone is the infamous $700 wheel kit.

00:03:01.840 --> 00:03:05.320
No, the $300 feet kit. Oh, wow.

00:03:05.320 --> 00:03:11.400
And as of earlier this month, the older Pro display XDR and its $1,000 stand.

00:03:11.400 --> 00:03:14.520
Apple is clearly Marie condoing everything

00:03:14.520 --> 00:03:17.580
from the 2019 Pro launch, but how am I supposed to grate

00:03:17.580 --> 00:03:21.160
my incredibly expensive mature Coolia cheese

00:03:21.160 --> 00:03:25.320
while rendering on Silicon that's three generations out of date now?

00:03:25.320 --> 00:03:29.920
You didn't think of that, Apple, did you? Oh gosh, I'm sorry, Riley.

00:03:29.920 --> 00:03:33.360
I'll make something else screamingly expensive for you.

00:03:33.360 --> 00:03:40.840
Thank you, Tim Cook. God. AMD has announced the Ryzen 9 9950X3D2 dual edition.

00:03:41.840 --> 00:03:45.880
And yes, that is a real product name. It's the world's first desktop CPU

00:03:45.880 --> 00:03:49.120
with 3D vCash stacked on both chiplets,

00:03:49.120 --> 00:03:52.480
meaning it packs a staggering 208 megabytes

00:03:52.480 --> 00:03:56.840
of total on-chip cash across its 16 Zen 5 cores.

00:03:56.840 --> 00:04:01.520
That's enough cash to make it feel at home in the vault of the Bellagio in Ocean's 11.

00:04:01.520 --> 00:04:06.240
And you don't even need to squeeze an ordinary gymnast into a room service card to access it.

00:04:06.240 --> 00:04:11.120
AMD is pitching this less as a gaming chip and more as a productivity monster

00:04:11.120 --> 00:04:14.400
for game devs and creators with five to 10% gains

00:04:14.400 --> 00:04:18.920
over the already beastly 9950X3D, no to,

00:04:18.920 --> 00:04:22.640
in tools like Blender, DaVinci Resolve and Unreal Engine.

00:04:22.640 --> 00:04:26.000
Notably, AMD didn't share a single gaming benchmark,

00:04:26.000 --> 00:04:29.000
which really tells you where they think this thing shines.

00:04:29.000 --> 00:04:32.440
The trade-off though, a 200 watt TDP,

00:04:32.440 --> 00:04:38.480
the highest of any AM5 chip and a boost clock that actually dropped 100 megahertz.

00:04:38.480 --> 00:04:46.080
No price yet, but it launches April 22nd. Oh, and both AMD and Intel are raising CPU prices up to 15%

00:04:46.320 --> 00:04:51.640
thanks to the whole global supply crunch fiasco. So, game devs can breathe a sigh of relief

00:04:51.640 --> 00:04:55.360
knowing they won't have to optimize their games or their budgets.

00:04:55.360 --> 00:04:59.880
At least you can get a deal with our sponsor. Much like Erlich Bachman's incubator,

00:04:59.880 --> 00:05:02.960
we have crammed way too many quick bits under one roof

00:05:02.960 --> 00:05:08.000
and are hoping nobody asks any follow-up questions. A new revision of the RCS messaging standard

00:05:08.000 --> 00:05:14.800
has been finalized and it's a big one. RCS Universal Profile 4.0 adds native video calling

00:05:14.800 --> 00:05:19.960
right inside your messaging app, including group calls with up to 32 people.

00:05:19.960 --> 00:05:23.200
If you can convince 32 people to use the same messaging app as you.

00:05:23.200 --> 00:05:27.300
It also brings smarter media sharing, letting devices figure out which formats

00:05:27.300 --> 00:05:30.680
the other person supports, as well as rich text formatting,

00:05:30.680 --> 00:05:34.760
like bold, italics, and strike through, finally. We'll have to wait for tech companies

00:05:34.760 --> 00:05:38.440
to add these features to their default texting apps, but it's still cool.

00:05:38.440 --> 00:05:43.720
So while the cost of living might be skyrocketing, you'll at least be rich in text.

00:05:44.120 --> 00:05:47.780
Hmm. Apple just handed over the real identity

00:05:47.780 --> 00:05:52.840
of the user behind one of its iCloud hide my email aliases to the FBI.

00:05:52.840 --> 00:05:56.720
After a man named Alden Rumel used an aliased email address

00:05:56.720 --> 00:06:00.760
to send violent threats to FBI Director Kash Patel's girlfriend.

00:06:00.760 --> 00:06:04.680
Two horrible things. This kind of compliance might seem uncharacteristic

00:06:04.680 --> 00:06:11.400
for Apple, who famously blocked the government from breaking into the iPhone owned by a suspected terrorist.

00:06:11.400 --> 00:06:17.480
But unlike their device level encryption, alias to account mapping is stored on Apple servers,

00:06:17.480 --> 00:06:22.320
and as a result, can be subpoenaed. The hits keep coming for Patel, though,

00:06:22.320 --> 00:06:27.480
as Iranian hacker group, Handila, breached Patel's personal Gmail account

00:06:27.480 --> 00:06:30.480
and posted some of his most cringe vacation photos

00:06:30.480 --> 00:06:34.200
on Telegram. Oof! Handila did him dirty here.

00:06:34.200 --> 00:06:37.760
The vibe of these is less international man of mystery

00:06:37.760 --> 00:06:41.200
and more high school jazz band study abroad trip.

00:06:41.200 --> 00:06:46.120
It's an important experience. Sony is hiking PS5 prices across the board

00:06:46.120 --> 00:06:50.460
for the second time in less than a year, starting on April 2nd.

00:06:50.460 --> 00:06:53.600
The PlayStation portal handheld jumps up $50.

00:06:53.600 --> 00:06:57.840
The standard PS5 and digital edition both jump up $100,

00:06:57.840 --> 00:07:00.920
and the PS5 Pro climbs a cool 150

00:07:00.920 --> 00:07:04.720
all the way up to $900 for a console.

00:07:04.720 --> 00:07:09.640
Sony blamed continued pressures in the global economic landscape, which is fair.

00:07:09.640 --> 00:07:14.960
Lucky for us, though, Sony has decided to bravely pass those costs onto their consumers.

00:07:14.960 --> 00:07:18.240
They got mouse to feed. We, you know, we will survive.

00:07:18.240 --> 00:07:23.200
Wikipedia has officially banned the use of AI to generate or rewrite articles

00:07:23.200 --> 00:07:27.040
with editors voting overwhelmingly in favor of the new policy.

00:07:27.040 --> 00:07:32.000
LLM generated text was found to regularly violate the site's core context rules

00:07:32.000 --> 00:07:37.040
around verifiability and sourcing, which, yeah, hallucinations will do that.

00:07:37.080 --> 00:07:40.240
Editors can still use AI for basic copy editing

00:07:40.240 --> 00:07:43.640
and translation assistance, but only with human review.

00:07:43.640 --> 00:07:48.760
So maybe your middle school teacher was right all along. That's one for you, Madam Ruzel,

00:07:48.760 --> 00:07:53.760
but I should still be allowed to chew gum in class, so I still don't like you.

00:07:53.760 --> 00:07:57.120
And a delivery robot made by Serv Robotics

00:07:57.120 --> 00:08:01.200
has smashed through the glass wall of a bus stop shelter in Chicago

00:08:01.200 --> 00:08:05.160
before sheepishly blinking and then shambling away,

00:08:05.200 --> 00:08:08.240
shamefully. The footage, which is hilarious,

00:08:08.240 --> 00:08:11.720
went viral on social media and it comes amid growing backlash

00:08:11.720 --> 00:08:16.040
against delivery robots in the city. According to the Chicago Sun Times,

00:08:16.040 --> 00:08:21.600
a separate Coco Robotics bot smashed into yet another bus shelter that same week,

00:08:21.600 --> 00:08:25.800
making the successful delivery to smashed bus stop ratio

00:08:25.800 --> 00:08:29.480
lower than residents would hope. It's not what we're going for.

00:08:29.480 --> 00:08:33.120
Just like I hope to see you back here on Monday for more tech news.

00:08:33.120 --> 00:08:38.040
I'll be here next week pretending I understand what half these companies actually do,

00:08:38.040 --> 00:08:41.920
which as far as I can tell, is something to do with something called

00:08:41.920 --> 00:08:45.320
middle out compression? I don't know. Anyway, see you then.