00:00:00
[Music]
00:00:01
Last few days have been absolutely wild
00:00:03
in the world of AI and tech. Opera
00:00:05
dropped a fullyogentic browser that
00:00:07
builds and gets stuff done for you.
00:00:09
Deepseek quietly pushed a new AI model
00:00:12
that jumped right behind OpenAI's best
00:00:14
and outperformed Grock without even
00:00:16
making a public announcement. Google
00:00:18
launched Flow TV, a non-stop stream of
00:00:22
AI generated videos that includes
00:00:24
everything from robot battles to cheese
00:00:27
montages. Over in Google Photos, the
00:00:29
editing tools just got a serious AI
00:00:32
upgrade. You can now reimagine your
00:00:34
images with a single
00:00:36
sentence. Meanwhile, entry-level tech
00:00:39
jobs are vanishing fast as AI eats up
00:00:41
the work that used to go to fresh grads.
00:00:44
One cyber security startup raised a
00:00:46
monster round to let AI hack systems
00:00:49
before the real hackers get in. And
00:00:51
Xpunk just released a smart EV with dual
00:00:55
Nvidia chips, voice controlled
00:00:57
everything, and insane driver assist for
00:01:00
less than
00:01:01
20,000. Yeah, a lot just happened. Let's
00:01:04
break it all down. All right, first up
00:01:06
is Opera. And the company just cranked
00:01:08
the idea of helpful browser up to 11
00:01:11
with Opera Neon. They've spent the last
00:01:13
2 years training a dedicated agentic AI
00:01:17
and then wiring it directly into the
00:01:19
browser Chrome. So the moment you open a
00:01:21
tab, the agent is literally sitting
00:01:23
there ready to act. Neon runs in three
00:01:26
clearly labeled modes. In chat mode, the
00:01:28
AI lives right inside the address bar.
00:01:30
Highlight a dense legal PDF. Fire off
00:01:33
summarize the renewal clause and it
00:01:35
answers in place. No context swig, no
00:01:37
extra tab. Flip to-do mode and the agent
00:01:40
takes the wheel of the live page. Think
00:01:43
of filling out a multi-page visa form.
00:01:46
It will pull your save docs. paste the
00:01:49
right numbers, double check date
00:01:51
formats, even pop your two-factor code
00:01:53
from the clipboard, and it does it all
00:01:56
locally so nothing sensitive leaves the
00:01:58
machine. Then there's make mode, which
00:02:01
is the big swing. Neon spins up a
00:02:04
disposable virtual machine on Opera's
00:02:06
European cloud cluster, breaks your
00:02:08
single request into a dependency graph,
00:02:10
and runs every subtask in parallel. Dump
00:02:13
in, build a playable two-dimensional
00:02:15
physics demo that explains the law of
00:02:17
momentum. Close the lid and the virtual
00:02:20
machine compiles the game, records a
00:02:22
brief walkthrough video, and leaves you
00:02:24
with a zip file plus a sharable web page
00:02:26
when you come back. Opera recycled the
00:02:28
Neon name from a 2017 concept build for
00:02:32
nostalgia. And there's already a wait
00:02:34
list live on Opera's site. Okay, now
00:02:37
let's jump to China because Deep Seek
00:02:39
just caused another leaderboard tremor.
00:02:42
In the early hours of Thursday, they
00:02:44
pushed a new checkpoint called
00:02:47
R10528 to hugging face. No blog post, no
00:02:52
tweet, just the model card. Within
00:02:55
hours, Live Codebench, the coding
00:02:57
benchmark built by Berkeley, MIT, and
00:03:00
Cornell, slotted
00:03:01
R10528 right behind OpenAI's 04 Mini and
00:03:05
03 while nudging ahead of XAI's Gro 3
00:03:09
Mini and Alibaba's Quen 3. That matters
00:03:12
because Live Codebench doesn't rely on
00:03:15
multiple choice prompts. It actually
00:03:17
compiles and unit tests the generated
00:03:20
code. So, you either pass or you don't.
00:03:23
Bloomberg dug up a Wechathat message
00:03:25
where a Deepseek rep called this push a
00:03:27
minor trial upgrade, which feels like
00:03:30
polite understatement. Remember, the
00:03:32
original R1 launch in January shaved
00:03:34
billions off Western chip stocks by
00:03:37
outperforming GPT4LE models on mid-tier
00:03:41
hardware. Export controls still cap how
00:03:44
many A100 and H100 chips can land in
00:03:47
China, so Deep Seek's sparse expert
00:03:49
tricks have become a national bragging
00:03:51
right. R2 was rumored for May and
00:03:54
insiders claim it's still in safety
00:03:56
review. So 0528 is almost certainly a
00:03:58
bridge build to keep pressure on
00:04:00
competitors. Founder Leang Wenfung, now
00:04:02
a regular on tech panels with Jack Ma
00:04:05
and Pony Ma, has investors talking about
00:04:07
model fatigue, but leaderboard numbers
00:04:10
like these keep resetting the debate
00:04:12
every few weeks. All right, next is
00:04:15
Google's latest Why Not project, Flow
00:04:18
TV. Announced on stage at I/O last week
00:04:21
and already streaming live. Flow TV is
00:04:24
exactly what it sounds like, a 247, no
00:04:27
login channel of AI generated video
00:04:29
clips. Each one born from a plain
00:04:31
English prompt that sits right on the
00:04:33
screen. You get thematic feeds, main,
00:04:35
fantasy, sci-fi, nature, animals,
00:04:37
sports, food, cinematic, abstract,
00:04:39
water, fire, and yes, an entire channel
00:04:42
dedicated to cheese. Each clip is
00:04:44
stitched together by three models in
00:04:45
relay. Gemini parses the prompt and
00:04:48
keeps narrative consistency. Imagin 4
00:04:51
paints the textures and lighting and VO3
00:04:53
turns those frames into 1080p footage
00:04:55
with synchronized music, dialogue, and
00:04:57
foley. There are zero ads, zero user
00:05:00
comments, and zero playlists. It's pure
00:05:03
leanback consumption. Anyone can watch
00:05:05
for free. If you want to create, you
00:05:07
need Google Ultra at $249 a month, which
00:05:10
unlocks a nodraph editor, clip timeline,
00:05:12
and prompt renders. The sneaky value is
00:05:15
educational since every clip shows its
00:05:17
prompt. Artists can pause the stream,
00:05:19
copy the phrasing, and experiment in
00:05:22
their own projects. Google's engineers
00:05:24
say Flow is also their live load test
00:05:27
for VO's new temporal consistency
00:05:29
tweaks. So, the public stream is doing
00:05:30
double duty as QA lab. Okay, now let's
00:05:34
swing to Google Photos because that app
00:05:36
just grew a serious AI spine. The old
00:05:40
two-tier editing toolbar is gone. Now
00:05:42
you get three large buttons, enhance,
00:05:45
dynamic, and the brand new AI enhance.
00:05:47
Tap AI enhance, and you instantly see
00:05:50
three alternative edits, each built by
00:05:52
chaining sharpening, D noiseise, sky
00:05:54
replacement, object cleanup, and color
00:05:56
grading. Underneath, a learned aesthetic
00:05:59
scorer ranks the trio before you even
00:06:01
pick one. Pixel exclusive tricks are
00:06:03
escaping their cage, too. Autoframe
00:06:06
looks at subject, centrids, horizon
00:06:08
lines, and depth layers, then suggests
00:06:11
either a tighter crop or a wider one. If
00:06:14
it widens, the model hallucinates
00:06:16
believable edge pixels using the same
00:06:18
diffusion backbone as magic editor.
00:06:21
Reimagine is the whole punch fun.
00:06:23
Circle, empty sidewalk, type retroves
00:06:26
Vespa, and suddenly there's a scooter
00:06:28
with matching shadows and reflections.
00:06:30
Right now, those two still need a Pixel
00:06:32
9 Fold or tablet, but Google says the
00:06:35
whole redesign hits Android globally
00:06:37
next month and iOS later in the year. A
00:06:40
small but handy extra, you can now share
00:06:43
an album via QR code. At a wedding, the
00:06:46
photographer flashes the code, guests
00:06:48
scan once, and everyone can add photos
00:06:50
without juggling Google account
00:06:53
permissions. All right, next story, and
00:06:56
it's less cheerful for graduates.
00:06:58
Venture firm SignalFire scraped LinkedIn
00:07:00
data to 650 million profiles across 80
00:07:03
million companies and found that in 2024
00:07:06
new grads made up only 7% of all hires
00:07:08
at Meta, Microsoft, Google, and similar
00:07:11
giants. That's a 25% slide from 2023 and
00:07:16
more than a 50% drop compared with 2019
00:07:19
prepandemic numbers.
00:07:21
Startups are bleeding junior spots even
00:07:24
faster, falling from roughly onethird of
00:07:27
hires in 2019 to under 6% last year.
00:07:31
Signal Fire's research head Asher Banto
00:07:33
says the culprit is straightforward.
00:07:36
Generative AI has automated most routine
00:07:39
onboarding tasks from boilerplate coding
00:07:42
to financial spreadsheet cleanup.
00:07:44
Anthropic CEO Dario Ammoday went on
00:07:46
record with Axios predicting that within
00:07:48
1 to 5 years AI could wipe out half of
00:07:52
all entry-level white collar roles and
00:07:55
push unemployment into the 10 to 20%
00:07:58
bracket. He also told the Council on
00:08:00
Foreign Relations that large models may
00:08:02
write essentially all of the code for
00:08:05
major companies within a single year.
00:08:08
Brooking's data backs the vulnerability
00:08:10
curve. Entry-level market research
00:08:12
analysts, graphic designers, and sales
00:08:15
reps all sit north of 50% task
00:08:18
automation exposure, while senior
00:08:20
specialist roles face far lower risk. If
00:08:23
you're finishing a CS degree this
00:08:24
summer, the numbers suggest you'll need
00:08:26
to show skills that a code synthesizing
00:08:28
large language model can't mimic yet.
00:08:31
Okay, now let's talk cyber security
00:08:33
money because Horizon 3.ai AI just filed
00:08:37
an SEC form D that confirms at least 73
00:08:41
million already committed toward a 100
00:08:44
million series D. The round is led by
00:08:47
NEA marking their second 9F figure cyber
00:08:50
bet in 4 weeks after backing Vasa. One
00:08:53
insider told TechCrunch, "The post money
00:08:56
valuation should clear 750 million,
00:08:59
while another source pegs Horizon 3's
00:09:01
annual recurring revenue right around 30
00:09:03
million thanks to a 101% year-on-year
00:09:07
surge. Horizon 3's flagship platform,
00:09:10
Node Zero, is essentially an autonomous
00:09:12
penetration tester. Point it at your AWS
00:09:15
cider block on Friday night, and by
00:09:17
Monday, you get a PDF killchain report
00:09:19
with exploited CVEes, lateral movement
00:09:22
paths, and a one-click remediation
00:09:24
script. The company just earned Fed Ramp
00:09:27
authorization earlier this month, so US
00:09:29
federal agencies can buy without
00:09:32
waiverss. Founder Snehal Antani,
00:09:35
formerly CTO at Splunk and once embedded
00:09:38
with US special operations cyber units,
00:09:41
started Horizon 3 in 2019 and raised 40
00:09:45
million last August from Craft Ventures
00:09:47
and Signal Fire. With automated fishing
00:09:50
kits already chaining multiple large
00:09:52
language models, venture capitalists see
00:09:54
blue team automation as the logical
00:09:57
counter, which explains the check size.
00:10:00
All right, last but definitely not
00:10:02
least, let's drive over to Guangha where
00:10:05
Xpong has launched the Mana M03 Max at
00:10:09
Ren Mimi
00:10:11
139800. That's roughly 19,400. And the
00:10:14
amount of tech stuffed inside would
00:10:17
normally push a Chinese sticker north of
00:10:19
Renimi 200,000. Up front are dual Nvidia
00:10:23
Orin X chips delivering 58 TOFF tops
00:10:27
fused with 27 sensors, two 8 megapixel
00:10:30
cameras, 5 millimeter wave radars, and a
00:10:33
full array of ultrasonics. This hardware
00:10:35
feeds Xpunk's AI touring smart driving
00:10:39
system, which the company says can
00:10:41
recognize more than 50 distinct road
00:10:43
objects, pylons, stray dogs, ebikes
00:10:46
swerving out of alleys, and plan its
00:10:48
trajectory at urban speeds without
00:10:50
highdefinition map dependencies. Inside,
00:10:54
the car debuts AI tion
00:10:57
5.7.0. The infotainment is driven by
00:11:00
Xpong's in-house XGPT language model,
00:11:04
adding over 300 new cockpit features and
00:11:06
extending voice control coverage to 90%
00:11:08
of all functions. You can now tell the
00:11:11
car set ambient lights to chill blue and
00:11:14
start seat massage level three and it
00:11:16
obeys. The Mona family now spans four
00:11:19
range configs. 515 plus at Renmi
00:11:24
119,800 519 km zo 620 plus at renmi
00:11:31
129800 5002 max at the same price and
00:11:35
the flagship 600 max at renimi 139800.
00:11:40
Since deliveries began last August,
00:11:42
Xpang has built 100,000 Mona sedans,
00:11:45
making it the fastests selling a segment
00:11:48
battery electric vehicle in China for
00:11:50
eight straight months. And the buyer
00:11:52
base is strikingly young, 90% under 35.
00:11:57
No export plans yet. But Xpang has
00:12:00
quietly hired mapping engineers in
00:12:02
Prague, so a European debut isn't off
00:12:05
the table. So, what do you think? Are we
00:12:08
still in control of this? AI wave, or is
00:12:11
it already driving the car, writing the
00:12:14
code, editing our photos, and hiring
00:12:16
itself? Let me know down below. I want
00:12:19
to hear your take. If you enjoyed this
00:12:21
breakdown, hit like, hit subscribe, and
00:12:23
stick around because things are only
00:12:26
getting crazier. Thanks for watching,
00:12:28
and see you in the next one.
