00:00 - 00:03
here's all the AI news that came out
00:01 - 00:04
this week that I thought was important
00:03 - 00:06
to talk about starting with what was
00:04 - 00:09
probably the biggest news in the world
00:06 - 00:11
of AI this week which was the release of
00:09 - 00:14
grock 3 from xai they launched this at
00:11 - 00:17
like 8:30 at night on Monday in what was
00:14 - 00:19
probably one of the most awkward launch
00:17 - 00:22
videos I've ever seen hi everyone my
00:19 - 00:24
name is Igor lead engineering at XI I'm
00:22 - 00:27
Jimmy Paul leading research I'm Tony
00:24 - 00:30
working on the reasoning Team all right
00:27 - 00:32
I'm Elana I don't do anything
00:30 - 00:34
I just show up
00:32 - 00:36
occasionally yeah as of right now grock
00:34 - 00:38
3 is ranked the number one large
00:36 - 00:39
language model over on LM Arena if
00:38 - 00:42
you're not familiar with LM Arena it's
00:39 - 00:44
basically a blind taste test for large
00:42 - 00:46
language models you give it a prompt it
00:44 - 00:48
gives you two responses from two
00:46 - 00:50
different language models not telling
00:48 - 00:52
you what model you're working with you
00:50 - 00:55
pick your favorite and this leaderboard
00:52 - 00:57
is created based on so many people doing
00:55 - 00:59
this blind test and picking their
00:57 - 01:02
favorite model again we can see that
00:59 - 01:04
grock 3 code named Chocolate is
01:02 - 01:07
currently number one from xai and to be
01:04 - 01:09
honest grock 3 is a really impressive
01:07 - 01:11
model I know there's a lot of people out
01:09 - 01:12
there that will never touch it they'll
01:11 - 01:14
throw the baby out with the bathat
01:12 - 01:16
because they know that it's elon's
01:14 - 01:18
company behind it and they just will not
01:16 - 01:21
touch anything Elon uses but just from
01:18 - 01:23
like a peer data standpoint it's
01:21 - 01:24
actually a really really good model we
01:23 - 01:27
can see that it pretty much beats out
01:24 - 01:29
all of the other state-of-the-art models
01:27 - 01:32
in competition math it beats out all the
01:29 - 01:34
other models in graduate level Google
01:32 - 01:36
proof questions and answers so basically
01:34 - 01:38
questions that you can't get the answer
01:36 - 01:41
by just Googling them beats out all the
01:38 - 01:42
other models in Live code beats out all
01:41 - 01:45
the other models in code generation from
01:42 - 01:47
the Live code Benchmark and in the
01:45 - 01:49
multimodal understanding it pretty much
01:47 - 01:51
matches the other models interesting
01:49 - 01:55
that in the code Benchmark they don't
01:51 - 01:57
compare it to Claude Sonet 3.5 which
01:55 - 02:00
most people who use AI for coding are
01:57 - 02:02
using Claude 3.5 Sonet Rex here points
02:00 - 02:04
out that they also didn't compare it to
02:02 - 02:07
like the regular 03 model the one that
02:04 - 02:08
Sam Alman demonstrated but we don't
02:07 - 02:10
actually have access to yet and if you
02:08 - 02:13
add that model into the mix well it
02:10 - 02:14
looks like 03 would still beat out Gro
02:13 - 02:16
but like I've mentioned people have been
02:14 - 02:18
getting really really impressive results
02:16 - 02:20
out of it they gave it a prompt create a
02:18 - 02:22
game that is a mixture of two classic
02:20 - 02:24
games make it in P game and make it look
02:22 - 02:27
pretty so they even let it pick which
02:24 - 02:29
two games to sort of blend together and
02:27 - 02:31
we can see down here that it actually
02:29 - 02:33
took pong and it took breakout and it
02:31 - 02:36
merged those two games together so it's
02:33 - 02:38
pong with like breakout in the middle
02:36 - 02:41
and this is basically a new game that
02:38 - 02:42
grock invented when given the prompt
02:41 - 02:44
create a game that is a mixture of two
02:42 - 02:47
classic games it's got a deep search
02:44 - 02:50
mode which is very similar to Google's
02:47 - 02:52
deep research open ai's deep research
02:50 - 02:54
perplexity deep research except they
02:52 - 02:56
dropped the re they're just calling it
02:54 - 02:58
deep search but this will actually
02:56 - 03:00
search the web and find a whole bunch of
02:58 - 03:03
sources to help answer the the questions
03:00 - 03:05
that you asked including sources on X
03:03 - 03:07
itself grock also has a think mode so
03:05 - 03:10
very similar to what you're going to get
03:07 - 03:12
out of 01 or 03 or deep seek R1 it will
03:10 - 03:15
actually use that sort of Chain of
03:12 - 03:17
Thought reasoning that test time compute
03:15 - 03:19
where after you give it a prompt it will
03:17 - 03:21
think through and sort of double triple
03:19 - 03:23
check its responses before finally
03:21 - 03:24
giving you its final output and so far
03:23 - 03:26
among the people that have been using
03:24 - 03:29
grock they've really been liking it
03:26 - 03:32
Andre karpathy who used to work over at
03:29 - 03:34
open AI says that grock 3 clearly has an
03:32 - 03:36
around state-of-the-art thinking model
03:34 - 03:38
he goes into quite a bit of detail about
03:36 - 03:40
what it did well and what it didn't it
03:38 - 03:42
didn't solve his Emoji mystery question
03:40 - 03:45
it solved a few Tic-Tac toe boards it
03:42 - 03:47
did really great with the gpt2 paper and
03:45 - 03:49
looking up information from the paper
03:47 - 03:51
and we can see in this image here he
03:49 - 03:54
asked it to generate an SVG of a pelican
03:51 - 03:56
riding a bicycle the top three images
03:54 - 03:59
are what grock 3 gave him compared to
03:56 - 04:02
what Claude 3.5 Sonic gave him GPT 40
03:59 - 04:04
mini and all these various other models
04:02 - 04:07
the cloud 3.5 Sonic 1 isn't too bad but
04:04 - 04:10
I would say grock 3 is clearly the
04:07 - 04:12
winner of these various models here we
04:10 - 04:14
also got a sneak peek of grock 3 voice
04:12 - 04:16
mode originally it was going to launch
04:14 - 04:18
with grock 3 but Elon said it might be
04:16 - 04:20
like a week late so maybe we'll see it
04:18 - 04:24
next week but here's the teaser they
04:20 - 04:27
gave us hey Gro what's up can you hear
04:24 - 04:29
me I'm so excited to finally meet you I
04:27 - 04:30
can't wait to chat and learn more about
04:29 - 04:33
each other I'll talk to you soon and
04:30 - 04:35
that's all we got but Elon did say that
04:33 - 04:37
it was going to understand emotions and
04:35 - 04:38
inflection and pacing and things like
04:37 - 04:41
that the voice about a gr doesn't it's
04:38 - 04:43
not just voice text it understands tone
04:41 - 04:44
inflection pacing everything it's wild I
04:43 - 04:46
mean it's like talking to a person now a
04:44 - 04:49
few other things of note about this
04:46 - 04:50
launch here scoble here points out that
04:49 - 04:54
the thing to really pay attention to in
04:50 - 04:56
AI is the learning speed xai grock is
04:54 - 04:59
learning way faster than any other
04:56 - 05:02
models a big part of that is the fact
04:59 - 05:05
that they originally started training on
05:02 - 05:08
100,000 gpus and they doubled it to
05:05 - 05:11
200,000 gpus that's 15 times more
05:08 - 05:13
compute power from grock 2 to grock 3
05:11 - 05:14
xai also announced that they're
05:13 - 05:17
launching a game studio and they're
05:14 - 05:19
going to be using AI to help them
05:17 - 05:20
develop these games we watching an AI
05:19 - 05:22
gaming studio we're announcing it
05:20 - 05:24
tonight let's go now the good news is if
05:22 - 05:26
you are somebody that wants to use grock
05:24 - 05:30
you can use it for free right now over
05:26 - 05:33
at gro.com on x.com or inside of the
05:30 - 05:35
grock mobile app now Elon says for a
05:33 - 05:37
short time grock 3 is available for free
05:35 - 05:39
to all we don't know how long a short
05:37 - 05:41
time is but if you want to get in and
05:39 - 05:43
play with it the thinking mode the Deep
05:41 - 05:45
search mode all of it is available to
05:43 - 05:48
you right now again I've personally
05:45 - 05:50
messed around with it over on X but if I
05:48 - 05:52
do something like deep search and I say
05:50 - 05:54
what YouTube channels should I follow to
05:52 - 05:56
stay up to date on AI click on deep
05:54 - 05:59
search we can actually see as it thinks
05:56 - 06:01
through and does this search for us and
05:59 - 06:04
after a little bit less than a minute of
06:01 - 06:06
thinking we can see follow AI explained
06:04 - 06:08
Matt wolf Wes Roth and David Shapiro for
06:06 - 06:10
the latest AI news and updates with a
06:08 - 06:11
little bit more details about them they
06:10 - 06:14
also talk about two-minute papers and
06:11 - 06:16
Lex fredman all great options if I put
06:14 - 06:18
it on think mode and ask it the most
06:16 - 06:20
cliche question ever how many RS are in
06:18 - 06:23
the word strawberry we can watch this
06:20 - 06:25
Chain of Thought thinking as it thinks
06:23 - 06:28
through in real time here and after
06:25 - 06:30
spending 22 seconds of thinking and
06:28 - 06:33
double-checking triple checking we
06:30 - 06:34
finally get the correct answer that the
06:33 - 06:36
number of ours in the word strawberry is
06:34 - 06:38
three I know Matthew Burman did a much
06:36 - 06:39
deeper dive testing a whole bunch of
06:38 - 06:42
different prompts in Gro make sure you
06:39 - 06:44
check out his video if you want a deeper
06:42 - 06:46
dive and see some more complex prompts
06:44 - 06:48
getting tested out but a quick spoiler
06:46 - 06:50
he found that it did really good with
06:48 - 06:52
pretty much everything he tried now
06:50 - 06:54
while grock was probably the biggest
06:52 - 06:57
news in the AI world this week there was
06:54 - 06:58
a ton of other announcements and so I'm
06:57 - 07:00
going to try to go through them really
06:58 - 07:02
quickly and break down what I think you
07:00 - 07:06
need to know this week perplexity op
07:02 - 07:08
sourced their R1 1776 model which is
07:06 - 07:10
basically deep seek R1 that they did
07:08 - 07:12
some additional post training on to
07:10 - 07:14
remove all of the bias and to add
07:12 - 07:16
accurate and factual information you can
07:14 - 07:18
use this model directly from their API
07:16 - 07:20
and they made the model weights
07:18 - 07:22
available over on hugging face and we
07:20 - 07:25
can see an example here where when
07:22 - 07:28
asking deep seek R1 what happened in tnm
07:25 - 07:30
and square in 1989 it refused to answer
07:28 - 07:33
the question but when using R1
07:30 - 07:35
1776 they actually gave the proper
07:33 - 07:38
answer and there's a whole bunch of
07:35 - 07:40
other examples of it being much less
07:38 - 07:42
censored in this new version as always I
07:40 - 07:45
will link up to this article if you want
07:42 - 07:47
to dive deeper into what R11 1776 can do
07:45 - 07:49
and since we're talking about making
07:47 - 07:52
large language models more uncensored
07:49 - 07:54
open AI is apparently trying to make
07:52 - 07:55
their models more uncensored according
07:54 - 07:58
to this Tech crunch article here open AI
07:55 - 08:00
is changing how it trains AI models to
07:58 - 08:02
explicitly Embrace int ual Freedom no
08:00 - 08:04
matter how challenging or controversial
08:02 - 08:06
a topic may be the changes might be a
08:04 - 08:07
part of open ai's efforts to land in the
08:06 - 08:09
good graces of the new Trump
08:07 - 08:11
Administration but it also seems to be a
08:09 - 08:14
part of a broader shift in Silicon
08:11 - 08:15
Valley and what's considered AI safety
08:14 - 08:17
the principle may be controversial as it
08:15 - 08:19
means the assistant May remain neutral
08:17 - 08:21
on topics that some consider morally
08:19 - 08:24
wrong or offensive however the goal of
08:21 - 08:26
an AI assistant is to assist Humanity
08:24 - 08:28
not to shape it so all of this basically
08:26 - 08:30
means that over time we're going to see
08:28 - 08:33
it answer more and more questions and
08:30 - 08:36
refuse to answer less questions although
08:33 - 08:38
I'm not totally certain on the timeline
08:36 - 08:40
of these changes here and while we're
08:38 - 08:42
talking about open AI let's shift over
08:40 - 08:45
to Microsoft who appears to be gearing
08:42 - 08:47
up for new models to come out from open
08:45 - 08:50
AI within the coming weeks according to
08:47 - 08:52
this article on the verge GPT 4.5 could
08:50 - 08:54
arrive as soon as next week Microsoft
08:52 - 08:57
Engineers are currently readying server
08:54 - 08:59
capacity for open ai's upcoming 4.5 and
08:57 - 09:01
GPT 5 models Accord according to a
08:59 - 09:03
source familiar with the company's plans
09:01 - 09:06
basically this article is speculating
09:03 - 09:08
that because Microsoft is gearing up for
09:06 - 09:09
this and engineers at Microsoft have
09:08 - 09:12
sort of confirmed they're gearing up for
09:09 - 09:14
this we can likely expect GPT 4.5 like
09:12 - 09:18
next week or the week after and we
09:14 - 09:21
should expect GPT 5 in late may now we
09:18 - 09:23
talked about this last week but GPT 4.5
09:21 - 09:25
is going to be their next non-thinking
09:23 - 09:27
model so it's not going to have that
09:25 - 09:29
whole Chain of Thought thing going but
09:27 - 09:31
then GPT 5 on it's going to of
09:29 - 09:33
consolidate all of the models and
09:31 - 09:34
they're not going to differentiate
09:33 - 09:36
between the models that sort of think
09:34 - 09:37
through things and don't they're just
09:36 - 09:39
going to let the models think through
09:37 - 09:41
things that need more thinking and when
09:39 - 09:42
prompts need less thinking it's going to
09:41 - 09:44
think through it less and since we're
09:42 - 09:45
already talking about Microsoft they've
09:44 - 09:49
shipped a handful of other things this
09:45 - 09:51
week including a new AI experience
09:49 - 09:53
inside of the Microsoft store so if I
09:51 - 09:55
open up my Microsoft store here on
09:53 - 09:57
Windows we can see over on the left
09:55 - 09:58
sidebar there's a new little like AI
09:57 - 10:01
icon and if we click on this it takes us
09:58 - 10:04
to the A AI Hub where we can see all
10:01 - 10:06
sorts of AI powered apps inside of
10:04 - 10:09
Microsoft store so things like a Reading
10:06 - 10:12
Coach Microsoft co-pilot cascader
10:09 - 10:15
clipchamp gamma AI which does slides
10:12 - 10:17
Adobe express canva it just sort of put
10:15 - 10:20
all of the AI related stuff that was in
10:17 - 10:22
the Microsoft store into like an AI Hub
10:20 - 10:24
section of the store today's video is
10:22 - 10:27
sponsored by in video a platform that
10:24 - 10:29
makes it super easy to generate videos
10:27 - 10:31
with a single prompt and I'm not talking
10:29 - 10:33
about like the 5 to 10 second videos
10:31 - 10:35
you're used to when generating AI videos
10:33 - 10:38
I'm talking anywhere from 30 seconds up
10:35 - 10:40
to like 10 minute videos all from one
10:38 - 10:42
prompt so check this out my favorite way
10:40 - 10:44
to use N Video is to use their workflows
10:42 - 10:46
feature where they have all sorts of
10:44 - 10:48
pre-built templates for things like
10:46 - 10:51
explainer videos lisal videos generative
10:48 - 10:53
ads and so much more but for this
10:51 - 10:55
example let's do a listicle video Let's
10:53 - 10:57
create a 3-minute video for YouTube
10:55 - 11:00
about the top three and let's give it
10:57 - 11:02
the prompt the top three foods and
11:00 - 11:04
drinks to get better sleep at night and
11:02 - 11:06
under settings we have options for
11:04 - 11:08
generated Clips generated images or only
11:06 - 11:10
stock media let's use generated Clips so
11:08 - 11:12
it actually generates the whole video
11:10 - 11:14
with AI for us and then we have so many
11:12 - 11:16
other options like the ability to dial
11:14 - 11:19
in the background music the language
11:16 - 11:20
subtitles voice actors Watermark music
11:19 - 11:22
preferences and generative style I'm
11:20 - 11:24
going to leave all of these on the
11:22 - 11:25
default and see what we get with the top
11:24 - 11:28
three foods and drinks to get better
11:25 - 11:30
sleep and we got a nearly 2-minute video
11:28 - 11:32
that's actually one of the better
11:30 - 11:35
generative videos that I've seen so far
11:32 - 11:37
first up we've got almonds these bad
11:35 - 11:39
boys are loaded with magnesium which is
11:37 - 11:41
like the chill pill of minerals it helps
11:39 - 11:43
kick inflammation to the curb and tells
11:41 - 11:45
your cortisol levels to take a hike now
11:43 - 11:46
right out of the gate the video was
11:45 - 11:48
really good but let's say I want to make
11:46 - 11:50
some tweaks to it I can do it with
11:48 - 11:52
simple text commands like let's make the
11:50 - 11:54
background music a little more upbeat I
11:52 - 11:56
click generate and it will change that
11:54 - 11:58
in the whole video and now when we play
11:56 - 11:59
it back I've got a secret weapon for
11:58 - 12:02
better sleep and it's it's not what you
11:59 - 12:04
think we're talking food and drinks
12:02 - 12:06
people let's dive into the crema Creme
12:04 - 12:08
of sleep inducing eats and pretty much
12:06 - 12:09
everything about the video you just
12:08 - 12:11
created is editable just click on this
12:09 - 12:13
edit button you could regenerate any of
12:11 - 12:16
the generative videos you can upload
12:13 - 12:18
your own media use stock media tweak the
12:16 - 12:20
music tweak the script however you want
12:18 - 12:23
and so many other settings so if you
12:20 - 12:26
want the simplest easiest way to
12:23 - 12:28
generate a video using AI like a
12:26 - 12:31
completely ready to publish video
12:28 - 12:32
definitely check out Nido AI you can
12:31 - 12:33
learn more at the link in the
12:32 - 12:36
description and thank you so much to
12:33 - 12:38
Nido for sponsoring this video Microsoft
12:36 - 12:41
also released some new research this
12:38 - 12:43
week called biom mu1 which explores the
12:41 - 12:46
structural changes driving protein
12:43 - 12:48
functions now this actually Builds on
12:46 - 12:50
top of the alpha fold database which was
12:48 - 12:53
created by Google deepmind this new AI
12:50 - 12:55
model can predict viable protein
12:53 - 12:58
structures which could really help
12:55 - 13:00
increase research in medicine and
12:58 - 13:02
biology and help us discover new drugs
13:00 - 13:05
and things like that and along a similar
13:02 - 13:07
topic this week Google research
13:05 - 13:10
introduced what they call an AI co-
13:07 - 13:13
scientist which is a multi-agent AI
13:10 - 13:15
system that's a virtual scientific
13:13 - 13:18
collaborator to help scientists generate
13:15 - 13:20
novel hypotheses and research proposals
13:18 - 13:22
now the way I've heard this described is
13:20 - 13:25
in the same way that you might use
13:22 - 13:27
cursor to help you write code and cursor
13:25 - 13:30
becomes your like AI assistant this is
13:27 - 13:32
like your AI assistant for scientific
13:30 - 13:34
research and it's already proved to be
13:32 - 13:36
pretty valuable in fact I came across
13:34 - 13:38
this article from BBC about how AI
13:36 - 13:41
cracks superbug problem in two days that
13:38 - 13:43
took scientists years a complex problem
13:41 - 13:44
that took microbiologists a decade to
13:43 - 13:46
get to the bottom of has been solved in
13:44 - 13:49
just two days by a new artificial
13:46 - 13:51
intelligence tool the team had spent
13:49 - 13:54
years working out and proving why some
13:51 - 13:57
super bugs are immune to antibiotics he
13:54 - 13:57
gave Google's
13:58 - 14:03
co-citizens investigating and it reached
14:00 - 14:05
the same conclusion in just 48 Hours it
14:03 - 14:06
was so accurate that the person wrote an
14:05 - 14:08
email to Google saying do you have
14:06 - 14:11
access to my computer and of course
14:08 - 14:13
Google confirmed they did not but it
14:11 - 14:16
managed to crack the code on a problem
14:13 - 14:19
in 48 hours that previously they took 2
14:16 - 14:21
years to crack so pretty impressive
14:19 - 14:22
breakthroughs happening right now and
14:21 - 14:26
while we're on the topic of Google
14:22 - 14:28
Google introduced their P Gemma 2 mix
14:26 - 14:31
which is a vision language model this
14:28 - 14:34
new pal Gemma 2 mix model can do things
14:31 - 14:36
like solve tasks such as long and short
14:34 - 14:38
captioning optical character recognition
14:36 - 14:41
image question answering object
14:38 - 14:42
detection and segmentation all from one
14:41 - 14:44
model we can see some examples here they
14:42 - 14:47
gave it this image and the input was
14:44 - 14:49
detect Android and it put a box around
14:47 - 14:51
the Android in this example they were
14:49 - 14:54
given this image detect chair and table
14:51 - 14:56
and you can see it put you know boxes
14:54 - 14:58
around two different tables and a chair
14:56 - 15:00
detect food plate Bowl you can see it
14:58 - 15:03
found found all of those things here it
15:00 - 15:06
is ocing the text on whatever the heck
15:03 - 15:08
this is here here it segmented out this
15:06 - 15:10
cat it can answer questions about images
15:08 - 15:13
and it's just a single model that can do
15:10 - 15:15
a whole bunch of things and one cool
15:13 - 15:17
thing about the Gemma series of models
15:15 - 15:18
from Google is these are their open-
15:17 - 15:21
Source models so you can actually
15:18 - 15:23
download the model from hugging face and
15:21 - 15:25
use it yourself and iterate off of it
15:23 - 15:27
and improve upon it Google released some
15:25 - 15:29
updates for advertisers where
15:27 - 15:32
advertisers can now generate lifestyle
15:29 - 15:34
imagery so basically you can use AI to
15:32 - 15:37
generate what looks like stock imagery
15:34 - 15:39
of real people for your advertising on
15:37 - 15:40
Google now we can see what this looks
15:39 - 15:43
like describe the kind of image you want
15:40 - 15:45
person cooking they click generate and
15:43 - 15:47
it gives them a whole bunch of images
15:45 - 15:49
that look like stock photos of a person
15:47 - 15:52
cooking there's also some new updates
15:49 - 15:54
out of Google around the iOS apps you
15:52 - 15:56
can now use lens to search your screen
15:54 - 15:58
while browsing on iOS if you have an
15:56 - 16:00
iPhone you'll find a new lens option
15:58 - 16:03
that lets you select and search what's
16:00 - 16:04
on your screen with chrome or the Google
16:03 - 16:06
app using whatever gestures come
16:04 - 16:08
naturally like drawing highlighting or
16:06 - 16:10
tapping so basically it's like that
16:08 - 16:12
Circle to search feature that's been in
16:10 - 16:14
all the Google phones now it's in
16:12 - 16:17
iPhones as well Google also rolled out
16:14 - 16:19
deep research inside of the Gemini app
16:17 - 16:21
on IOS and Android so if you're a Gemini
16:19 - 16:24
Advanced user you can now access it
16:21 - 16:27
across Android and iOS to generate your
16:24 - 16:30
deep research reports mrol released a
16:27 - 16:33
new Regional model this week which is
16:30 - 16:35
focused on the Arabic language the new
16:33 - 16:37
model is called mistal saba and it is
16:35 - 16:39
specifically designed for arabic
16:37 - 16:41
speaking countries if you've been
16:39 - 16:42
following along to all the open AI drama
16:41 - 16:44
that's happened over the last couple
16:42 - 16:47
years you'll probably remember Mira
16:44 - 16:49
moradi who was previously the CTO of
16:47 - 16:52
open AI she was around she took over for
16:49 - 16:54
like a day when Sam Alman got fired and
16:52 - 16:57
then came back well recently she kind of
16:54 - 16:59
unexpectedly up and left open AI now
16:57 - 17:01
she's resurfaced out of of stealth mode
16:59 - 17:03
to tell us about her new company called
17:01 - 17:05
thinking machine lab she apparently
17:03 - 17:07
started this company with a handful of
17:05 - 17:10
other people that came over from open Ai
17:07 - 17:12
and the goal is helping people adapt AI
17:10 - 17:14
systems to work for their specific needs
17:12 - 17:16
developing strong foundations to build
17:14 - 17:18
more capable AI systems and fostering a
17:16 - 17:19
culture of open science that helps the
17:18 - 17:22
whole field understand and improve these
17:19 - 17:24
systems the goal is simple Advance AI by
17:22 - 17:25
making it broadly useful and
17:24 - 17:27
understandable through a solid
17:25 - 17:30
foundations open science and practical
17:27 - 17:32
applications looking at their website
17:30 - 17:34
there's a little bit more detail here it
17:32 - 17:35
sounds like they're planning on open
17:34 - 17:38
sourcing what they create they're
17:35 - 17:40
emphasizing human AI collaboration
17:38 - 17:42
instead of focusing solely on making
17:40 - 17:44
fully autonomous AI systems so we don't
17:42 - 17:45
know exactly what they're building yet
17:44 - 17:48
but it kind of sounds like they're going
17:45 - 17:50
to be making their own Foundation models
17:48 - 17:52
that are likely going to be open source
17:50 - 17:55
and likely not like AI agents that just
17:52 - 17:56
go do everything for you more like AI
17:55 - 17:58
that assists you in whatever you're
17:56 - 18:00
trying to achieve I also came across
17:58 - 18:02
this article this this week from scitec
18:00 - 18:04
daily scientists unveil AI that learns
18:02 - 18:07
without human labels this new AI
18:04 - 18:09
algorithm called torque clustering
18:07 - 18:11
enhances an AI system's ability to learn
18:09 - 18:13
and identify patterns and data on its
18:11 - 18:17
own without human input kind of the Holy
18:13 - 18:19
Grail in AI right now is self-improving
18:17 - 18:21
AI this is something we haven't really
18:19 - 18:23
seen yet where you have an AI system
18:21 - 18:27
that just kind of goes off and learns on
18:23 - 18:29
its own right now there's a lot of data
18:27 - 18:31
labeling that has to happen so the AI
18:29 - 18:33
essentially knows if it's getting things
18:31 - 18:35
right or not and this new torque
18:33 - 18:38
clustering apparently enhances ai's
18:35 - 18:40
ability to learn and identify patterns
18:38 - 18:42
in data independently without any human
18:40 - 18:43
intervention by uncovering hidden
18:42 - 18:45
patterns it can provide valuable
18:43 - 18:48
insights such as detecting disease
18:45 - 18:50
Trends identifying fraudulent activities
18:48 - 18:52
and understanding human behavior and
18:50 - 18:54
this article goes on to basically say
18:52 - 18:55
that torque clustering could lead to AGI
18:54 - 18:57
torque clustering could support the
18:55 - 19:00
development of General artificial
18:57 - 19:02
intelligence AKA a GI particularly in
19:00 - 19:04
robotics and autonomous systems by
19:02 - 19:06
helping to optimize movement control and
19:04 - 19:08
decision-making it is set to redefine
19:06 - 19:11
the landscape of unsupervised learning
19:08 - 19:12
Paving the way for truly autonomous Ai
19:11 - 19:15
and the open- source code has been made
19:12 - 19:18
available to researchers again
19:15 - 19:20
self-improving AI that learns is kind of
19:18 - 19:23
what most of these AI companies are
19:20 - 19:25
shooting for but it's also the sort of
19:23 - 19:27
scariest possible scenario when AI
19:25 - 19:29
starts learning and teaching itself
19:27 - 19:32
things that's where most of of those
19:29 - 19:34
Sci-Fi movies that show AI destroying
19:32 - 19:37
the world seem to start it's like as
19:34 - 19:39
soon as we get to that point where AI is
19:37 - 19:41
now going and teaching itself stuff and
19:39 - 19:43
learning and trying to understand stuff
19:41 - 19:45
and constantly self-improving itself
19:43 - 19:47
that's kind of the point where it seems
19:45 - 19:49
like humans might lose control a little
19:47 - 19:52
bit so this is sort of like equally
19:49 - 19:54
exciting but also slightly scary that
19:52 - 19:56
we're just on the brink of that
19:54 - 19:58
self-improving AI but let's move on to
19:56 - 20:01
some lighter subjects like the fact that
19:58 - 20:03
m moft and Xbox created what they call
20:01 - 20:06
Muse which is a generative AI model for
20:03 - 20:09
gameplay this AI model was trained on
20:06 - 20:11
multiplayer battle arena game bleeding
20:09 - 20:13
edge this allows the model to create
20:11 - 20:15
consistent and diverse gameplay rendered
20:13 - 20:17
by AI demonstrating a major step towards
20:15 - 20:19
generative AI models that can Empower
20:17 - 20:22
game creators we can see some examples
20:19 - 20:25
on the screen here of actual gameplay
20:22 - 20:26
footage generated by the AI they say we
20:25 - 20:29
are already using Muse to develop a
20:26 - 20:30
real-time playable AI model train on
20:29 - 20:32
other first party games and we see
20:30 - 20:34
potential for this work to one day
20:32 - 20:36
benefit both players and game creators
20:34 - 20:38
from allowing us to revive nostalgic
20:36 - 20:40
games to faster creative ideation
20:38 - 20:42
they're going to make it available for
20:40 - 20:44
us to mess around with in co-pilot Labs
20:42 - 20:47
soon but who knows how soon moving on to
20:44 - 20:50
the world of AI video Pika Labs who just
20:47 - 20:51
keeps on shipping out new features just
20:50 - 20:53
released another one this week called
20:51 - 20:56
Pika swap we can actually see in this
20:53 - 20:58
demo video here that you give it a video
20:56 - 21:00
you give it an image and it will swap
20:58 - 21:03
out things in the video with the image
21:00 - 21:05
that you gave it it's pretty crazy
21:03 - 21:08
looking in fact I haven't tested it yet
21:05 - 21:10
but let's head on over to p. art we can
21:08 - 21:12
see here we have the option to upload a
21:10 - 21:15
video and upload an image and it's kind
21:12 - 21:17
of got a prompt pre-built here swap the
21:15 - 21:18
original object with a completely new
21:17 - 21:21
object the new object should vary in
21:18 - 21:22
multiple attributes such color style
21:21 - 21:25
visual appearance and much more be
21:22 - 21:27
creative I'll give it this AI generated
21:25 - 21:28
video of a Ferrari driving on the moon
21:27 - 21:30
it looks like I can describe I want to
21:28 - 21:33
modify or use a brush let's go ahead and
21:30 - 21:36
use the brush here and then just like
21:33 - 21:39
modify or brush out the car in the video
21:36 - 21:42
here and then for an image I've got this
21:39 - 21:43
image of our side by side dune buggy
21:42 - 21:45
thing that we like to take out to the
21:43 - 21:47
desert I'll throw that in and see if uh
21:45 - 21:48
we can make that look like it's driving
21:47 - 21:50
on the moon let's submit the prompt and
21:48 - 21:51
see what we get out of it and here's
21:50 - 21:53
what we got out of it I mean it made the
21:51 - 21:56
dude buggy look really dark and gave it
21:53 - 21:58
some extra Wheels I don't cherry-pick
21:56 - 22:00
things on this channel this is the
21:58 - 22:02
actual first output that I got from it I
22:00 - 22:05
mean it did replace the Corvette it just
22:02 - 22:06
doesn't it it it made the buggy look
22:05 - 22:08
really really dark I'll have to play
22:06 - 22:11
around with this some more and really
22:08 - 22:13
figure out what kind of images and
22:11 - 22:16
prompting combinations work best again
22:13 - 22:18
this was my first shot at it Pika also
22:16 - 22:21
released an iPhone app so if you have an
22:18 - 22:22
iPhone you can now generate all of these
22:21 - 22:25
fun things that you can do inside of
22:22 - 22:27
Pika right directly on your IOS app some
22:25 - 22:29
cool new research came out this week
22:27 - 22:31
from Alibaba group group called animate
22:29 - 22:34
anyone to and we can see in some of
22:31 - 22:37
these demos it takes a sort of driving
22:34 - 22:39
video here a reference image and then it
22:37 - 22:40
puts the person from the reference image
22:39 - 22:43
into the driving video so we can see
22:40 - 22:45
this video from I believe Mr Bean here
22:43 - 22:47
writing his bicycle and then they
22:45 - 22:49
swapped it out with this person here we
22:47 - 22:51
can see this person skateboarding they
22:49 - 22:53
uploaded this image and it made this
22:51 - 22:55
person in the image doing the
22:53 - 22:57
skateboarding here's one of this martial
22:55 - 23:00
artist replaced by The Joker here's one
22:57 - 23:03
where Aldo is doing parkour from this
23:00 - 23:04
original parkour video pretty cool stuff
23:03 - 23:06
kind of similar to what pic was trying
23:04 - 23:08
to do but this one looks like it's more
23:06 - 23:10
focused on people specifically and this
23:08 - 23:12
one's just research right now not
23:10 - 23:15
something that we actually have access
23:12 - 23:18
to Spotify is now working with 11 Labs
23:15 - 23:20
so that authors can use 11 labs to
23:18 - 23:22
generate the speaking of audiobooks and
23:20 - 23:25
Spotify is actually going to allow those
23:22 - 23:28
11 Labs generated audio books on their
23:25 - 23:30
platform Nvidia launched a new platform
23:28 - 23:33
that helps people learn sign language
23:30 - 23:34
apparently American sign language is the
23:33 - 23:36
third most prevalent language in the
23:34 - 23:40
United States and you can actually use
23:36 - 23:43
this app right now by going to ss- a.com
23:40 - 23:45
and you can either learn ASL or you can
23:43 - 23:48
record yourself signning to kind of
23:45 - 23:50
teach the model welcome to signs make
23:48 - 23:52
sure you have the space to do the signs
23:50 - 23:54
adjust your camera and move either
23:52 - 23:57
closer or further away so that your face
23:54 - 23:59
fits within the box if you are
23:57 - 24:02
right-handed put your right hand into
23:59 - 24:04
the blue circle if you are left-handed
24:02 - 24:06
put great we will make sure we use your
24:04 - 24:07
right hand as your main hand for signing
24:06 - 24:09
the first word I will teach you is the
24:07 - 24:12
word hello let me show you how to do it
24:09 - 24:14
stretch your fingers and tuck your thumb
24:12 - 24:17
in now move your right hand so it
24:14 - 24:19
touches your head and now move your hand
24:17 - 24:24
outwards and that's it now it's your
24:19 - 24:26
turn to try let's go great now move your
24:24 - 24:28
awesome you get the idea it's like
24:26 - 24:30
almost like a dualingo for sign language
24:28 - 24:32
where you can go and teach it science if
24:30 - 24:34
you know sign language or you can go and
24:32 - 24:36
learn American Sign Language if you
24:34 - 24:38
don't and it watches your camera and
24:36 - 24:41
make sure that you're doing the signs
24:38 - 24:42
correctly so pretty handy feature if ASL
24:41 - 24:44
is something you want to learn all right
24:42 - 24:48
moving into Hardware now this week Apple
24:44 - 24:51
debuted their new iPhone 16e this new
24:48 - 24:53
iPhone is a less expensive iPhone that
24:51 - 24:55
has Apple intelligence built into it up
24:53 - 24:58
until now I believe you needed like an
24:55 - 25:01
iPhone 16 Pro or better to actually get
24:58 - 25:03
apple intelligence but now this 16e also
25:01 - 25:05
has it and it's a stripped down model
25:03 - 25:08
that it's got not as strong of a camera
25:05 - 25:09
not as strong of specs is like the Pro
25:08 - 25:11
Models and honestly the big
25:09 - 25:13
differentiator of this model just seems
25:11 - 25:16
to be the price it's like a lower
25:13 - 25:18
barrier to entry way to get an iPhone
25:16 - 25:20
the iPhone 16e will be available in
25:18 - 25:23
white and black these various storage
25:20 - 25:24
ranges and they're starting at $599 so
25:23 - 25:27
if you want an iPhone and you want to
25:24 - 25:29
get the least expensive iPhone you can
25:27 - 25:31
get without buying an old older used
25:29 - 25:33
model the iPhone 16e might be for you
25:31 - 25:35
and if you're really trying to get the
25:33 - 25:37
Apple intelligence this is the least
25:35 - 25:39
expensive way to get a device with apple
25:37 - 25:42
Intelligence on it the company who main
25:39 - 25:45
who made this AI pin that Marquez and a
25:42 - 25:46
lot of reviewers just kind of destroyed
25:45 - 25:49
when it came out well they're going
25:46 - 25:52
under HP is basically acquiring their
25:49 - 25:55
assets for 116 million and essentially
25:52 - 25:56
if you have a Humane AI pin they're
25:55 - 25:58
going to like brick it it's just not
25:56 - 25:59
going to work anymore they're phasing
25:58 - 26:02
out one of the lamest things about it
25:59 - 26:04
too is they're only refunding people
26:02 - 26:05
that are still within the refund period
26:04 - 26:07
so if you were like an early adopter of
26:05 - 26:09
it you were one of the first people to
26:07 - 26:10
go and support Humane cuz you like the
26:09 - 26:12
idea you're kind of screwed you're not
26:10 - 26:14
getting any money back on it you're just
26:12 - 26:16
going to have a Humane pin that doesn't
26:14 - 26:19
work anymore which kind of sucks but
26:16 - 26:20
another company that kind of got a bad
26:19 - 26:23
review from Marquez and a lot of the
26:20 - 26:25
other reviewers rabbit is continually
26:23 - 26:27
trying to improve their product and make
26:25 - 26:29
it right however this latest
26:27 - 26:31
announcement actually doesn't really
26:29 - 26:34
have anything to do with their little
26:31 - 26:36
handheld orange device it's basically
26:34 - 26:38
what the rabbit can do but on Android we
26:36 - 26:39
can see that they're finally showing off
26:38 - 26:41
their large action model that will go
26:39 - 26:43
and take actions on your behalf but
26:41 - 26:44
they're actually showing it off on
26:43 - 26:47
Android they're not actually showing it
26:44 - 26:49
off being used on the device I think the
26:47 - 26:52
idea here is that you build out the
26:49 - 26:54
automations in Android and once these
26:52 - 26:57
automations are built out then you can
26:54 - 26:59
use the little orange handheld rabbit R1
26:57 - 27:01
device to to then trigger the
26:59 - 27:03
automations even if you're not at a
27:01 - 27:05
computer or not on an Android device but
27:03 - 27:07
also at the same time if you do have an
27:05 - 27:10
Android device I don't actually know
27:07 - 27:12
what you need the rabbit for because it
27:10 - 27:14
seems like you can launch the large
27:12 - 27:16
action models just using an Android and
27:14 - 27:18
not necessarily even need the little
27:16 - 27:20
orange device I don't know I'm a little
27:18 - 27:22
confused I do have a rabbit I've still
27:20 - 27:24
never played with it uh it came out
27:22 - 27:25
almost a year ago exactly now so maybe
27:24 - 27:28
it's about time to finally pull it out
27:25 - 27:30
of the box and give it a review like a
27:28 - 27:33
year after I got it all right let's talk
27:30 - 27:35
some cool robots now Adam Silver who is
27:33 - 27:37
the commissioner of the NBA showed off
27:35 - 27:39
some ways they're using robots to help
27:37 - 27:40
NBA players and the Golden State
27:39 - 27:43
Warriors here's another example of how
27:40 - 27:45
the Warriors are using physical AI to
27:43 - 27:47
support their players yeah at first it
27:45 - 27:49
was definitely weird having all these
27:47 - 27:51
robots moving around and doing stuff
27:49 - 27:53
these robots can replicate any of them I
27:51 - 27:58
just treat them like any other player
27:53 - 27:58
get the out of here give me another one
28:16 - 28:20
we also learned this week that meta is
28:17 - 28:22
planning on getting into AI powered
28:20 - 28:24
humanoid robots meta plans to work on
28:22 - 28:26
its own humanoid robot Hardware with an
28:24 - 28:27
initial focus on household chores
28:26 - 28:30
they've started discussing its plan with
28:27 - 28:33
robotic companies like unry Robotics and
28:30 - 28:34
figure Ai and at least initially isn't
28:33 - 28:37
planning on building its own meta
28:34 - 28:39
branded robot so we'll just have to keep
28:37 - 28:41
an eye on this and watch how it unfolds
28:39 - 28:44
but since we just mentioned figure let's
28:41 - 28:46
check out the demo that figure just
28:44 - 28:48
released this week as well around their
28:46 - 28:50
humanoid robots these are the Helix
28:48 - 28:52
robots and what's really cool about
28:50 - 28:54
these is that they work completely
28:52 - 28:56
autonomously and they actually work
28:54 - 28:58
together like you don't actually hear
28:56 - 28:59
them communicate but they're somehow
28:58 - 29:02
communicating with each other even
28:59 - 29:04
though this is the very first time that
29:02 - 29:08
you've ever seen these
29:04 - 29:10
items I'd like you to use your new Helix
29:08 - 29:13
Ai and try to reason through where you
29:10 - 29:16
think they belong in the scene and then
29:13 - 29:16
work together to put them away
29:34 - 29:37
okay before I continue this video
29:35 - 29:39
there's one comment on YouTube that just
29:37 - 29:41
has me cracking up as I watch this video
29:39 - 29:43
and I want you to experience the same
29:41 - 29:45
joy that I had watching this video the
29:43 - 29:47
first time the very first comment here
29:45 - 29:48
is when you ask your two Stone roommates
29:47 - 29:50
to put away the groceries all right keep
29:48 - 29:53
that context in mind let's get back to
31:17 - 31:21
so it's really cool because it never
31:19 - 31:23
seen those objects before it figured out
31:21 - 31:26
what the objects were where they should
31:23 - 31:27
go and then how to work together to get
31:26 - 31:30
that while we're on the topic of robots
31:27 - 31:32
I'll go ahead and wrap up with this one
31:30 - 31:34
this has got to be one of the creepiest
31:32 - 31:38
robots I've seen out of Clone this is a
31:34 - 31:40
protoc clone the world's first bipedal
31:38 - 31:42
musculoskeletal Android I'm not going to
31:40 - 31:44
play the music on this video because I
31:42 - 31:47
don't actually know the copyright status
31:44 - 31:49
of it but it's like really dark creepy
31:47 - 31:52
music as you're seeing this like
31:49 - 31:54
humanoid robot with human looking
31:52 - 31:58
muscles hanging from a ceiling it just
31:54 - 31:59
makes it way creepier I mean it it's
31:58 - 32:01
very cool that we're building these
31:59 - 32:03
robots that actually move and have
32:01 - 32:06
muscles and skeletons like real humans
32:03 - 32:08
they're just making it very creepy with
32:06 - 32:10
the music they're using in that video
32:08 - 32:11
I'll link to it in the description if
32:10 - 32:13
you want to watch the original and
32:11 - 32:15
that's what I got for you I do have one
32:13 - 32:16
little housekeeping thing I mentioned
32:15 - 32:19
last week that I'm going to be giving
32:16 - 32:21
away an RTX 90 for free all you got to
32:19 - 32:23
do to win the 90 is be subscribed to
32:21 - 32:25
this channel subscribe to the Future
32:23 - 32:28
tools newsletter and I mentioned there
32:25 - 32:30
will be a third criteria well that third
32:28 - 32:33
criteria is you got to register for
32:30 - 32:34
nvidia's GTC conference the virtual
32:33 - 32:36
version of the conference is totally
32:34 - 32:38
free to register for you just go to
32:38 - 32:43
GTC you can register for free to watch
32:41 - 32:46
the online sessions once you've
32:43 - 32:48
registered for NVIDIA GTC I have a
32:46 - 32:50
little form that I created a Google form
32:48 - 32:52
where you just give me your email first
32:50 - 32:53
and last name country what session
32:52 - 32:56
you're planning on attending and then
32:53 - 32:58
upload a screenshot to confirm that you
32:56 - 33:01
did register for GTC and you'll be
32:58 - 33:04
entered to win an RTX 5090 there's no
33:01 - 33:05
purchase necessary all you got to do is
33:04 - 33:07
make sure you're subscribed to the
33:05 - 33:09
channel the newsletter and that you
33:07 - 33:13
register for the GTC event for free and
33:09 - 33:15
you'll be entered to win a $2,000 RTX
33:13 - 33:18
590 which are practically impossible to
33:15 - 33:20
even come by right now but Nvidia is
33:18 - 33:22
hooking up one person from this channel
33:20 - 33:23
who registers so make sure you get
33:22 - 33:26
registered and that's what I got for you
33:23 - 33:27
today thank you so much for tuning in if
33:26 - 33:30
you want to stay looped in on the latest
33:27 - 33:32
news you want to learn how to do cool
33:30 - 33:34
tutorials and workflows with AI you want
33:32 - 33:36
to know about all the latest cool AI
33:34 - 33:38
tools make sure you like this video And
33:36 - 33:39
subscribe to this channel I'll make sure
33:38 - 33:41
stuff like this keeps on showing up in
33:39 - 33:44
your YouTube feed and as always be sure
33:41 - 33:46
to check out futur tools. where I keep
33:44 - 33:48
the AI news page up to date on a daily
33:46 - 33:50
basis I share all of the cool AI tools
33:48 - 33:52
that I come across every single day and
33:50 - 33:54
of course this is where you register for
33:52 - 33:56
the free newsletter where every week
33:54 - 33:58
twice a week I'll send you an email with
33:56 - 34:00
just the coolest to tools and most
33:58 - 34:02
important news for you you also get free
34:00 - 34:04
access to the AI income database a
34:02 - 34:06
database of cool ways to make money with
34:04 - 34:09
AI all you got to do is sign up for free
34:06 - 34:10
over at Future tools. once again I
34:09 - 34:12
really really appreciate you hanging out
34:10 - 34:16
with me nerding out around the latest in
34:12 - 34:18
Ai and uh Cool Tech and yeah just really
34:16 - 34:20
appreciate you spending time with me
34:18 - 34:23
today and hopefully I'll see you in the
34:20 - 34:23
next one bye-bye
video to article