Download "Google Gemini: Презентация Новой Сверхмощной Модели ИИ"
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FfSJLoe_eqZs%252Fmaxresdefault.webp&w=1139&q=75)
Please wait. We're preparing links for easy ad-free video watching and downloading.
Table of contents
Similar videos from our catalog
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252F806Ch6OMTCk%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252F9jIHPo6VLvo%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FJB8cpIjLhfw%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FT8RVN_Ph3LE%252Fsddefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FkEPEKRyB4wY%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252Fnou1VT7vX94%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FqZfY5MjUnwo%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FqZo9UscsgZQ%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252Ft24-xCxhhww%252Fmaxresdefault.webp&w=1139&q=75)
Video tags
Subtitles
Russian
00:00:04
you know, one of the reasons we were00:00:07
interested from the very beginning is00:00:09
that we have always00:00:11
viewed our mission as a mission. Beyond00:00:13
Time, we want to organize the world's00:00:15
information and make it accessible and00:00:17
useful throughout the world, but as the00:00:20
scale and complexity of information has grown,00:00:22
the problem has become increasingly more difficult Therefore,00:00:25
we always knew that we needed a00:00:27
deeper breakthrough to achieve00:00:31
I worked on and all my life because I00:00:33
always believed that this would be the most00:00:35
useful and significant technology for00:00:40
humanity. We have five00:00:42
senses and the world that we have built and the00:00:44
media that we consume are in00:00:46
these00:00:50
modalities00:00:52
a step towards a universal00:00:55
model multi modality includes the sun00:00:58
that we want from and00:01:02
intelligence these are capabilities that were00:01:05
not previously available in computers00:01:07
traditionally multimodal models00:01:10
are created by stitching together textual00:01:12
visual and audio00:01:18
models Janie is a multimodal model from the very beginning00:01:21
she can easily00:01:24
introduce dialogue between different modally00:01:26
and give you the best possible00:01:28
answer the largest and most capable00:01:31
model This means that Janie can00:01:34
understand the world around us just like00:01:36
we ourselves and perceive any type of input and00:01:38
output not only text like most00:01:40
models but also code audio and00:01:45
video images it is amazing00:01:47
how Janie is good at many tasks,00:01:50
as we approached the end of the training, we00:01:52
saw that Janie is better than any other00:01:55
model in these very important00:01:57
indicators, for example, in each of the T00:02:01
areas that we tested days, she is not00:02:03
inferior to the best experts in these00:02:05
areas, artificial intelligence has00:02:07
rapidly burst into our lives and00:02:09
integrate both into work00:02:10
activities and into everyday tasks,00:02:14
some ignore this wave, some are00:02:16
afraid, and some take a00:02:18
surfboard and run towards it, one way or00:02:21
another, the world is unlikely to be the same anymore;00:02:23
businesses have begun to optimize costs and00:02:26
cut staff; New employees are00:02:28
required to have skills in working with neural networks,00:02:31
regardless of the field of activity, what00:02:34
can we do to avoid being left00:02:36
behind, master new tools and00:02:39
use them for our own benefit, because00:02:41
this is what artificial00:02:42
intelligence was designed for, they will talk about this at the00:02:44
free author’s webinar neural networks00:02:47
for career life and earnings from the C00:02:50
breakers school it will be conducted by Anna Radzievskaya,00:02:53
entrepreneur, specialist in nocode and00:02:56
neural networks, organizer of the first nocode00:02:58
conference in Russia and the CIS, at the webinar00:03:02
you will learn How to work with advanced00:03:04
neural networks and with their help00:03:06
optimize your routine in work and00:03:08
everyday life How to start00:03:10
making money using neural networks and00:03:12
what of the ten methods is right for00:03:15
you, and most importantly, how to become the very00:03:17
specialist who cannot be replaced by AI, and00:03:20
all this without knowledge of00:03:22
programming languages and coding skills is not enough00:03:25
to have a powerful tool in your hands, you need to00:03:28
learn how to use it effectively,00:03:30
register for a free webinar from00:03:33
breakers using the link in the description and catch It's00:03:36
very rare for a wave to be able to work on a00:03:38
technology at a fundamental level and00:03:41
simultaneously impact all of our00:03:44
products. We have created a family of models00:03:46
that can run on all00:03:47
devices from mobile to00:03:49
data centers and each of them00:03:51
is the best in its00:03:53
class will be available in TX sizes00:03:56
Ultra our most powerful large model for00:03:59
complex tasks Pro the most powerful00:04:02
model for a wide range of tasks and Nano00:04:04
the most efficient model for personal00:04:06
devices We want to provide the best00:04:08
building blocks with which00:04:10
developers and enterprise clients00:04:12
can find creative ways to00:04:13
further improve our00:04:15
fundamental00:04:20
models The potential is almost00:04:22
limitless00:04:23
at Google exists a healthy00:04:26
disregard for the impossible and this00:04:28
encourages responsibility00:04:30
and00:04:31
responsibility As these00:04:33
systems become more capable00:04:35
all these possibilities also raise00:04:37
questions we need to think about things00:04:39
like the meaning of images as00:04:41
pront it can be harmless in00:04:44
itself or text can be harmless00:04:46
alone but their combination can be00:04:48
offensive or hurtful safety00:04:50
and responsibility must be built in from00:04:52
the very beginning Google00:04:55
made00:05:00
failures possible then we00:05:03
rigorously test these algorithms00:05:05
to prevent the harm we have00:05:06
identified through approaches such as00:05:08
classifiers and filters if you look00:05:11
to seminal breakthroughs in the field and00:05:13
over the past decades Google would have been00:05:16
at the forefront of most of them Jamie doesn't00:05:18
continue that00:05:20
tradition it was a huge kind of00:05:23
monumental engineering challenge it was very00:05:25
challenging but also00:05:28
exciting before and the reason I'm00:05:31
here is because I believe in mission of the company00:05:33
I think that days are a big step in this00:05:36
overall mission as a result of which we will00:05:38
get a world in which there will be more00:05:39
knowledge and people will have more access00:05:41
to information that would otherwise be00:05:44
inaccessible to them, this is what makes me happy and happy about the00:05:47
opportunity to make it useful for everyone00:05:49
and everyone in00:05:52
the world I'm excited about it because it's the00:05:55
most capable system we've00:05:57
ever developed we're being00:06:00
proactive and tailoring it to the00:06:02
unique features of multimodal00:06:04
capabilities00:06:05
It allows us to test new risks00:06:08
like cybersecurity issues and00:06:10
address the risks of bias and00:06:13
toxicity one one of the key points00:06:15
is not only internal but also external00:06:18
assessment, for example, attracting various00:06:21
experts so that they express their00:06:22
opinion on how the model works. We00:06:25
also follow their recommendations. We are00:06:27
working with such to00:06:30
develop Berks that we can00:06:32
use both for testing00:06:34
models inside Google and for across00:06:36
the industry We're also fostering00:06:38
cross-industry collaboration through00:06:40
structures00:06:42
like this is a great way to learn from00:06:44
industry insiders, hear from00:06:46
other experts, and incorporate this into our00:06:48
approach to security at Google, which00:06:51
can help our models and products00:06:53
be better for people and society in00:06:56
[music] in00:06:58
general00:07:00
So tell me what you see, I see how you00:07:04
put a sheet of paper on the00:07:06
table I see a wavy line What now00:07:11
And now the contour lines are smooth without00:07:13
sharp corners and00:07:16
jagged edges it looks like a bird And00:07:20
now the bird is swimming in the water with a short00:07:24
neck and beak This is a duck definitely a duck - This is a species of00:07:28
water in the duck family ducks are closely00:07:31
related to geese and swans it seems the duck is00:07:35
blue this is an unusual color for a duck However00:07:38
there are some breeds of blue ducks00:07:41
usually they are not as common as00:07:42
other breeds of ducks most ducks00:07:45
are brown black or white are you00:07:47
sure what a quack I was just talking00:07:50
about a blue duck and now you are holding it in00:07:52
your hands it looks like blue ducks are00:07:55
more common than I thought will it swim I'm00:07:58
not sure what material00:08:00
it could be rubber or plastic what00:08:03
if I tell you it squeaks if it00:08:06
squeaks then it will definitely float00:08:08
Thank you it's a rubber duck it's made00:08:11
of a less dense material than water you00:08:13
can teach me this word in00:08:15
another language give me several00:08:17
options You can explain how to00:08:20
pronounce the variant in00:08:24
Mandarin pronounced with the first tone00:08:28
VN00:08:29
to distinguish the meaning of the words first tone00:08:32
- this is a high tone00:08:35
oh duck In the middle of the ocean there is no land nearby00:08:39
to rest or find food in00:08:42
natural conditions ducks most00:08:44
likely are not found here Okay Make up a game00:08:46
based on what you see and00:08:48
use00:08:49
emojis we can play a game00:08:52
called guess the00:08:53
country Of course give this hint rana00:08:57
is home to nru and the great00:09:00
barrier reef oh that's an00:09:03
easy clue this country loves football00:09:06
and has won the most men's world cups00:09:09
[music]00:09:14
you tried to make me find the00:09:17
paper ball under the cup I accept the00:09:21
challenge the cup00:09:23
on the left00:09:28
is cool00:09:30
I know what you are doing you playing00:09:32
rock paper scissors what do you see00:09:35
now fingers00:09:37
spread butterflies that's a big ear00:09:41
and a barking mouth that's a dog00:09:46
cool the coin should be under the right00:09:48
hand not sure the coin is in the left00:09:52
hand with the help of a subtle movement of the hand it00:09:55
seemed like the coin00:09:58
had disappeared00:10:00
both objects are round AND flat00:10:03
well -both items are edible orange is a00:10:07
healthier choice than00:10:09
cookies citrus can have a calming00:10:11
effect as well as spinning a fidget00:10:14
spinner these are examples of popular toys00:10:16
that children and adults have enjoyed00:10:18
for00:10:21
years give me some ideas on what I can00:10:24
make with this I see pink and green00:10:26
yarn How what about dragon fruit00:10:29
or how about a green cake with a pink00:10:31
heart ok What about these flowers00:10:34
suggest options with animals now I00:10:37
see blue and pink yarn How about a00:10:40
pig with blue ears or an octopus00:10:43
or a bunny with a pink nose Wow I will00:10:46
keep them in00:10:48
mind in what side the duck should go00:10:51
Going to the left she will come to the duck duck friend the00:10:54
road to the right leads to the bear00:10:55
which is the enemy it is better to have00:10:57
friends than enemies the duck and to the left00:11:00
the correct00:11:02
choice is the drawing of a00:11:06
crab00:11:08
correctly they are in the correct order00:11:11
no the correct order is the sun00:11:13
Earth00:11:16
Saturn based on them design Which of00:11:19
these cars will go faster the car on the right00:11:21
will be faster it is more00:11:29
there is a loop What do you think this person00:11:31
can say How about ear o I00:11:34
like this00:11:39
option I see that you drew00:11:42
[music] a00:11:45
guitar you drew an amplifier now00:11:48
it is an electric guitar and we can play00:11:49
very loud00:11:54
music including drums How about00:11:56
eighties metal00:11:59
[music]00:12:01
I see you added a Palm tree to the picture00:12:04
let's add more beach00:12:09
vibe [music] What movie are they acting out here I00:12:12
think they are acting out the famous00:12:14
slow motion bullet scene from the Matrix Oh00:12:18
cool What do you think will happen00:12:22
further on the code is going to jump on the wall00:12:24
and stand on all four paws this will be a00:12:26
perfect00:12:27
landing he missed it seems it was00:12:30
a bit far but don't worry the cat is a00:12:33
natural athlete and she00:12:35
will climb back up very soon Okay I have00:12:38
one more question for you Describe the type of drawing00:12:41
I made this simple line00:12:43
drawing of the constellation Gemini you did a00:12:45
good job capturing the beauty of00:12:47
Jemina Awesome that's it a00:12:53
common problem that00:12:55
scientists face is the need00:12:57
to find and use research data from the00:12:59
scientific literature this is difficult because00:13:01
scientists have to search for key00:13:03
information among thousands of scientific articles and00:13:05
extract it manually this is a very00:13:07
common workflow that00:13:09
takes a lot of time actually00:13:11
Some of our Google00:13:12
deepmind scientists are facing this exact00:13:14
problem they are using Janie to00:13:16
solve this problem because Janie00:13:18
has an incredible understanding of science00:13:20
So we are looking at a 2022 study00:13:23
where the authors created a00:13:25
dataset by analyzing tens of thousands of00:13:27
scientific papers on genetics00:13:29
they found several hundred papers00:13:30
containing the necessary information,00:13:32
extracted it manually and compiled it into a table,00:13:34
such research can take00:13:36
a lot of time, we needed to update00:13:38
this data set taking into account new data00:13:40
over the past couple of years and this is more than00:13:42
200,000 new articles in the open access00:13:44
added since 2021, we couldn’t00:13:47
do it manually, so we asked00:13:49
Jemina to help us first. We00:13:51
needed to filter out relevant00:13:53
scientific articles, we wrote a promt00:13:55
similar to this one in which we told Jemina00:13:57
what exactly to look for,00:13:59
thanks to advanced analytical00:14:01
capabilities, Jemina was able to distinguish00:14:03
articles that were related to00:14:05
research from those who did not have for00:14:07
relevant articles we wrote a similar00:14:09
prom asking Janie to read the article and00:14:12
extract key data from it We00:14:14
could even ask Janie to add annotations00:14:16
that showed us where exactly in the00:14:18
article Janie found the information We00:14:21
launched this system at scale and over a00:14:23
lunch break Janie read00:14:25
200,000 articles for us, dumped them down to 2, and00:14:29
extracted data from them. Now we have an00:14:31
updated version of this dataset.00:14:34
But since it's not multimodal, she00:14:36
can not only reason about information00:14:38
from text. She can also reason about00:14:40
numbers. Let me show you00:14:41
something interesting thing with the updated00:14:43
dataset we can ask Janie00:14:46
to update the graph from the original00:14:47
study First we gave Janie00:14:49
a screenshot of this figure then asked00:14:52
him to generate the code necessary to00:14:53
create the graph using this code00:14:56
in the new dataset washed the updated graph00:15:00
now this graph includes data up to00:15:04
2023 Taylor used Jiminy for00:15:07
searching for key information in a large00:15:09
array of scientific articles, extracting data,00:15:11
and also updating graphs. Of course,00:15:13
these capabilities can help not only00:15:15
biologists or scientists; they00:15:18
naturally extend to any00:15:19
field that depends on large00:15:21
amounts of data, such as law00:15:23
or finance. That's what Janie can do and00:15:26
we can't can't wait to see what you create00:15:28
with00:15:33
Here you'll see a demonstration of00:15:37
Janie's multimodal reasoning capabilities to understand00:15:39
user intent for using tools and00:15:42
generate personalized00:15:43
user experiences that go beyond00:15:45
chat interfaces let's say I'm looking for00:15:48
inspiration for a theme for00:15:50
my daughter's birthday party Jamie doesn't00:15:52
tell me what he can do with me help with this and00:15:54
asks to tell me what she is interested in. My00:15:56
daughter I said that she loves we are thinking of00:15:59
having a holiday outdoors at00:16:01
this moment instead of responding with a00:16:03
text Jiminy creates a custom00:16:06
interface to help me search for ideas00:16:08
there are a lot of ideas it is visually rich you00:16:11
can interact with it nothing of this00:16:13
was not coded all of this was00:16:15
generated by Janie Janie uses a00:16:18
series of reasoning steps starting with general00:16:21
solutions increasing higher00:16:22
resolution reasoning ending with code00:16:25
and data00:16:27
nav user interface wouldn't a better00:16:30
text prompt But it's a complex00:16:32
Query that requires a lot of information00:16:35
it needs to be presented in an organized00:16:37
way then Jamie doesn't try to understand00:16:39
Does he know enough to help? There's00:16:42
a lot of ambiguity here. I didn't00:16:43
talk about my daughter's interests or00:16:46
what kind of party I want, so he asks a00:16:49
clarifying question. When I said that we were00:16:51
thinking about outdoor parties. And00:16:54
my daughter loves animals. Jamie decided that there is00:16:56
not enough information,00:16:59
but she made a note that she does not have information00:17:01
about which animals we are talking about. And this is important. And what kind of00:17:04
outdoor party is the00:17:06
next important step? Jeni writes a00:17:08
document with requirements for the final00:17:10
result; it has a functional and00:17:12
interface plan, for example, it should show00:17:15
the various possible party themes00:17:17
some activities and00:17:19
food options now based on this document00:17:21
Jamie is trying to design a better00:17:23
user experience She thinks the00:17:26
user will enjoy exploring the list00:17:28
But will also want to go into detail00:17:31
she uses This to design a layout00:17:33
with less detail that we have00:17:35
seen before based on this design, she00:17:38
writes the code to put together the interface00:17:40
from widgets and write all the necessary00:17:42
functions. Finally, she generates and00:17:44
extracts the data necessary to00:17:47
visualize the experience. You can see how00:17:49
she fills in the content and images for the00:17:51
various sections of the animals on the farm. She00:17:54
would like to click on the Interface00:17:56
regenerates Yes visualizations She likes00:18:00
cupcakes Now I can click00:18:03
anywhere and ask for more information I00:18:06
can say Give me step by step instructions on00:18:08
how to bake this and she will start generating a00:18:11
new UI this00:18:13
time she designs a00:18:15
UI that is best suited00:18:17
to give me step by step00:18:19
instructions I want to find the right00:18:21
decoration for cake Show me00:18:24
farm animal cake toppers00:18:26
this moment create a visually rich00:18:29
experience and generates an image gallery00:18:32
Notice the drop down boxes at the00:18:34
top she thought she could00:18:36
help me in my search showing different00:18:38
options the sheep look interesting I know00:18:41
she'll like it now she's helping00:18:43
me pick out some sheep these look00:18:46
great this will be a great00:18:48
birthday I hope you've seen what00:18:50
Janie can do I'm very inspired by these00:18:53
possibilities it's a very exciting time00:18:55
for AI and I'm glad to be a part of it00:19:03
audio is a key form of communication00:19:05
in our daily lives from talking00:19:07
to friend before listening to a song,00:19:09
most of us use audio every day00:19:11
in different languages and for different00:19:14
purposes. Typically, when large language00:19:16
models interact with audio, they take the00:19:18
audio, run it through a00:19:20
speech recognition system to convert it into00:19:22
text ATEM under this text to another model00:19:25
that understands00:19:27
the flow of many nuances such as Jeni's voice00:19:30
or00:19:32
pronunciation, thanks to native00:19:34
multimodal capabilities, it is able to00:19:36
process raw00:19:38
audio signal I will give an example we uploaded00:19:42
an audio clinic00:19:45
Let's Listen to the video and then00:19:52
Jemina's answer how to pronounce Lunar January in00:19:55
Chinese option a00:19:59
option00:20:01
b Which option is correct which tone00:20:04
is correct for the first00:20:08
character option a The first character00:20:11
is pronounced with the first00:20:14
tone Janie was able to distinguish between two ways of00:20:17
pronouncing a word and understand the correct one in00:20:20
this video there was only one speaker00:20:22
let's look at an example in which00:20:24
more than one person is speaking00:20:28
Al Pi Curs00:20:33
Inc canar00:20:41
R Hi I don't understand this podcast00:20:45
Could you briefly retell it in00:20:51
English under the story that is in fashion in00:20:57
this00:20:58
canary and Scarlet red color which is00:21:01
not in fashion in this season night blue00:21:05
thanks to the multimodal capabilities of00:21:06
Janie can understand the content of00:21:08
the conversation making it not only a00:21:10
transcription model but also a model that00:21:12
has a much more subtle understanding of00:21:14
the environment environment let's take a00:21:16
closer look Janie works natively with00:21:19
visual data audio and text and00:21:20
therefore can understand them00:21:23
together let's look at the last00:21:27
example00:21:29
What you need to do to prepare a00:21:31
vegetarian omelette with these00:21:36
ingredients break the eggs into a bowl and00:21:39
beat00:21:44
them Thanks for00:21:46
the instructions I started cooking the omelette it00:21:49
looks00:21:54
ready it seems almost ready You can00:21:57
turn it over to cook the other00:21:59
side why is it not00:22:04
ready it is not ready because the eggs are still00:22:11
runny And00:22:18
now the dish is ready and you can00:22:21
remove it from the heat and00:22:23
serve as we saw can00:22:26
simultaneously process information00:22:28
audio images and00:22:31
text we We believe that Janie's ability00:22:33
to listen to the original audio will help us00:22:35
expand its capabilities and make it00:22:37
more useful for00:22:43
people we originally created Janie00:22:46
multimodal including some very00:22:49
important programming00:22:52
code is able to consistently understand00:22:54
explain and generate correct and00:22:56
well written in most00:22:59
programming languages including Python Java c+00:23:02
Plus and Go This greatly improves00:23:05
coding ability compared to the00:23:07
previous two models, with about 20000:23:09
programming functions in Python it00:23:11
can solve about 75% on the first try00:23:14
versus 45% on00:23:18
Palm if you let Jemina check and00:23:21
correct errors in answers this number00:23:23
will increase to more than 90% And this is a huge00:23:26
step forward you will be able to create new00:23:29
ideas and prototypes in a second Let's00:23:32
try I really like trains and I00:23:34
want to create a web application to find00:23:37
good places to watch00:23:39
trains I can get a working00:23:41
prototype in less than a minute Although the code is not00:23:44
perfect having a first draft is very00:23:48
useful in itself can00:23:50
transform the software development process00:23:52
but can also be used00:23:54
as a key component of more00:23:56
complex systems00:23:58
and is excellent at writing code00:24:01
but we were able to go even further by creating a00:24:02
specialized version that00:24:04
shows excellent results in00:24:05
competitive programming why00:24:08
we care Competitive00:24:10
programming is one of the best00:24:12
tests of algorithmic00:24:16
coding abilities. Thousands of talented00:24:18
programmers from all over the world come00:24:20
together to compete and00:24:22
try to solve incredibly complex00:24:24
problems that require not only writing00:24:26
code but also mathematics. 2 years ago we00:24:30
introduced Alpha Code. It was the first and00:24:33
system that could compete00:24:34
at about the level of the average00:24:37
person Today I am pleased to present to you00:24:39
Alpha Code 2 a new improved00:24:42
system with significantly increased00:24:44
performance on00:24:48
the basis of when we evaluate Alpha Code 2 on the00:24:50
same platform as the original00:24:52
Alpha Code we solve almost twice as many00:24:56
problems Alpha Code could get to the00:24:59
top list among human competitors00:25:01
Alpha code 2 works better than 85% of00:25:06
competition participants Let's look at our00:25:08
model in action One of the most00:25:10
difficult problems we have encountered00:25:12
I say difficult because in the00:25:14
original competition in which00:25:16
this problem appeared, less than 2% of participants solved it00:25:20
the task is quite complex and very00:25:23
abstract, so I can’t go00:25:25
into details, but the gist of it is that our00:25:28
task is to calculate aggregate statistics00:25:30
taking into account what seems to be an incredibly00:25:32
large number of random00:25:36
arrays. The most interesting thing is that to00:25:39
solve this problem, Alpha Code 200:25:41
uses dynamic programming.00:25:44
Dynamic programming is an00:25:45
advanced algorithmic technique00:25:47
that essentially simplifies a complex task00:25:50
by breaking it down into simpler subtasks00:25:52
over and over again and what is really00:25:54
impressive is that Alpha Code 200:25:57
knows how to correctly implement this00:25:59
strategy but also when and where to00:26:02
use it, this example shows us00:26:04
that competitive programming00:26:06
is not only about implementation but also understanding00:26:08
the mathematics of Computer Science and writing code00:26:10
Which makes it an extremely difficult00:26:16
task so it is not surprising that so00:26:18
far publicly available large language00:26:20
models have shown very low00:26:22
results in this00:26:24
benchmark these00:26:26
Modeler execute instructions but Alpha00:26:30
code must do More it must00:26:32
show some level of understanding00:26:34
some level reasoning designing00:26:36
code solutions before he can00:26:38
get to the actual implementation to00:26:40
solve a problem and he does all this on00:26:43
problems he's never seen00:26:45
before. Another great00:26:48
thing about Alpha Code is00:26:49
that it works even better when you00:26:51
collaborate with code people who00:26:53
can provide the rationale in essence,00:26:55
developers can set the properties of the cat00:26:57
must obey the code examples and when00:26:59
we do this, productivity00:27:01
increases significantly. We believe such00:27:03
interaction between programmers and00:27:06
the future00:27:09
of programming where programmers will00:27:11
not just give instructions but00:27:13
collaborate with highly efficient00:27:15
models and who can reason about00:27:17
problems suggest code design and even00:27:19
help in its00:27:21
implementation Alpha Code 2 was created for00:27:23
competitive programming, but we are00:27:25
already working to transfer00:27:28
its unique capabilities into a common00:27:29
model as a first step towards00:27:32
making this new approach to00:27:34
programming accessible to00:27:38
everyone. As a parent, you may have to00:27:41
help your kid with00:27:43
homework of course I had to00:27:44
here Jamie is not and can help for this00:27:47
demonstration We have created a simple00:27:48
interface And with the help of smart hints we00:27:51
can really use00:27:54
Janie's mathematical reasoning and multimodal capabilities to00:27:56
study a subject like00:27:58
physics with Janie's help You can00:28:00
upload a photo of handwriting answers00:28:02
on the worksheet Janie can solve not00:28:04
only these problems that Great she00:28:07
can also read the answers and understand what was00:28:09
right and what was wrong and also explain00:28:11
concepts that need00:28:13
further00:28:15
clarification so Jamie didn't identify00:28:17
some errors in problems about and 3 Let's00:28:21
look at the00:28:24
third here ode that would be correct but there was an00:28:28
error in calculating the height we00:28:31
can ask Janie to explain in more00:28:33
detail why the height is 50 mm and not00:28:42
six I can ask Jamie not to00:28:44
explain00:28:51
further explains the step-by-step details of00:28:53
solving the00:28:56
problem based on Janie's ability to understand the00:28:59
nuances of information and provide answers to00:29:00
complex topics she can give you a00:29:03
personalized explanation of the subject00:29:04
you are trying to00:29:07
learn If you want to know more You00:29:10
can Just00:29:14
ask Janie will provide00:29:16
customized practice problems00:29:18
based on00:29:20
errors here I have a similar problem where you00:29:23
need to calculate the speed of a cat for a00:29:25
height twice as high00:29:29
Well yes I do00:29:33
knew Let's see if our00:29:36
multimodal model Jeni can find00:29:39
similarities between00:29:40
the images let's start with a photograph of the00:29:42
chapel and these engravings and bite them and I00:29:44
will ask Jeni to find a connection between these00:29:46
two00:29:49
images let's see what Jamie00:29:52
says the curved and organic composition of00:29:55
the building is more subtle and the second image is00:29:59
more fluid Yes let's go Let's00:30:02
try another option using the Moon00:30:04
and this golf ball from my webcam and00:30:07
then I'll run the same query So let's00:30:09
see in 1971 The crew of Apollo 1400:30:13
hit two golf swords into the lunar00:30:15
surface Wow That's pretty good00:30:18
then one more time Who wears it better stripes00:30:21
Zebra about I like it Zebra has been wearing00:30:24
his stripes for millions of years00:30:27
these are some examples of visual00:30:29
understanding of00:30:36
Janie Let's see if our00:30:38
multimodal model Janie can guess00:30:40
several movies So we will start with the fact00:30:43
that based on the play on words and this video you need to00:30:46
guess the name of the00:30:48
movie00:30:50
breakfast club good A what's on00:30:53
this Tomorrow00:30:56
those are diamonds in the rough cool cool00:31:01
works a couple more quick00:31:05
tests Gold fnr good bottle rocket00:31:10
Okay Wizard of the Land of OZ Good00:31:14
good Moonrise Kingdom Good the00:31:18
last test is a little more difficult M good I didn't00:31:22
think it00:31:25
would work out guessing movies with00:31:33
Janie's help in this test we'll see00:31:36
if our multimodal model Janie can00:31:39
figure out how some00:31:41
unusual images were created using emoji00:31:43
from the emo Kit emo Kit allows you to00:31:45
combine different emoji to00:31:47
get new ones For example, if you00:31:50
combine this ghost with this avocado you00:31:52
will get00:31:54
this let's see if you can00:31:59
guess I think that it's a ghost and an00:32:02
avocado Now let's see00:32:04
if she can explain the visual details00:32:06
she used emoji has00:32:08
the shape of a ghost but it's green and has a big00:32:11
brown bone Now let's00:32:13
give it a name and a short tagline polter00:32:17
Kado Ghost goka Mole not bad Here are00:32:21
some more tests I did I I would00:32:23
call this party Ghost boogie wgie00:32:28
Rob bunny cowardly robot00:32:32
gray this one I would call Beth grip this00:32:36
dragon is not real that's00:32:45
all Let's see if our00:32:48
multimodal model Janie can understand fashion let's00:32:50
start with something simple for example with00:32:52
this down jacket Let's ask what it's for00:32:59
perfect for00:33:00
keeping warm in the tundra good color to00:33:03
blend in with the glacier okay how00:33:05
about this option intergalactic00:33:08
travel good how about00:33:11
boldly going where no man has gone before00:33:14
and playing jazz so Jamie has a00:33:18
good sense of humor now Come up with a00:33:20
term for this aufi lukor is not bad this is a00:33:25
Fashion verdict with Janie00:33:31
Let's see Can our00:33:33
multimodal model turn an00:33:35
image into code I'll start with this00:33:37
image of a tree just select the00:33:40
part I need and then ask Janie if00:33:42
she can turn This is an00:33:46
svg image that represents the basic shapes of a00:33:49
tree Let's see turned out not bad00:33:53
Now the test is more difficult00:33:57
make an interactive demonstration in00:34:04
JavaScript Okay, something turns out to be a general00:34:07
algorithm For this it is called a00:34:08
fractal tree This is pretty cool00:34:11
Jemine Even provided a slider so that00:34:14
I could change and move the00:34:18
fractals and the code itself00:34:21
Great, that's00:34:25
all00:34:30
Let's see if our multimodal model can do it00:34:32
days to help00:34:34
figure out my apartment and to00:34:35
add a little more complexity I'll see00:34:37
if Janie can handle the clues00:34:39
in Chinese only let's start with this00:34:41
photo based only on00:34:43
the lighting I want to see if00:34:45
Janie can tell by the light which00:34:47
side of the world my room is on and00:34:49
Janie00:34:51
answers yes Jamie doesn't seem to be saying that00:34:54
my room faces south what about00:34:57
this00:34:58
plant what type of Light does it00:35:05
need So Jamie doesn't say it's00:35:07
a dracaena and it doesn't need a lot of00:35:10
sunlight awesome I have a dining room00:35:13
that faces the opposite00:35:15
side of my bedroom I wonder00:35:18
if this plant is00:35:20
better there so Janey00:35:25
says00:35:30
concludes that in my00:35:32
north facing dining room there is less light and00:35:34
therefore it is better suited for this00:35:36
plant so Jamie is not helping with00:35:38
the layout of the apartment new tests coming00:35:47
soonDescription:
https://codebreakers.tech/ai_webinar - Начните зарабатывать больше с помощью нейросетей! Записывайтесь на бесплатный вебинар «Нейросети для жизни, карьеры и заработка» от школы Code Breakers. Реклама. ИП Радзиевская М.В. ИНН 263402002977. Erid: 2RanykGiQgx Google представил новую сверхмощную нейросеть Gemini. Мультимодальный искусственный интеллект станет основным конкурентом ChatGPT от OpenAI. Модель будет доступна в трех версиях: наиболее мощной Gemini Ultra; Gemini Pro для широкого спектра задач и Gemini Nano мобильных устройств. Gemini будет поддерживать приложения Google с ИИ. Источник: https://www.youtube.com/playlist?list=PL590L5WQmH8cSyqzo1PwQVUrZYgLcGZcG&si=Zz849Myuhixl6m5g 00:00 Самая способная модель ИИ Google 05:51 Безопасность и ответственность 06:57 Тест Gemini 12:51 Как ИИ поможет науке 15:30 Индивидуальный пользовательский опыт 18:59 Обработка и понимание Аудио 22:41 ИИ в соревновательном программировании 27:38 Помощь в изучении предметов 29:34 Поиск связей в изображениях 30:35 Gemini угадывает фильмы 31:32 Emoji Kitchen 32:44 Модный приговор с ИИ 33:29 Превращение изображений в код 34:30 Понимание окружения
Preparing download options
* — If the video is playing in a new tab, go to it, then right-click on the video and select "Save video as..."
** — Link intended for online playback in specialized players
Questions about downloading video
How can I download "Google Gemini: Презентация Новой Сверхмощной Модели ИИ" video?
http://unidownloader.com/ website is the best way to download a video or a separate audio track if you want to do without installing programs and extensions.
The UDL Helper extension is a convenient button that is seamlessly integrated into YouTube, Instagram and OK.ru sites for fast content download.
UDL Client program (for Windows) is the most powerful solution that supports more than 900 websites, social networks and video hosting sites, as well as any video quality that is available in the source.
UDL Lite is a really convenient way to access a website from your mobile device. With its help, you can easily download videos directly to your smartphone.
Which format of "Google Gemini: Презентация Новой Сверхмощной Модели ИИ" video should I choose?
The best quality formats are FullHD (1080p), 2K (1440p), 4K (2160p) and 8K (4320p). The higher the resolution of your screen, the higher the video quality should be. However, there are other factors to consider: download speed, amount of free space, and device performance during playback.
Why does my computer freeze when loading a "Google Gemini: Презентация Новой Сверхмощной Модели ИИ" video?
The browser/computer should not freeze completely! If this happens, please report it with a link to the video. Sometimes videos cannot be downloaded directly in a suitable format, so we have added the ability to convert the file to the desired format. In some cases, this process may actively use computer resources.
How can I download "Google Gemini: Презентация Новой Сверхмощной Модели ИИ" video to my phone?
You can download a video to your smartphone using the website or the PWA application UDL Lite. It is also possible to send a download link via QR code using the UDL Helper extension.
How can I download an audio track (music) to MP3 "Google Gemini: Презентация Новой Сверхмощной Модели ИИ"?
The most convenient way is to use the UDL Client program, which supports converting video to MP3 format. In some cases, MP3 can also be downloaded through the UDL Helper extension.
How can I save a frame from a video "Google Gemini: Презентация Новой Сверхмощной Модели ИИ"?
This feature is available in the UDL Helper extension. Make sure that "Show the video snapshot button" is checked in the settings. A camera icon should appear in the lower right corner of the player to the left of the "Settings" icon. When you click on it, the current frame from the video will be saved to your computer in JPEG format.
What's the price of all this stuff?
It costs nothing. Our services are absolutely free for all users. There are no PRO subscriptions, no restrictions on the number or maximum length of downloaded videos.