Created
February 16, 2024 19:53
-
-
Save gsans/99d090e83cad349c8ed0bf79612b038e to your computer and use it in GitHub Desktop.
JSMP 5: Gerard Sans on OpenAI – The Start of a New Era in AI (Transcript)
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
foreign Poland conferences are coming soon | |
this year promises to be exceptional we will see the angular team on the stage | |
we will see quick create solid.js Creator experts from Google Microsoft | |
Amazon Cisco all zero and many many more join an amazing group of developers like | |
you today let's come together to celebrate angular and JavaScript go to | |
ngpaulin.pl and sign up now [Music] | |
what's up everyone this is founder of AMG Poland JS Poland England mastered to | |
death and workshopfest.dev welcome back to the JavaScript Master podcast today we've | |
got a special guest from London UK M.C speaker trainer Community leader ladies | |
and gentlemen Gerard sounds [Music] | |
[Applause] hi guard how are you | |
hello hello Eric I'm very good okay so for those who don't know you yet please | |
tell us about yourself well I'm Gerard Sands I'm a developer evangelist for web | |
and cloud and I've been doing some research on artificial intelligence and | |
web 3 for the last year and I'm here to share some of my learnings | |
Will AI replace our jobs | |
opening question will AI replace our jobs | |
well this is this is uh for some people uh very uh scary uh time we have seen a | |
lot of improvements in the last few years and some of them affect how people | |
approach their jobs and probably the most significant changes has happened on | |
text to image generation and this is where we use a prompt and an AI is able | |
to generate images from scratch and we have seen | |
[Music] crazy improvements there these AIS can generate illustrations can | |
generate photorealistic images and it just gets better and better and a lot of | |
people working as a designers or maybe as a digital | |
artist they can see that their jobs are on the line and there's no clear | |
separation of what we as people will | |
accept from a digital artist or from maybe an AI artwork which of these are | |
Writers artists and developers | |
more likely to be affected writers artists or developers well that's | |
something that it's been very surprising for me I um I started looking at some of | |
the text that AIS were able to uh to create and this is very similar to what | |
I was explaining from images but with text and the idea behind this | |
is that these AI models have been trained using a very large amount of | |
data and public books from the last few centuries so this is | |
anything from all the possible styles that you can | |
imagine commercial books self-help cooking recipes but also philosophy math | |
and all of the science and if you think | |
about big authors like Shakespeare or other | |
um popular authors in history these are also in in these models so actually you | |
can create text that resembles any of the most famous writers in the world and | |
of course including marketing and any other | |
areas so that's that's one one thing that if I was a writer I would use these | |
tools to maybe avoid what it's called the writer's | |
block so you can get inspiration of | |
of a history that you are kind of stuck and you can get some text to continue | |
a paragraph that you have written but people is using this to generate | |
anything from a blog post to cooking recipes or anything in between so it's | |
it's quite it's quite exciting but also it's quite scary for for different jobs | |
like technical writer or even a novelist | |
these these are all in the line and most people know about this technology but | |
they don't think that maybe as developers they will be affected by this | |
but it was only last year that open AI | |
it really is a codex which is the model behind GitHub copilot and it's able not | |
only to write code from scratch but also do refactors which which is basically | |
anything that you need as a developer you can write a code and then iterate | |
over it as many times as you need and this is what these tools are doing today | |
so it's uh it's quite a new era of AIS I would say | |
What is OpenAI | |
this is crazy so um yeah completely insane so what is open AI openai is a | |
company that was founded by Elon Musk son Oldman and Greg Brockman and | |
what they say about their mission is that they want to build a safe | |
artificial general intelligence that benefits all humanity and they created | |
open AI in back in 2015. as a non-profit company | |
and it took them few years until 2018 to create the first GPT model which is a | |
generative pre-trained Transformer but then they went ahead and published | |
and a public API which was in 2020 which is | |
when everyone started trying these technology and they discovered many | |
applications the initial ones were around writing but | |
they also discovered that gpt3 was able to go JavaScript | |
and that was when uh when they started working in in codex | |
which was released on 2021 so this is a very recent technology it's just been | |
around for a couple of years and the problem is just fantastic what products | |
Products from OpenAI | |
are being offered today there are three different products and | |
the one that I was talking at the beginning is gpt3a and this is the model that is | |
able to generate tax and the way it will work is that we provide a prompt | |
and gpt3 will continue that prompt in a way that makes | |
sense with the with the actual text that he founds | |
and you can use that for example to create a story | |
but it has also been used to | |
follow a conversation and this is some of the things that were | |
not plan this model was created to generate text | |
and predict the next word following a prom but then they realized that this | |
model could identify patterns like a series of | |
questions and answers and then that will start giving you answers back in a way | |
of uh of a conversation and they found that gpt3 was able to do | |
a lot of different things you can for example ask gtt3 for cooking | |
recipes you can ask for how to cook a certain meal and gpt3 will give you the | |
ingredients and how to uh how to cook it and that's not really | |
a text generation anymore in the sense of writing a story or writing a book | |
but it's more like identifying a pattern identifying a question and providing an | |
answer to that question which was completely unexpected uh for the | |
scientists that were building this uh these models but then they started doing more tests | |
and they identified a lot of different usages for for gpt3 and that was kind of | |
the beginning one of the things that they realized using gpt3 is that it could actually | |
write HTML and JavaScript and that brings the second product from | |
open AI which is codex when they realize that gpt3 was able to | |
write JavaScript they um they train gpt3 | |
only with a source code so that was when openai contact | |
Microsoft and Microsoft gave access to GitHub and the parts I think 59 million | |
repositories and it learned a list of 12 | |
different languages including SQL and Bash commands | |
What is GPT3 | |
well said what is gpt3 trying to solve well gpt3 | |
was um natural language processing model that was intended to | |
generate tax and the idea behind was | |
taking the prom and trying to guess what will be the next work that follow that | |
problem the thing is you can repeat that sequence so you can generate as many | |
words as as you need and then that will be the main operation for gpt3 so you | |
can provide a text as an input and then it can provide any number of words | |
usually what you would do is give it a stop sequence so for example a new line | |
will be the stop sequence so it will stop after the first paragraph but you | |
can you can create complex patterns as a chat so you can make a question and | |
answer text and gptc will follow that or you | |
can ask for instructions and it will provide the instructions to achieve a goal | |
like cooking a recipe cooking recipe sounds | |
GPT3s Languages | |
so weird for me created by Ai No it's uh it's it's quite | |
it's quite um new the idea behind is that these | |
models are very large models and | |
gpt3 is actually 175 billion parameters | |
uh large is is the size of the whole model but that means that it has more | |
parameters than the size of the atoms of the universe so this is a very this is a | |
very big model that we cannot imagine and it has a lot of things inside and of | |
course one of the things it has is a cooking recipes but this is not this is | |
not the most weird thing that you can find inside this model because it | |
literally codes in JavaScript but not only JavaScript I was | |
mentioning some of the languages it learned but it has learned | |
typescript JavaScript python Ruby go C sharp Swift rust | |
PHP all of these languages after just a | |
little bit less of half a terabyte of information | |
but this is also not the the most weird thing probably the most weird thing is | |
that it it can speak more than a hundred languages when it was only trained in English | |
which is also very surprising some languages gpt3 learned with less | |
than one percent of uh data which I don't think nobody knows how | |
that happened it was supposed to learn English but because there were like some you know uh | |
in the data in the training data it took information from different sources and | |
one was a common crawl so it took the top thousand websites | |
in Alexa and it took all the content and that's how it learned JavaScript because | |
a lot of the raw content of a common crawl is HTML and JavaScript but then it | |
also took hundreds of years of public books and | |
funny enough it also took the information from Reddit | |
so there's a lot of information from users and the links that these users provided | |
so there's anything any any information you can imagine is in there | |
and then probably one other resource that was in the training data was | |
Wikipedia so anything that is in Wikipedia is also knowledge of gpt3 | |
so it's a massive amount of data and yeah it can do funny things like | |
translate between languages so you can write something in French and it can give you the translation in Polish | |
why not it's it's completely it's completely insane and if you look at the | |
complexity of a language like English then you can understand how it's able to learn JavaScript because JavaScript is | |
much easier than it's yeah yeah exactly exactly but if it has | |
learned uh languages like Japanese yeah if | |
how the question is more like how how we'd learn Japanese with almost no | |
information yeah um and we're gonna see more I think that | |
this is just at the beginning and we haven't seen uh all the all the power I mean we have | |
seen a little bit we we have seen like the tip and uh because people hasn't got time to | |
experiment we are talking of uh of something that it's massive | |
um in artificial intelligence that's uh called latent space so this this phase | |
of 175 billion parameters is so big that you will not have time in your entire | |
life to explore it um where we are talking about uh gpt3 | |
giving you cooking recipes this is only a small uh grain of salt in the imagine | |
if it was a beach or maybe all of the beaches you know the whole world the | |
sand it will be just taking like a little a little uh grain of of sand I | |
mean this is nothing compared to the size of of the whole gpt3 so the the | |
idea that I want to share with uh the people is that we haven't seen anything | |
I mean we have seen some things that are surprising because nobody was expecting | |
a model to understand um JavaScript but now it's been more than | |
half a year being used in production with GitHub copilot and the service is | |
still running so I guess it's working so it's giving code to professional | |
programmers around whatever they are building I mean that's that's very impressive | |
CodeX | |
you're listening JavaScript Master podcast listen code repeat everything | |
you need to know to become an JavaScript super developer | |
while these codex's main purpose and for for codex is | |
generating uh code so we have talked about gpt3 being able to write as | |
Shakespeare which is a quite quite fun but it can also write in any in any | |
style that you can imagine so if you if you can find enough information | |
which gpt3 can have been trained on so it must be a popular writer it cannot be | |
someone that has uh not many books or materials but for codex is anything that | |
it's on GitHub and part of the training was all of these 59 million repositories | |
so you can imagine that anything in GitHub it will be there so codecs can | |
help you use all of these libraries the only thing that you need to take into | |
account is that codex was trained with a snapshot | |
so it's not being trained daily because it's very expensive | |
process but it was trained last year so there's a snapshot of last year whatever | |
was in GitHub last year so that that means that it's angular version 13. it's not | |
angular version 14 because that happened after the release was after so anything that | |
it's after the snapshot codex is not aware of | |
what is Dali and how people are using it | |
well Dali is probably the most popular of the of the three Dali is everywhere | |
is social media is on Tick Tock is on Twitter every day there's people that | |
goes into Dali and generates an image so you can you can see a new image every | |
day if you follow these users Dali is following this Trends so we have seen | |
how these models generate text how this model generate code and this is exactly | |
the same for images the only thing that they did here is | |
they took all the knowledge around languages so everything that gpt3 knows Dali knows | |
that means that it's a writer it's also | |
a developer that knows all these 59 million GitHub repositories but now it | |
has this information mixed with images and they have trained | |
gpt3 for Dali with around 400 million | |
images which is completely insane again and the main | |
the main feature that it can it can | |
it can generate any sort of images following the same as gpt3 and codecs | |
you need to provide a prompt and then from this problem it will generate the | |
image following the instructions and this is this is very surprising because | |
you can not only generate image from from scratch but you can also edit | |
images and give instructions so for example one | |
of the experiments that I did recently was to create a photorealistic picture of a woman | |
which can create no problem for tally so | |
it creates the picture following all your instructions you can provide | |
information from the composition how the model is positioned in depiction what is | |
the lining what is the camera that it's being used and | |
I I got something no I got the picture I was after and then I changed the image | |
and ask tally to add I black leather | |
jacket to the to the model and it did so you can you can ask things | |
as if it was some kind of a system that is helping | |
you to create these uh this setup um so you can ask things like at | |
sunglasses to the model or maybe add a scarf | |
but not only it will add these elements to the picture but it will take into | |
consideration the volume it will take into consideration the lining it will | |
just place that item in the model and give you the resulting image which | |
is completely insane and it cannot it cannot only do | |
photography because photography is probably the most difficult but you can do pixel art it can do 3D models it can | |
do games it can do illustrations of course photography is completely crazy | |
but you can do like drawings like hand drawings uh | |
it's it's it's really something it's really something else people is also | |
creating um games um because games usually when you need | |
to create a 3D model object it's quite um it's quite a lot of work | |
and sometimes you need to create a texture imagine you are creating a | |
castle and there's some pattern in the in the walls which is uh from Brock I | |
mean imagine there's some there's something else maybe it's wood it's a wood wall and it's difficult for an artist to | |
create this kind of textures but you can give it to Dale and it will create the | |
texture for you you can then use it in a 3D model and in a few minutes you have | |
got everything and you are ready to go there's people building games like super fast these | |
days that sounds sounds really amazing yeah there's also some fancy uh use | |
cases one of them is for example now that I can see you uh in the call I can | |
see that you have like a a sofa in the background so you could take a picture | |
people is taking picture of a room and you can ask dally to add Furniture | |
so imagine there's something missing maybe you want to add some artwork in | |
the box like some pictures so you can you can give this task to Dali and it | |
will just decorate your room in different styles if you want | |
people is also using Dali to recreate historic | |
scenes like imagine you want a picture from | |
Normandy you know normally when in the second world war yeah so you can ask | |
Dali to give you a selfie from a soldier in Normandy it's it's | |
anything that you can imagine dollies is able to use all the | |
information that you give it it's more like a die a direct a creative director | |
and it just puts all the information uh | |
it makes and it makes it work of course the Technologies is very new we are | |
talking about a technology that last year didn't exist so in just a year it | |
has done a lot of improvements so we will see many many other things happening in all of these different | |
products but Dali is probably the most because it's visual it's the most | |
surprising can you use open night today how would | |
OpenAI Today | |
that work well there's a there's an API that you | |
can use some of the apis are open to anyone tpt3 is completely open everyone just | |
needs to go create an account and start using it and the only thing is that it's | |
not for free so you will have to buy some credits and then | |
activate that access with the API key | |
like you Google with other services and then if you want to have access to | |
codecs you need to apply for a for a better beta access | |
and also for that dally takes a little bit longer so if you apply today it will | |
take a little bit longer but you can use all of these Technologies without | |
without any problem what are the benefits of using those tools are they | |
Benefits and Drawbacks | |
any drawbacks I think the benefits today is that they will boast your efficiency | |
so some of this some of the tasks that you can do with gpt3 like for example | |
imagine you want to create a technical blog post that will give you like a big | |
post so if you needed two days to edit | |
your your text and go through different revisions with gpt3 you can probably do | |
that in in few hours or maybe in a couple of hours with the other products | |
it may depend what what is that you want to do sometimes the quality is not there | |
yet but you can you can use parts of the of the results of the outputs to speed | |
up your project so you you can get uh more efficient you can do things faster | |
I I mentioned for gaming how they can create these textures that before you | |
could have took days for digital artists to create so it just depends on what | |
exactly is that you want to use it for but yeah it's going to give you that productivity | |
um as drawbacks I would say that depending of how you use it there's no | |
easy way to know if the information that maybe egpt3 is giving you it may not be | |
factual so if you are writing a blog post and some of the information is not | |
factual it can bite you back because somebody will find out and then you will | |
be embarrassed that you have used a source that is not it's not a search | |
so it's not something that you can trust so probably depending how much you rely | |
on AI you need to check the facts you cannot just take the output from | |
gpt3 and use it in a publication just uh release it you will have to to check the | |
facts or you may be embarrassed for other things it just depends what you | |
try to do sometimes when you try to do something very complex there's some | |
issues with the technology is not perfect so you can get some artifacts for example in in Dali when it | |
creates an image faces like human faces and | |
human models in the pictures they don't look totally they look odd at times sometimes | |
you need to edit the pictures so they are not all the time like 100 perfect | |
but I think this these errors or these | |
drawbacks are gonna be fixed and then it's going to be much much faster to use | |
these tools the million dollar question what the future of AI may look like in | |
The Future of AI | |
five years I I just did I just did the talk around | |
this topic and I describe um the AI Revolution so this is um from | |
the work of different people and the first stage it's called cognitive | |
Revolution and what is the cognitive Revolution this is where we use AI | |
assistance to get our job done faster and depending of what you do that will | |
work best or or not I mean even for GitHub copilot it's not that the tool is | |
creating everything for you that is just giving you some kind of um a | |
productivity uh boost then The Next Step which is a little bit more | |
crazy and it's probably not going to be uh that soon so this one using AIS | |
systems is going to be very soon some people is already using this today but | |
the next step is the robotics Revolution this means that we will see the first | |
robots using Ai and being of tournaments | |
and maybe we have seen some of these autonomous robots I think I saw the | |
first one in in Silicon Valley some years ago they go around a building and | |
they can give you information around what you can visit | |
but of course this is not something that you can find everywhere but in the | |
future we will see more examples of autonomous robots and this is the second | |
stage the third stage is when AIS get | |
smarter than humans and this is called the singularity | |
and when we reach the singularity this is uh similar to the character in Star | |
Trek Data I don't know if you if you remember data from Star Trek this is a | |
Android that helps Star Trek characters when they when they | |
struggle and they go into Adventures when they have questions about what they | |
are seeing like maybe there's some danger and they need to decide make a decision if they they go into a planet | |
or not then they will ask data and data will be like some kind of | |
uh super intelligent being that will give some advice in that situation and | |
this is similar also uh to the intelligence of a spook uh character | |
so these are some these are some some of the things that may happen in this AI | |
Revolution and and people is is not um | |
certain when all of these different stages will happen but uh there's a | |
quote from Elon Musk that was from not | |
long ago and he said that we were going to be replaced by AIS in | |
the next five years so that's that's what Elon Musk uh has to say around this | |
topic do you think AI can help us build something we can't do ourselves for now | |
AI for Rocket Engines | |
like a super engine for a rocket oh yeah definitely definitely yes | |
um what what is happening is that with these models we have created some kind | |
of uh some kind of a super intelligence not not in the sense of because there's also | |
the debate if they are sentient or if they are autonomous or if they are like | |
uh some kind of a god entity I mean I don't think they are a god entity or | |
anything like that but the the main thing is that they have all the | |
information accessible at the same time so one of these uh | |
technologies that is being used behind the scenes is neural networks and a | |
neural network it's not doing calculations all the calculations are | |
already done and that's how these neural networks are so fast that you can use them for | |
example in self-driving cars and other systems they are in the range of | |
milliseconds of the response time so they are they don't need to calculate they are not it's not working like a | |
search imagine that some people think that these AIS they work like a search engine so they have a query and there's | |
some computation and then it's coming back with an answer but it's not actually working like that the neural | |
networks is more like a human it's it's not even thinking it's giving you the | |
answer right away when you use the systems gpt3 and codex it's not taking seconds not | |
even milliseconds to answer back and that means that all of the information is accessible to them like | |
right now it's they don't have to think because that's part of the training | |
process which is very expensive and it takes weeks but one the training is | |
finished the model is already is already done | |
and that means that it has all the skills at the same time so when you access gpt3 you are talking to | |
something that is it has Shakespeare inside but it also has everything else | |
it also it's also a developer it's also a chemist it's also a mathematician it's | |
also it has everything so when you ask him for example uh how | |
we will build a rocket engine it will give you the answer right away it's not | |
something that he needs to think like okay I need to think now he will give you the answer right now | |
the moment you ask it's not it's not gonna tell you oh let me let me uh think | |
about it and next week I'll give you an answer no no it gives you the answer right now and this is for anything when | |
it looks at code when it looks at uh at the image or text the model is giving | |
you the answer in the same moment so that means that it has access to to all | |
of the knowledge all of the knowledge that it has been trained of course if the if the | |
information is not in the training data it's not like it's it's a god it's not a God but it's very capable | |
it's like very very capable but it's not it's not a god some kind of | |
a new um but it has it has all of the information | |
the same as it could be Google but it can work with the information it's not | |
only um it's not only like a library where you have the information and then it retrieves the information for you it has | |
all the mix of all the all the information and it has the results | |
um at that same moment so it doesn't it doesn't do calculations intermediate calculations so it's quite impressive | |
it's a new technology and we are still discovering uh how we can use it so what | |
the conclusion oh the conclusion is that I don't see nobody knows what is | |
happening with uh with these models but it's exciting because people find new uh | |
use cases and they find new things that you can use these models and there has | |
also been news around open sourcing some of these Technologies so you can use | |
them in your projects so that's also very exciting after all these discussions can you give some advice to | |
Advice for New and Old Timers | |
people who are starting their career in it world today | |
and for those who are Old-Timers I think I think a good advice is to to keep your | |
mind open because everything is changing everything is changing people think that | |
things don't change but in in the area of AI this is uh probably the the better | |
example AI was is not a technology it's not a area that has been changing a lot | |
and it's kind of an old area but the | |
discovery is done in the last few years they have exploded and this is new | |
it's not there it's more like JavaScript you know they're all JavaScript and the new JavaScript that happened a few years | |
back and people were surprised like oh can you do this with JavaScript I mean before we didn't use JavaScript in this | |
way so that's a little bit of a new generation of AI and this is this this will change | |
everything and a lot of the jobs that we do today they won't exist in the future | |
so if you are not retiring uh I mean if you are an old-timer you maybe still | |
have five ten years or 15 years more of career I think you should you should | |
look into what's happening in in the blockchain what's happening in web3 keep | |
your mind open because I mean we can become obsolete like very fast for the | |
Alzheimer's for the new uh Generations I think that they also need to be careful | |
where they put their energy because if they put their energy learning a | |
specialization and then they realize that anyone could be able to use any | |
technology because they will use AI assistance they will be wasting time why why you want to | |
memorize things or understand things when an AI can help you and you can use | |
their eye or anyone so then it's not it's not clear that you want to go | |
through a specialization of 15 years and then find out that someone that has no | |
specialization is using a tool that is making that instead of 15 years is just | |
five years so then you have lost 10 Years Learning something that it wasn't necessary | |
um so that's that's something to to be careful let's change the subject a bit | |
Last visit to Poland | |
how do you remember your last visit to Poland OH I am | |
I I remember as a crazy as a crazy time I mean it's always it's always fun | |
and it's more like a wheel wheel so it all starts for me because I was doing a | |
workshop and also participating both in NG Poland in Gs Poland that was uh | |
that was a crazy A crazy three days and | |
yeah I I really have a great time meeting all the all friends and also | |
people that comes and asked me uh questions probably the most surprising | |
from last year was the award that I received for the community contributions | |
to JavaScript so that was very special and then also all the fun we have in the | |
in the back States and we we also make uh jokes and I remembered and that Derek | |
put me in the spot few times and I was like what's going on I was going crazy | |
because I didn't know what was happening but this is uh also one of the things | |
that I that I like from from the conference my last question for today | |
JSMP 2022 | |
what we can expect at your emceeing at NG Poland and JS Poland | |
2022. oh it's gonna be full of surprises and I think there will be some | |
artificial intelligence uh coming but I don't think I can I can share it with | |
you at this time but the experiments uh I've shared with tarek | |
and and he was pretty was pretty surprised so we'll see if I can I can | |
make it happen during the conference otherwise as always uh I will make the | |
attendees do something it could be jumping it could be uh doing a squat I | |
don't know we'll do a Mexican wave I mean who knows what will happen on the day | |
Gerard thank you so much for joining today's podcast and sharing this | |
exciting knowledge with us I look forward to our meeting in Walsall soon | |
yeah thanks a lot for inviting me and I'm really looking forward to another | |
edition of the conference thank you so much bye bye bye bye | |
everyone finally please subscribe to our podcast | |
leave a like and a comment to help us continue to grow | |
foreign [Music] |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment