 Any topic any discussion on AI must start with addressing the elephant in the room Which is chat GPT, right? Chat GPT is one of the fastest Services in the history of business to reach a hundred million users Just two months all of you are aware of that. I'm sure all of you have used chat GPT in the last three months So what I would like to do is to unpack What does it take to build something like a chat GPT? And what is the future of AI? Let's go back a little bit. Let's go back 10 more years in 2011 Steve Jobs unveiled Apple iPhone 4 s and an integrated feature of that iPhone 4 s was Siri the language assistant and Siri was making some incredibly comical mistakes. You can see some of them on the screen here incredible mistakes, right? And people thought you know what this is not going to work, but just in five short years after that Sundar Pichai in 2016 mentioned that 20% of all Google searches were voice searches The technology had evolved a lot in just those five years. What happened there? Just to understand this a little bit and to build this mystery a little bit further in 2020 Sam Altman who is the CEO of Open AI made this statement in one of the one of the interviews that remember, you know data Important but even more important than data is compute. How much compute you're throwing, right? So what did? Chad GPT exactly do For that you have to understand this equation one of the most important equations to understand in the world of AI is this equation which is error is a function of data compute and the techniques that you're bringing to the party Data all of us understand what data is right the more the better higher quality of data better it is if you bring in lots of data and then throw in incredible amount of compute and Bring the right techniques together. You can bring the error rates down What happened with Chad GPT and Siri and voice assistants and so on is that? Data compute and techniques multiply the multiplicative effect of these three meant that the accuracy improved from 17% error rates back in 2011 when when Steve Jobs launched iPhone 4s with Siri Error rates were 17% in voice recognition Therefore, it was not able to understand a language accents, etc And now the error rates are in this are sub 4% which is which means it's better than human accuracy and Then adoption really took off and that is one of the reasons why Chad GPT reached hundred million users in such short time Here's what they did if you see the data in GPT one they used much even smaller data But in in GPT two they use 1.5 billion parameters. This means 1.5 billion thing Nobs that they can tweak 1.5 billion knobs that they can tweak to get the right effect From 1.5 billion by by 2021 in GPT three they went to hundred and seventy-five billion parameters 175 not 175 billion knobs that they're tweaking to fine-tune the engine They really threw in incredible amounts of this is such a large model That's why these are called large language models, right now What do the what is a compute they threw in they threw in two hundred and eighty five thousand CPU course and 10,000 GPUs put them together To deliver this performance and the data that the data the ruin was 570 gigabytes. This is by some estimates. It's not official estimate, but 570 gigabytes Which is 30 times the size of Wikipedia 30 times the size of the all of Wikipedia So they brought in tremendous amounts of data they threw in large amounts of compute They built very large models with a new technique called transformers Transformers as a technique was developed in 2017 by Google researchers. They use transformers They put in this tremendous amounts of compute and they made some more, you know technique wise innovations So they did data compute and techniques and therefore brought the accuracy down Accuracy to such a level that it became magically good all things all good things in technology appear like magic when they're launched And this is one of those magical things that's happened But the behind the magic are these three things data compute and technique now Now if if you're in if you ask the question What is the hottest programming language in? 2020 the hottest programming language and 2021 also and 2022 also the hottest programming languages were JavaScript and Python 2023 it appears the hottest programming language is now English Because now if you put if you write code if you write English and that's in the comments in the code Code gets generated automatically. This is github co-pilot, right? So what's happening is it is dramatically accelerating given AI is become so good That's actually accelerating tech imagine if you're a 15 billion dollar IT services company And you suddenly realize that you can beat 100x 10x to 100x in terms of productivity The productivity of the IT industry is 10 lines of code per day 10 lines of good code per day Right if they can write 10 lines of code in the in the first one minute What are they going to do for the rest seven hours and 59 minutes and in their day? Just imagine what's going to happen to that, but the hottest programming language is now English, right? So the question that keeps coming out coming back is is AI going to replace jobs? Will AI replace human beings? No, AI is not going to replace human beings But what's going to happen is that humans that use AI will replace or outperforms humans that don't use AI That is the key thing to take out is that this tells you More than ever ever before that humans that use AI will outperform humans that don't use AI This AI is a paradigm shift in computing because if you think about what was computing for the last hundred years It was about telling machines what to do and machines faithfully reproducing what to tell them to do that has changed These machines can think this was the this good old-fashioned AI that you see on the top is what AI was back then It was about teaching machine lots of rules to understand and learn from no the new AI doesn't work like that Machines can be trained can learn to think that with examples supervised learning you give examples it learns from example Unsupervised learning you just give it lots of data it figures out anomalies patterns, etc And in reinforcement learning it does actions it get rewards and it gets better and better and the new thing That's come up which chat GPT has done is self supervised learning where it learns with the own data It learns in the way chat GPT learns is you take a sentence It tries to predict the next word in the sentence and by doing Predicting the next word and learning from that it gets feedback it gets learns from it It's so it's a form of self supervised learning. We're just creating all this magic So AI this AI is very different and given this AI You know obviously everything is changing if you look at image recognition or if you look at speech recognition Or if you look at games if you look at dress cancer diagnosis in each of these areas AI is matching and exceeding human performance in a broad range of tasks And we feel like we are very very close to that place where a Generalized machine can start to match and exceed human performance in not just in narrow tasks, but in broad range of tasks Now I gave you one of the three equations There are two more equations that you have to think of if you have to drive great advantage using AI Number one you we saw error equation data into computing to techniques What's the second one if you want to drive great results with AI? Just algorithms is not good enough You have great algorithms that can drive great results But you need solid engineering because if you have to manage large vast amounts of data That's sitting in various parts of the world you have to bring all of them to the problem So it requires you to bring in significant engineering to bring the data to the problem And then you have to automate decision making which means that you have to in milliseconds in microseconds You have to make decisions happen. So you need engineering skills in that too So AI plus engineering and then you need design to understand exactly What is the problem to solve and how to drive adoption one of the things that chat GPT has also done is design if you see one of the reasons it has really taken off is because The it is a very well behaved bot. It gives you very well structured a nice English It knows exactly what problems to answer what it doesn't want. It is confident It is good. It is it has great language is great manners And that is also because the design has has incorporated those elements into it So if you bring in AI engineering and design, you can drive great results with AI and the last one if you especially You're building a culture of analytics or AI inside your organization You need solid talent a culture of experimentation and a governance model because AI is not necessarily Foolproof AI has still lots of bugs as you've seen with chat GPT Also, it makes lots of mistakes and some of them very confidently So you need a right kind of governance mechanism so that you can drive Responsible AI and then also make sure that the models are running and not and not deteriorating over time So if you use these three equations, you can drive great AI advantage I'll give you one example of what fractal has done in our business called cure dot AI Now we incubated cure dot AI a few years ago And the problem was to see how to bring down Errors in radiology if you go to a radiologist today, you get 23 percent error rates in finding You know any abnormalities in in an x-ray. So we built a System by which you can take an x-ray and it detects 30 different abnormalities in x-ray And this came in very handy. We built it very quickly during COVID We actually evolved that model to work to identify COVID and this was as early as March 2020 In fact in February 2020, we had the closing party for cure and then I asked the CEO of cure Prashant Hey, why don't we build a COVID detection algorithm and literally within four weeks? We built a COVID detection algorithm by March 2020 COVID took off and even as late as early as March 2020 the hospitals in Italy were using The the cure dot AI algorithm to find out if patient is progressively getting worse So they would look at daily differential x-rays every day you look at x-ray and you look at what is the difference between yesterday's x-ray and today's x-ray Most doctors are not able to see the difference because x-rays change very little day by day But an algorithm is able to spot small little differences in those x-rays and then using that they were able to find out Which which patients are getting worse which patients will which patients will not make it or will have to go to the ICU and so on So you can use AI not just to build you know build products new You know, you know do things like chat GPT You can actually use them in real-world applications where you can bring in deep science and deep tech together, right? You can Not just this there are countless other problems where AI is now making fundamental breakthroughs and deep science think of protein folding How does the three-dimensional structure of protein is something that AI has recently cracked So even though chat GPT was a seminal moment for 2022 in November 2022 The bigger moment in in in my world in case of AI was what deep mind did in July on July 31st, 2022 They released the three-dimensional structures of 200 million proteins Just imagine if you want to solve any major problem in the world any health care problem in the world You have to decode the structure of the protein that underlying that disease or the virus and so on and that protein structure is is a hard problem So deep mind using incredibly good deep reinforcement learning actually figured out the the three-dimensional structure of 200 million proteins and remember before that Figuring out the structure of a protein meant one PhD one PhD working full-time would develop the three-dimensional structure of one protein And they'd solve that problem by solving it for 200 million proteins at once So deep science and AI are beginning to merge So just to conclude a AI is a paradigm shift in computing it will alter the way we work live and play It will and AI is not going to replace human beings But human beings that use AI are going to replace human beings that don't use AI and finally AI is going to not just make a Difference to deep tech is going to make a difference to deep science and the way we live. Thank you