<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Rajesh Kumar S A on Medium]]></title>
        <description><![CDATA[Stories by Rajesh Kumar S A on Medium]]></description>
        <link>https://medium.com/@business-of-tech?source=rss-f393fb52be53------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*UaDFuh5aqxzKVrz4.jpg</url>
            <title>Stories by Rajesh Kumar S A on Medium</title>
            <link>https://medium.com/@business-of-tech?source=rss-f393fb52be53------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 21:55:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@business-of-tech/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Tokens Are the New Bits]]></title>
            <link>https://business-of-tech.medium.com/tokens-are-the-new-bits-33a139e575b7?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/33a139e575b7</guid>
            <category><![CDATA[inference]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[performance-optimization]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llmops]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 09:51:33 GMT</pubDate>
            <atom:updated>2026-08-31T09:51:33.407Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Post 1 of the “New Physics of Computing” series</em></p><p>In 1948, Claude Shannon gave the world a unit of measurement for information: the bit. For seventy years, everything in computing was priced, engineered, and fought over in bits. Bandwidth was bits per second. Storage was bits per dollar. Moore’s Law was, at its heart, a promise about how cheaply you could move and store bits.</p><p>That era isn’t over. But it has a successor.</p><blockquote>Today, the unit that decides who wins and who loses in AI is the <strong>token</strong>.</blockquote><h3>What’s a token, anyway?</h3><p>If you’ve ever used a chatbot, you’ve watched tokens stream past without knowing it. A token is a small chunk of text — sometimes a whole word, sometimes a piece of one. “Cricket” might be one token. “Bengaluru” might be two or three. Every question you ask gets chopped into tokens, and every answer gets generated one token at a time, like beads pulled off a string.</p><blockquote><strong>every single token costs something real</strong></blockquote><p>Electricity. GPU time. Cooling. Money. When a model writes you a 500-word answer, somewhere a rack of processors did trillions of arithmetic operations, and someone paid a power bill for it.</p><p>In the bit era, we asked: <em>how fast can you move data?</em> In the token era, we ask: <em>how cheaply can you produce intelligence?</em></p><h3>The new game: squeeze everything</h3><p>When bits were the currency, the heroes were the people who compressed files, designed faster modems, and packed more transistors onto silicon. The token era has its own version of this obsession, and it runs on two fronts.</p><h3>Front one: training</h3><p>Training a large model is like teaching a student by making them read a library — except the library has trillions of tokens and the tuition bill runs into millions of dollars. So the game becomes:</p><ul><li><strong>Feed better tokens, not just more tokens.</strong> A model trained on carefully curated data can outperform a bigger model trained on internet sludge. Data quality is the new data quantity.</li><li><strong>Waste nothing.</strong> Every GPU-hour where chips sit idle waiting for data is money on fire. Teams obsess over keeping thousands of processors fed simultaneously, the way a kitchen obsesses over never letting a burner go cold.</li><li><strong>Learn more per token.</strong> The frontier isn’t just bigger models — it’s models that extract more understanding from every token they see. Better curricula, smarter training recipes, architectures that spend compute only where it matters.</li></ul><h3>Front two: inference</h3><p>Inference is the other half. While training happens once; inference happens every time anyone, anywhere, asks the model anything. Billions of times a day.</p><p>Serving a model efficiently is a genuinely beautiful engineering problem:</p><ul><li><strong>Answering has two phases.</strong> First the model <em>reads</em> your whole prompt (fast, parallel, like skimming a page in one glance). Then it <em>writes</em> the answer one token at a time (slow, sequential, like typing with one finger). These two phases stress the hardware in completely different ways, so clever systems handle them separately.</li><li><strong>Memory is the bottleneck, not math.</strong> Modern GPUs can do math absurdly fast. The struggle is shuttling data in and out of memory quickly enough. The model’s “short-term memory” of your conversation — its cache — can eat more space than the model itself. Managing that cache well is worth fortunes.</li><li><strong>Batching is everything.</strong> One user’s question doesn’t fill a GPU. A hundred users’ questions, cleverly interleaved, might. The art is serving everyone at once without anyone noticing they’re sharing.</li></ul><p>Each of these tricks compounds. A 2x here, a 1.5x there, and suddenly one company serves the same model at a fifth of a competitor’s cost. In a business where the product <em>is</em> tokens, that’s not an optimization. That’s the whole ballgame.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MT-qD1OVRvIur8WO2uWZKQ.png" /></figure><h3>There’s a Moore’s Law for tokens — and it’s faster called super-moore’s law or llmflation</h3><p>Here’s the proof that all this squeezing is working, and working absurdly well.</p><p>Moore’s Law promised that computing power would double roughly every two years. It held for half a century and built the modern world. But look at what tokens are doing.</p><p>In mid-2020, generating a million tokens with the best available model cost around $30. Today, comparable or better capability costs a few <em>cents</em>. That’s a several-hundred-fold collapse in about six years. Moore’s Law over the same stretch would have delivered maybe 6x. Analysts have started giving the phenomenon its own names — Andreessen Horowitz calls it “LLMflation” and pegs the decline at roughly 10x per year for equivalent capability; some researchers have tracked curves steeper once they stack pricing and efficiency gains together.</p><p>Think about what that means. A dollar spent on intelligence next year buys roughly what ten dollars buy today. No technology in history has deflated this fast.</p><p>But — and this is the twist that makes the token era genuinely different from the bit era — <strong>cheaper tokens don’t mean cheaper answers.</strong></p><p>Economists call it the Jevons paradox: when something gets cheaper, we don’t spend less on it, we use wildly more of it. Cheaper coal in the 1800s meant <em>more</em> coal burned, not less. And the newest AI models lean into this hard. Reasoning models “think” before they answer — burning tens or hundreds of times more tokens per question than their predecessors, deliberately, because thinking longer produces better answers. The price of a token collapses while the number of tokens per answer explodes.</p><p>So the scoreboard is shifting. The metric that decides winners isn’t really cost per token anymore. It’s <strong>cost per answer</strong> — how cheaply can you deliver a <em>correct, complete result</em>, thinking included. A lab that halves its token price but triples its tokens-per-answer has gone backwards. The one that bends both curves at once runs away with the market.</p><p>For last few years people with data had all the fun. Now its people with core engineering who will have fun. Its a repeat of 90s and early 2000s of internet era — apache2 and nginx were getting built with browsers playing their part in squeezing out the performance. Will keep posting!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=33a139e575b7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Intelligence is the service — Laying the Foundation]]></title>
            <link>https://business-of-tech.medium.com/intelligence-is-the-service-laying-the-foundation-a5c42eaf9f5b?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5c42eaf9f5b</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[challenge]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Fri, 14 Jun 2024 09:14:51 GMT</pubDate>
            <atom:updated>2024-06-14T15:06:16.961Z</atom:updated>
            <content:encoded><![CDATA[<h3>Intelligence is the service — Laying the Foundation</h3><p>Heading an AI business is arguably the coolest and most exhilarating job today. That being said, when you are in the thick of action its easy to get lost in the details. Zooming back and seeing the bigger picture puts a lot of things in perspective. Here is my take on the AI revolution — parallels in the past, where we are currently with AI, what are the gaps and where I believe we are heading as an industry.</p><p>In terms of defining an era of computing, the AI era intriguingly combines elements of the PC era, where enhanced computing was accessible to everyone, with the Internet era, which democratized access to vast amounts of knowledge and computational resources. Coming from the Internet era myself, I will try to draw parallels to identify recurring patterns that could help techpreneurs, programmers and product managers build innovative solutions.</p><blockquote>Dialing back in time</blockquote><p>The concept of ‘X’aaS, or ‘anything as a Service’, only emerged in the last few decades, driven by the internet. From computing power offered by AWS to the convenience of hailing a cab via Uber. Everything was shared. Rewinding back two more decades and you will hit upon the mother of all inventions, internet. Of course, when it started, it demonstrated a simple document sharing capability. One was able to read a document which was present in a server in California while sipping tea in London. <strong><em>Miraculous!</em></strong></p><p>The fundamental idea of internet was that a computer had an address that uniquely identified it across the world. Once identified, there was a port through which we could pass information. Addresses! Once the addresses were set, a name was tied to that address and then of course you had currency of measurement of usage in bytes per seconds transferred from one computer to another. This infrastructure which was simple yet powerful was absolutely free. <strong><em>Unbelievable!</em></strong></p><p>For the first decade (80s), internet’s most profound application was mail. It was not until another innovation literally exploded the usage. Browsers (90s). The moment when netscape navigator came about, internet was never the same. With the advent of html, a markup language to present information in a nicer cleaner colourful structured way, people started using it to open up a lot more information. And then, server side scripting changed the landscape completely. People were able to run applications by using compute remotely! The following 2 decades (2000s and 2010s) were about building applications to more real world scenarios.</p><p>Given the above context, I would argue that AI is now in the same place as that of internet of 80s or PC in 70s. The first ever evidence of general purpose intelligence is demonstrated. You have a system that can apply intelligence to any task. <strong><em>Amazing!</em></strong></p><blockquote>Now lets bring out the differences between AI era and the previous two eras</blockquote><p>Unlike the internet — Firstly, AI has not been entirely free. In the internet era, the largest network was free. Smaller networks were paid (VPNs). But, in AI era, largest model is paid but smaller models are free. This is where I believe its following a trend similar to PC era. Powerful PCs are costly even today. Smaller compute nodes are almost free.</p><p>Secondly, there is a no protocol like HTTP / TCP for AI. In case of internet, http tells the network where to route the information to and what to do with that information. The closest AI has today is “prompt engineering”. This is almost like a programming language to an operating system but the same programming language behaves differently on different OS. Prompts are not portable across models! And are not consistent all the time. So, this highlights a fundamental challenge in the field of AI: the lack of a universal standard for how models interpret and respond to inputs. Just as HTTP and TCP standardized the transmission of data over the internet, enabling seamless communication across various systems, AI desperately needs a similar protocol for prompts. Currently, the effectiveness of prompt engineering varies not just from one model to another but also from one use case to another, leading to a lack of predictability and portability. This inconsistency can be likened to speaking the same language in different countries and being understood in varying degrees. Until there is a universal standardization, AI’s potential will be hindered by its unpredictability and the steep learning curve required to tailor prompts effectively for different AI models.</p><p>Thirdly, there is no standard user interface for applications. PC era had GUI elements defined even in the smartphones. The internet era adopted universal standards like HTML for structuring web pages and CSS for styling them, AI does not have a standard Human Computer Interface yet. Chatbot is not an ideal HCI, its like a pin hole to view a very large beautiful painting! Here, AI era is neither like PC era or Internet era. Each AI application — from chatbots to more complex decision-making systems — still tends to have a unique interface, which can be confusing and inefficient for users. The development of a standardized HCI in AI would revolutionize how we interact with these technologies, potentially making AI applications more accessible, predictable, and user-friendly.</p><p>In a nutshell, AI is a game changer. It has ushered in a new era of computing. An era as big as internet or PC or both combined. I firmly believe that we have not seen true AI applications yet barring the chatGPT-esque ones. And the reason we do not see widespread applications yet is that we are still seeing the <strong><em>foundational layers being built and the middleware (plumbing) infrastructure being set up</em></strong>. There is still a need for some standardisation on protocols (prompts) and browser-like clients (better HCI).</p><p>One has to remember that en-masse user adoption of technology has happened only after 3 things are solved — when tech gives consistent solution to real world problems, when there are no security threats or unintended consequences and finally when there is an easy to use interface. Now that foundations are laid, more experimentation has to happen at the application layer which means more practical applications have to be built to come up with a generic repeatable winning formula of model + infra + HCI.</p><p>In the next blog <a href="https://medium.com/@business-of-tech/intelligence-is-the-service-building-momentum-fe71b35ea5fb">https://medium.com/@business-of-tech/intelligence-is-the-service-building-momentum-fe71b35ea5fb</a>, I am going to talk about how we are building momentum in using AI and the challenges therein</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5c42eaf9f5b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Intelligence is the service — Building Momentum]]></title>
            <link>https://business-of-tech.medium.com/intelligence-is-the-service-building-momentum-fe71b35ea5fb?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/fe71b35ea5fb</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[challenge]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Fri, 14 Jun 2024 09:14:07 GMT</pubDate>
            <atom:updated>2024-06-14T15:09:18.822Z</atom:updated>
            <content:encoded><![CDATA[<h3>Intelligence is the service — Building Momentum</h3><p>In the previous <a href="https://business-of-tech.medium.com/intelligence-is-the-service-laying-the-foundation-a5c42eaf9f5b">blog</a>, I covered the road to AI with some stark differences to earlier tech revolutions. In this, I will go deeper into why such differences have occurred and what are some of the current challenges to make AI mass market.</p><p>Lets take the 2 tracks — PC development and Internet Development. PC was becoming smaller with smartphones and IoT while internet was adding more data and expanding to every corner of the world. So, PCs optimised compute while internet optimised data transfer. Interestingly, the winners in both era were different — PC world ended up being operating system centric with 3–4 operating systems windows, ubuntu, android, macOS with no room for other players. OS is the most important software. OEM integration virtually sealed the door for smaller players. So, the secondary winners were app developers on these platforms. In case of the internet, it became a lot more democratic with room for everyone to grow if they were good. Winners were those who drove more traffic (data) than others like google, meta, amazon, netflix etc. but there was room for lot more. So, we have a lot more small time winners. All of it on a single application — “browser”.</p><blockquote>More compute gave better PC apps, more data gave better Internet apps.</blockquote><p>AI needs a mix of both — large compute and large data. And interestingly, it combines compute and data so efficiently that you dont know whether to treat it as an operating system or as a data/knowledge source. And as of today, its neither reliable as an OS nor as a knowledge source. Moreover, this combination results in high-level cognitive processes like reasoning, summarising etc.. Hence, its difficult to understand how it can be definitively used. (more on this below)</p><p>Thinking from a PC era perspective, once the training of a neural network based AI model is done, you get an <strong><em>AI model file</em></strong> that acts like a circuit board but the circuit is created using data. Metaphorically speaking of course, neural networks are software constructs that learn from data, whereas circuits are hardware designed to carry out defined electronic functions. So, in this context, larger the circuit board, larger is the capability to solve problems.</p><p>Thinking from Internet era perspective, once the training of the model is done, the model file is like a compressed form of all of the data used to train it. So, larger the model file, more information is compressed. You can query it to get you any answer.</p><p>PC era folks want to make an AI chip and burn that model file as a circuit. Internet era folks want to build larger models and give away as APIs or have a chip in the cloud. So, you have larger models consolidating to small set of companies while lot of smaller models are being built by everyone.</p><p>This explains Apple’s and Microsoft’s dilemma of having local and cloud. But security (as they say) cannot be the reason because if you are connecting to the cloud to solve “complex” problems, you have essentially built a compound wall with a hole. Its either fully local or fully remote with security being handled irrespectively. The real reason I believe is that they are hedging their bets. They are unsure if the cloud model (internet era) works or chip model (PC era) works. So, they are keeping the options open.</p><blockquote>Application! Application! Application!</blockquote><p>While larger trends are a little orthogonal in nature, there is momentum, people are building more apps, trying with newer use cases and perfecting the tech for it to be applied nicely. At application level, there are more challenges primarily because we are dealing with a class of algorithms that programmers and product managers have not dealt with before.</p><blockquote>Probabilities. The difference between deterministic and stochastic models</blockquote><p>Majority of computer programmers have always dealt with deterministic systems. Stochastic systems have a probability of outcome which needs to be managed. This is a paradigm shift. The programmers and product managers must embrace it and so should the end users / corporations.</p><blockquote>Use cases. Hammer searching for nails ?</blockquote><p>Like it or not, AI is a new tool. This means that usage of this tool needs to be understood. The way our human brain is wired, according to me, is that as soon as we get a new tool, we tend to use it in various ways to understand the limits of the tool. (Its like test driving a car). Once we know that this particular “hammer” is good for a “certain nail”. We will use it for that. AI is in that exploration phase wherein we are trying to find what is the right “nail” and how big the hammer needs to be.</p><blockquote>Hallucination, Bias and Cost. Still a problem</blockquote><p>The 3 main issues that have persisted over many years now. If we consider that the current AI architecture mimics human brain, then hallucination and bias is going to stay. We hallucinate in that we will not give the exact same answer to the same question every single time and we all have our own biases. So, we have to live with it and make use of the tool (AI) to solve our problems. The other problem is cost. Inference costs are quite high even today. But there are lots of optimisation coming along including at the chip level. So, it should become affordable pretty soon.</p><blockquote>AI is still a feature — genie in the bottle</blockquote><p>This is perhaps a big issue. So far, barring chatGPT or Siri or Gemini, we have not seen an AI only product. And we have definitely not seen a AI only enterprise product. Everyone is releasing AI as a feature inside their products. I think this has to break. The moment the application stack of AI is finalised — lets say with GPT as an OS, certain prompts as a standard for a task and a standardised HCI — we should see an explosion in pure AI apps. But the AI as a feature is adding a lot of momentum. In a weird way, when PC era got transitioned to internet era, we had a “thin” client (with fewer features) on the cloud and then a “thick” client (with full features) on desktop. The dawn of AI era has been with a “thin” client (with simple features) on desktop and a “thick” client (with ability to solve complex task) on the cloud.</p><blockquote>Collaboration is missing at application layer</blockquote><p>While people are opening the model weights, we are still unable to get en-masse collaboration in terms of content or training data. We are unable to get everyone to build their own apps and make them available for wider audience. The biggest success of internet era and of PC era was that there were industry accepted standards / protocols and there were simple discovery platforms. AI is missing this badly. AI is too dependent on a few large players. To be truly disruptive, we need ability for people to build different models and expose it with a set protocol. HuggingFace has tried to solve this. Kudos to them. But its still missing a middleware that can check which is the best model to choose and apply for particular task. Or to that extent, prompt standardisation is not there across models.</p><p>While there are lots of problems, they are surmountable. There is a strong belief among developers / product managers to build AI applications and use AI in our daily lives. The power of AI is too alluring to not get attracted to it. This momentum will continue.</p><p>But what is the definition of excellence in the AI era ? Where is this momentum leading us to ? Tried to answer them in the next part of this blog series. <a href="https://medium.com/@business-of-tech/intelligence-is-the-service-defining-excellence-3190e96e6be6">https://medium.com/@business-of-tech/intelligence-is-the-service-defining-excellence-3190e96e6be6</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fe71b35ea5fb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Intelligence is the service — Defining Excellence]]></title>
            <link>https://business-of-tech.medium.com/intelligence-is-the-service-defining-excellence-3190e96e6be6?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/3190e96e6be6</guid>
            <category><![CDATA[future]]></category>
            <category><![CDATA[challenge]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Fri, 14 Jun 2024 09:13:16 GMT</pubDate>
            <atom:updated>2024-06-14T15:13:10.183Z</atom:updated>
            <content:encoded><![CDATA[<h3>Intelligence is the service — Defining Excellence</h3><p>As part 3 and the last of this series of <a href="https://business-of-tech.medium.com/intelligence-is-the-service-building-momentum-fe71b35ea5fb">blogs</a>, I would like to define what excellence in the field of AI would really mean (in my own way).</p><p>Lets ask ourselves, what is intelligence ? According to me, its the ability to convey information, to express the most complex workings of nature, to perform complex actions, to have an ability to achieve a goal via different means / methods.</p><p>If such is a definition of intelligence, how do we charaterise levels of intelligence ? What do we consider more intelligent than others ? Again, according to me, its the ability to convey maximum information with <em>minimum words</em>. Its to express the most complex workings of nature in the <em>simplest of terms</em>. To perform complex actions in a <em>simple and effective way</em>. To have an ability to achieve a goal via methods that are <em>easy and repeatable</em>.</p><p>So, this means there is a measure of intelligence. Thanks to the brilliant minds who have worked on this so far, we have a unit of measurement and that is tokens/second. I believe that tokens per second can be a measure of intelligence, not just a metric for computing the cost of using LLMs. Read on to know how...</p><p>There is a famous quote in management, “Measurement is the first step that leads to control and eventually to improvement”. So, how does tokens/ second measurement actually improve intelligence ?</p><p>As usual, lets look at some parallels in the PC era and internet era.</p><p>Those who have taken computer science interviews 20 years can easily recognise this. Every solid interview had a time complexity and space complexity question. Every algorithm had to be analysed based on these parameters. Why ? because a good application would have the best space and time complexity. Because, the best PC apps were supposed to take less CPU and less memory.</p><p>The internet era had bandwidth as an important measurement. Apps &amp; websites that loaded faster were deemed to be better. It was imperative to optimise the data transfer across the globe. Also, excellence was not in making too many features and overwhelming the user but about simplicity of UI. Minimal amount of data transfer to ensure work is done with of course minimal number of steps. So, this resulted in web apps becoming better in data handling and user experience. Workflows became simpler and users were able get their work done faster.</p><p>So, to summarise, programmers measured the amount of CPU / Memory utilisation in PC / Desktop / Server apps to make them better and so they became extra-ordinarily better from the time they started. Web and app developers measured bandwidth utilisation and user time spent which progressively optimised them exponentially.</p><p>Coming to AI. Token usage is a metric that all the LLM companies measure and charge. So, any application that can solve complex problems with minimal token usage would be deemed to be of exceptional quality.</p><p>Optimising token usage can be done by having smaller fine tuned model for specific tasks or can be based on RAG or by just making the model so “intelligent” that it needs very little prompting to perform the task. At least, this would be my definition of excellence when it comes to AI applications.</p><p>In Hindi, there is a saying <strong><em>“Samajhdar ko ishara hi kafi hain”. The wise only need a hint.</em></strong> It essentially says that a person who is wise or intelligent can understand or grasp information from very subtle indications or minimal details. I guess its apt for AI applications to be measured and rated by this.</p><p>There are some AI / ML systems that have already reached the “excellence”. In case of recommendation systems, there are systems that take minimal input from users to optimise their experience and give great outcomes. When it comes to generative AI, we are still witnessing the phase 1 and 2 which is laying of foundations and building initial applications. Once there is consensus on the stack and the interface, then there will be race towards excellence.</p><p>Given the metric of measurement is tokens per second, I believe AI era will tend more towards how internet era shaped. There will be standardisation in prompting and user interfaces. There will be more complex models in the cloud accessible to consumers from any device. Although, they might not be free. Thats only because the bandwidth constraint will be minimal. In order to support such large usage, there will be optimisations at the chip level that can give maximum tokens per second. No wonder some companies are raising hundreds of billions of dollars for such things.</p><p>In conclusion to this 3 part series, I believe the AI era is different and its shaping up in its own way. Like every new era of computing, the AI era will bring in a significant shift for everyone. For programmers, its going to change the way we build apps, for users the way they consume content, for workers the way they do their day-to-day job and for children the way they learn. Intelligence inherently brings in more efficiency and simplicity. Intelligence is a service that everyone will need to embrace; one might call it artificial, but its effects are undeniably real.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3190e96e6be6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Of MVPs, Captains, Coaches and Owners]]></title>
            <link>https://business-of-tech.medium.com/of-mvps-captains-coaches-and-owners-eea859db37c?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/eea859db37c</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[management]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Tue, 09 Jan 2024 11:54:48 GMT</pubDate>
            <atom:updated>2024-01-09T11:54:48.116Z</atom:updated>
            <content:encoded><![CDATA[<p>Most of us who follow sports can relate to the terms mentioned in the title. There are always most-valued-players in a team, their captains guided by bunch of coaches and finally the owners who fund the team. Now, this is no different in any other team as well. Be in software development, research or ops.</p><p>As an MVP, they need to be most valued team member. Bring in an extraordinary skill or work ethic that just differentiates them from rest of the pack. Everyone should aspire to be one.</p><p>As a captain, you need to lead a team to complete a mission. Leading teams means operating the team to their strengths and weaknesses. Ultimately making the team to win the mission it was set out to do.</p><p>As a coach, you are expected to improve your team members. This means, identifying weakness in team members and improving them. Making them do their best work. Coaching is about setting the team for success. They do not play in the field but they would have channeled the energies of the team members to perform at the most optimal level.</p><p>As an owner, you are expected to make the coach successful. Be an enabler, make judgement calls on funding or on business. Ensure the team operates without any fear and shackles.</p><p>A team can win games if MVP is firing well, can win few matches in a row if captain is inspiring, can win a tournament if coach is really good but can win multiple championships if it has a winning combination of all. From ownership to players.</p><p>If one looks at sports be it football or cricket, we can easily draw parallels and examples. I would not get into cricket, cause there is too much passion associated with individuals. The debate will get polarised missing the fundamental points.</p><p>In football, we can clearly see this. Zidane for example was the Most-Valuable-Player of his team. He became a captain of the famous French team and the Galacticos. Was very successful in that too. He was a coach for Real Madrid and coached them to 3 consecutive champions league trophies. In tech industry too, if we look at Bill Gates, he was a lead coder initially, he lead the teams, hired the right next level and invested in right bets. These are exceptions. If you study the pattern, many successful coaches were actually average players. Many captains were not really MVPs of the their team. So what gives ?</p><p>I believe these are specific skills in themselves. Hard technical skill initially, leadership skill, coaching skill and finally business skill.</p><p>One can be exceptional coder but not a team lead. One can lead teams very well but not become a manager to grow the team. One can manage the team and grow it but not know how much and when to invest / divest. Combination of all this in one individual can be deadly, but excellence in each or multiple areas is not bad either. While the world tends to push people in this hierarchy, each one of us should find our sweet spot and excel.</p><p>One should aspire to be the best at each level. But, to the best at any level, one should have knowledge of other levels too. Excellence is always appreciated at any level.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eea859db37c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Secret Invasion is happening already]]></title>
            <link>https://business-of-tech.medium.com/secret-invasion-is-happening-already-7a398f58d083?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/7a398f58d083</guid>
            <category><![CDATA[generative-ai]]></category>
            <category><![CDATA[marvel]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Sat, 22 Jul 2023 18:26:52 GMT</pubDate>
            <atom:updated>2023-07-22T18:26:52.644Z</atom:updated>
            <content:encoded><![CDATA[<p>Many of us would have started to watch the latest Marvel Web Series — Secret Invasion. I don’t know if I was the only to see this, but, the negative outcomes of the latest generative AI techniques can be easily visualised in this series. It is eerie and way too real.</p><p><strong><em>Spoiler Alert (next 2 paragraphs) </em></strong>— The series is all about Shape Shifters — Skrulls. They can assume anyone’s form and work the scene for their advantage. While skrulls have been allies with humans for a long period of time, a bunch of them go rogue and decide to destroy humanity. &lt;<em>that escalated quickly</em>&gt;</p><p>Let me cite a few correlations now — in the series, people show CCTV footage of Nick Fury shooting Maria Hill when the real Nick Fury was just a bystander and the shape shifted Nick Fury just carries on his act. Rest of the world saw the CCTV footage and started accusing the real Nick Fury. In another instance, there is a case where a missile launch is averted by a phone call of a military person who was impersonated. There are so many such instances showcasing impersonation and manipulation of audio visual sources. In all of this, the real person would not have a chance to prove their innocence since the proliferation of information is way too fast and inorganic.</p><blockquote>Here is the issue, all of this is possible even today.</blockquote><p>The Tesla sales call or Google Duo booking an appointment are a few publicly showcased examples of AI impersonating humans and effectively passing the test as a human. There are reports of fake Whatsapp video calls convincing people to pay money. If fake news was a threat 4 years ago, fake people are going to be 10X that threat.</p><p>In the series (Secret Invasion), they call it the greatest threat to humanity because one cannot identify the shape shifter. The problem is similar. In the next few months, there will be no way to distinguish between real videos / audio vs fake ones. There is no way to know if you are dealing with real humans or skrulls. Now, I am not a naysayer of AI but watching this series just makes the threat very real.</p><p>Some people mentioned that AI is equal to nuclear power and when such comparisons were made, I would tag them as exaggeration. But looking closely at exponential progress that AI in the recent past, AI’s power is immense. Since there is no physical form for AI, it is hard for people to estimate the magnitude of its prowess. Only visualisations showcased in sci-fi movies give an idea of this power.</p><p>While Google, OpenAI and others have decided to watermark the outputs generated by AI, there is no guarantee that other companies will do such a thing. OpenAI forming SuperAlignment team to address AI going rogue etc. is also great. But, how can we ensure this from everyone who is developing AI models ?</p><p>Not sure if Hollywood is getting influenced by Silicon Valley or the other way around. But, them being so close to each other makes fiction and fact more blurry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/692/1*1ObQrXZrAm0iSktise2FCA.png" /><figcaption>generated via canva</figcaption></figure><p>Skrulls are ultimate deepfakes!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7a398f58d083" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A New Operating System!]]></title>
            <link>https://business-of-tech.medium.com/a-new-operating-system-dad67060a47f?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/dad67060a47f</guid>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Thu, 15 Jun 2023 20:12:51 GMT</pubDate>
            <atom:updated>2023-06-25T08:50:26.010Z</atom:updated>
            <content:encoded><![CDATA[<p>There is new era of computing thanks to this new system. No, I am not talking about spatial computing! I am talking about the other tech that has gone wild in recent past.</p><blockquote>Obviously LLMs!</blockquote><p>I am going to dig into fundamentals of computer engineering for this. <em>(bear with me</em>). We humans always liked to get tasks done from machines. As computer engineers, we wrote code, which basically meant, we instructed a computer to do a task. There is a fascinating history about how we have evolved in the art of “asking computers to do a task”.</p><blockquote>Lets do a little history check before we dial back to present day…</blockquote><p>World War II accelerated our advancement in so many areas that it saddens that we need wars to push ourselves in science. But thats how world works I suppose. Alan Turing famously built the first programmable computer along with his friend in 1945, that effectively won the war for Allied forces. (oh.. there is a <em>history-is-written-by-winners</em> kinda theory where apparently a German engineer had built a computer in his garage in 1941 itself). Whoever built it, they set a revolution in motion.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tGPtrT3exvC4fNA1UVoFTg.jpeg" /><figcaption>First ever programmable computer source : <a href="https://www.theverge.com/">www.theverge.com/</a></figcaption></figure><p>Once humans knew how to program a computer to do a task for us. We used it for everything. But we started with industrial use cases. Computer engineers writing massive instruction sets in assembly languages and ALGOL etc. Come to think of it, I had written assembly code to run on an 8085 chip in college! Perhaps the best way to learn and appreciate today’s ease (I am not that old, but my college was really really old that had all these chipsets lying around). While programmers were having fun with silicon chips, Steve Jobs thought, maybe we should have everyone use a computer! Great idea! The era of personal computers was about to begin. But, you cannot teach everyone assembly code language…</p><p>Along came<strong> GUI</strong> (for the younger readers, this means Graphical User Interface) Revolutionary in the 70s. Made it possible for every person to “instruct their personal computer to perform a task”. While programmers were still coding in C, they released applications in an Operating System. This OS was now taking instructions from a keyboard, mouse, joystick etc.. and communicating back through a screen. Every such end to end operation was packaged as an application. Applications were just a bunch of programs which were in turn a bunch of instruction sets that were supposed to work together.</p><p>Applications need code and an OS. Development of applications needed coding (lots of it). So, application ecosystem had 3 components that came together and advancement in each had exponential benefits on other component : operating system, UI and programming</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0bVvUzrsYo9kbNADSnNCIQ.jpeg" /><figcaption>source : www.g2.com</figcaption></figure><p>As computing power became cheaper and more efficient, desktops turned to laptops. Laptops needed mobility and connectivity to access larger centrally maintained data and to communicate with each other. Thankfully, networking and communication had generational shifts in parallel which brought about the concept of <strong><em>internet</em></strong>. Laptops shrunk further to become smart phones. Making everyone to have a computer in their pocket. Internet ensured data access, marking a new era of application programming. Connectivity became a standard requirement.</p><p>Operating systems evolved from MacOS / Windows to Android / iOS. UI went from simple desktops to smartphones to wearables. Programming languages also went through a similar journey that brings us to current point of discussion.</p><p>The goal of computer programmers has always been to make it easier for next generation of programmers to write code easily. We started writing instructions sets in binary — 0s and 1s. Then moved to assembly code. Then moved to C with compilers. From then on, we went with high level languages. Abstractions to hide the bare 0s and 1s. Making the programming languages simpler and simpler and simpler… Until…</p><blockquote>Natural Language is now our computer programming language!</blockquote><p>We have achieved it. We are able to instruct a computer to do a task in our language. People have started to realise the potential. There seems to be a perception of intelligence where a system is easily able to understand our language. So, we have all started to exploit this “intelligence”. This is the disruption that will change the way we use computers.</p><p>This new programming language can now create very powerful applications. Applications that uses LLM as an operating system. We can probably start considering LLMs as processing systems. And human-computer-interface might not be UI anymore. It can be voice or display or other sensory outputs like even touch.</p><p>Innovators and problem solvers of the world are getting a grasp of this. There are fundamental questions of why do we need UI, why do we need a tool that requires a series of clicks ? For example : We can probably just say “file my taxes” just like how we talk to a tax consultant. Or have detailed conversations on some research topic. Or create art works that are from our imagination instead of understanding how GIMP or photoshop tool works.</p><p>We will see a phase where tools are going to be reimagined, intelligence is going to get augmented, applications are going to give us amazing experiences. Our next generation will feel like how we felt when we saw people using cassettes and tape recorders.</p><blockquote><strong>PCs made GUI as a standard requirement for every application. Internet made connectivity as a standard requirement for every application. Generative AI is making intelligence as a standard for every application.</strong></blockquote><p>A new era! Exciting times!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dad67060a47f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why should robots have all the fun ? [Part 1]]]></title>
            <link>https://business-of-tech.medium.com/why-should-robots-have-all-the-fun-part-1-c46153a77735?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/c46153a77735</guid>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Tue, 14 Mar 2023 20:18:13 GMT</pubDate>
            <atom:updated>2023-03-14T20:18:13.336Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why should robots have all the fun ? [Part 1]</h3><p>You need to step out of the generative AI frenzy to look at where the world is heading towards. When you do that, you will realise AI is just taking over the joyful exploration that humans thrived at. Do we still have fun in creating and discovering new things or are robots having all the fun?</p><p>When you take a step back, you see the bigger picture. Looking at the bigger picture, you can trace a path with which you can probably project the trajectory. Hindsight is always 20–20. I will look at this picture from a point-of-view of 3 distinct domains where humans engaged in joyful exploration— arts, discovering new lands and scientific research.</p><p>Let’s get some context. So far, robots have been automating mundane things. Factory floor automations, UIPath kind of automations and largely tasks that does not involve creativity. AI folks have been trying to crack the creative space for quite some time. I have seen the progression first hand in the last one decade (partly aiding it albeit in a very small way). But, what has happened in the last couple of years is nothing short of remarkable. A tipping point of sorts thats going to have far reaching ramifications to our way of life.</p><p>Part 1 — Arts (pictures, poems, music &amp; philosophy)</p><p>Digitisation — Past three decades were all about digitisation and making it available over a network. A process that is now almost 100% complete across the world. What started with mail, went on to news, books, music, movies, commerce and every aspect of our daily lives. Obviously with all this digital data, there was explosion in “accessible” digital content. When I say “accessible”, I mean in terms of the amount of content that is available to any person on earth. Obviously this created a problem, hence Search &amp; Discovery algorithms became more and more advanced. Majority of the scientific applications were around choosing the right content for the right set of users. Machine Learning and AI played a heavy hand in this. Past decade accentuated the problem further with content creation platforms such as tiktok and youtube exploding the creator base. So, AI algorithms became more advanced. Finding patterns to understand users more quickly and accurately.</p><p>People were having fun in creating and discovering new videos, songs, poems, stories, pics, paintings. Now, with the new era of generative AI, AI is creating the content. Does this just take away all the fun? People roamed the world, did adventures, read previous works — to find inspiration, to fuel their creativity. But, the next generation of humans do not need to anything. AI has already understood our needs, now it will create what we want and display it to us. Will that be real ? Generated by some other human or completely fictional ? Case in point :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3x9qo1g3O5e4-uyPGrEONg.png" /><figcaption>Sunflower in New York Streets — Generated by Dall-E</figcaption></figure><p>Some say, generative AI is a new tool. Like painting transitioning to photography. Photography might transition to generative AI in computer vision.</p><p>Generative AI will allow everyone of us to express ourselves and create more content. But this also means the skill of painting or writing is going to wane down. People will be publishing lot more generated content than original materials. But, as I mentioned previously, we do not need to type what we want in a prompt. Recommendation systems are already aware of what we like, we might as well get what we want. .</p><blockquote>Are we on the precipice of a cultural revolution ?</blockquote><p>The phenomenon is not limited to pictures obviously. Its more pronounced in news or blogs or textual content. Literature. Another case in point :</p><p>Output for “write a small poem on sunrise with an undertone of life &amp; death philosophy”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/772/1*ou7h1zBDbw9t42P3q0wvdA.png" /><figcaption>obviously I had to use chatGPT</figcaption></figure><p>The poem is objectively beautifully composed. So, now how do I have any motivation to search for more. Maybe there was one more poem that someone has written but due to the speed at which I could get this, I wont search anything beyond.</p><p>Coming to music. Generating music is supposedly tough for many people. But generative music is here and its has become a rage already. A simple search online gives so many options to create generative music! According boomy.com, their users have already create 12% of all songs present in the world today. You don’t need any skill for this, all you need is to say what sort of a mood, tempo, genre, instrument and theme you want the music for. Thats it. So, basically, you don’t need to have any skill to produce your own art.</p><p>Finally, let’s talk philosophy! Yes, AI and philosophy. In some of my interactions with chatGPT, I found its replies to be quite profound in its comprehension and reasoning. If I am not looking at these replies with a context that its generated by a machine, I will believe the other side to be extremely thoughtful and hold them in very high regard. Probably ask for life advices, resorting to it during tough times because “hey, this fellow is indeed knowledgeable and reasonable”. An example of this</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WkHn5Rdlr0wTjdRa8kYBZA.png" /><figcaption>chatGPT</figcaption></figure><p>For the uninitiated, in Hinduism, the entire Bhagavat Gita is a Q&amp;A between Arjuna and Lord Krishna. Arjuna asked very tough existential questions and Krishna gave answers that are amazingly relevant even to this day. In one of the most quoted statements of the Bhagavat Gita, Krishna says</p><blockquote>Karmanye vadhikaraste Ma Phaleshu Kadachana,<br>Ma Karmaphalaheturbhurma Te Sangostvakarmani</blockquote><blockquote>Translating to</blockquote><blockquote>You have the right to work only but never to its fruits.<br>Let not the fruits of action be your motive, nor let your attachment be to inaction.</blockquote><p>Now, read the highlighted part of chatGPT’s response. It is scary and amazing at the same time. Did these large language models copy or generate the sequence in the most logical sense ? We would not know. But, philosophy is one of the basic subjects of humanities. People spend years together to read, meditate and internalise some of the ancient wisdom to preach and help people in distress. Its fun for people, passion and life’s mission. We seem to have a machine becoming better at it already. This is just a third version of GPT models. To give a comparison, google what Windows 3.0 was in 1990 to what it is today.</p><p>In an eerie Hitchhiker’s reference — answer to the ultimate question of life, universe and everything is Generative AI (our version of 42) but like in H2G2 one has to figure out the right question!</p><p>In conclusion, I am going to put a few questions (cause thats the new thing — becoming good in asking questions, prompt engineering)</p><p>Are robots having all the fun by creating pictures, poems, music and answering deep philosophical questions ? Considering they are good at it, why not push it further ? Are we going to spectate and marvel at this creation ? Will this degenerate our little grey cells up top ?</p><p>I will leave this part with one more problem to ponder over, for those who grew up with googling everything, how will you search or discover something that has not been created yet ?</p><p>P.S: This was not created by any large language models.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c46153a77735" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Analyze Analyze Analyze — AI models]]></title>
            <link>https://business-of-tech.medium.com/analyze-analyze-analyze-ai-models-595cba7ef89d?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/595cba7ef89d</guid>
            <category><![CDATA[advanced]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Wed, 02 Mar 2022 11:37:08 GMT</pubDate>
            <atom:updated>2022-03-02T15:23:52.139Z</atom:updated>
            <content:encoded><![CDATA[<h3>Analyze Analyze Analyze — AI models</h3><p>AI development has accelerated rapidly in the last 3 to 4 years. Back in 2017, very few knew what AI or deep learning specifically can do. What sort of data to collect ? what sort of model architecture to choose ? There was a fundamental doubt on the capability of this technology</p><p>Fast track to 2022, every business leader is cognizant of AI. Most of them have set up data science teams and most of them have the basic data necessary to kick start AI model development. Now, they are able to realize that AI model development is more complicated. Building a robust AI model that handles as many cases as possible — is a much more complicated problem.</p><p>We at Streamoid, faced very similar challenge few years ago. We were able to get the AI use cases very early and were also able to collect lot of data. So, we encountered the problem of building robust model very early on. The only way to solve this problem was to analyze the models. Understand where they were failing and correct them. More analysis leads to better models.</p><p><strong><em>Analyze the edge cases</em></strong> — You can know more about the model on how it handles edge cases, unseen data and wrong data (data it is not meant to process) than how it handles normal cases. These edge cases are more powerful indicators of the model than normal data. So, we built a feedback loop that integrated to the end application and captured all those cases where the prediction was wrong. Where the model failed. This hard negative mining helped us capture new edge cases and build more robust models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mk0KXRLnZ7qrsM9bx71LNg.png" /><figcaption>Fetches All Wrong Predictions from Prod</figcaption></figure><p><strong><em>Analyze a model using traditional statistics</em></strong> — The traditional statistics are still the bedrock of modern AI techniques. Everything can be explained using statistics. Typically, when the problem does not fit a model statistically, it will not produce a robust model. So, by analyzing the data, models can be fine tuned. At Streamoid, we used TSNE plots, standard confusion matrices, PR scores and along with that violin plots and other out-of-distribution analyses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*GgUw9pFznpusq88ousn_3g.png" /></figure><p><strong><em>Analyze the behavior of the models</em></strong> — AI models are like any other algorithms, we need to understand the boundaries in which they operate well. Some of the models might be based on deep learning hence its harder to understand. But, understanding the nature of the algorithm would be critical to its application. There are techniques like TracIn that explain model’s output based on the training data. Many visualization techniques that allow AI developers to understand bias. Many standard datasets that the AI model can be benchmarked against. For internal use cases, at Streamoid, we standardized the evaluation datasets and also integrated some of the visualization techniques in a no-code tool such that anyone can use it and understand the models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*FO7bqJd3g10FMMoPjgSNTA.png" /></figure><p>AI Studio, the tool we built at Streamoid, helps in analyzing and explaining a model’s action in the most visual way. As a data scientist, you are supposed to constantly analyze your models. Treat your models like any other software wherein every wrong prediction is a bug. Only with this, AI will be successful in production.</p><p>Contact Streamoid Sales via malini@streamoid.ai or ashutosh@streamoid.ai for early access to the AI Studio tool.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=595cba7ef89d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to manage data science team — 2]]></title>
            <link>https://business-of-tech.medium.com/how-to-manage-data-science-team-2-99557e1a7475?source=rss-f393fb52be53------2</link>
            <guid isPermaLink="false">https://medium.com/p/99557e1a7475</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rajesh Kumar S A]]></dc:creator>
            <pubDate>Wed, 02 Mar 2022 07:21:20 GMT</pubDate>
            <atom:updated>2022-03-02T15:20:03.898Z</atom:updated>
            <content:encoded><![CDATA[<h3>How to manage data science team — 2</h3><p>Continuing with the previous post where I mentioned various types of data science teams and different challenges associated with them, this post is going to be on managing a data science team that is a hybrid of domain experts and data scientists. Also, with businesses adopting AI, I am going to talk about how product managers and leadership team can be on the same page to track the effectiveness of AI.</p><p>When it comes to managing such teams, communication becomes key and everyone should have visibility to everything. As in, domain expert needs to know how her input data is getting converted to AI model. Data scientist needs to analyze the models and clearly understand what type of data is needed to improve the model to perform well in real world scenarios. Product manager needs to know how to use the AI in an application to improve some core metrics. Leadership team needs to ensure AI investments are paying off by checking its effectiveness in real world. All of this has to happen on a daily basis.</p><p>So, we came up with a tool that can help everyone. Most critical feature for each stake holder.</p><p>Data Scientist —<strong> <em>visualization of data</em></strong><em>.</em> For a data scientist, visualizing the data is key to making progress. So, we had to first organize the data in a proper way and visualize the same. Next, we had to automate some sample scripts that allow data scientists to easily analyze the data whenever they want.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*GgUw9pFznpusq88ousn_3g.png" /><figcaption>Violin, TSNE, Confusion Matrix, PR Scores — all automated</figcaption></figure><p>Data Sciences Manager — <strong><em>Experiment tracking</em></strong> and measuring the improvement. When multiple people are working on different projects, its important to ensure they are tracked properly. Also, managers need to make sure AI models go to production and testing team’s responses are captured systematically.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JVKVRSJnVR40tNqSFvAOVA.png" /><figcaption>Simple Experiment Tracker</figcaption></figure><p>Domain Experts — This is special case where AI is disrupting a certain domain like fashion or medicine or finance. In this case, experts need explainable AI where models’ predictions and models’ output is explained well enough for experts to accept and integrate to their workflow. <strong><em>Explainable </em></strong>and well visualized data will help domain experts to give guidance to data scientists to fine tune the models and data labelers to collect the right data</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*RpjmHGav2w5w0WQJ519jjA.png" /></figure><p>Data Labelers — While labeling a simple object like car or tree is easy, there are cases where deep domain knowledge is necessary for data labeling. Data labelers also need to know “<strong><em>what data to source</em></strong>”. Data sourcing has to be done based models’ performance.</p><p>Product Manager — <strong><em>Application deployment</em></strong>. When one has a use case for AI, its important for them to convey the same openly to the team. The important message would be “what is the problem you are trying to solve” ? So, there needs to be an option in the tool wherein product managers can clearly explain the use case and domain experts / data scientists alike can always refer back and correct the models</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*Yu76iIQ4NwpAOw43d93xeA.png" /></figure><p>Business Leader — <strong><em>What is the ROI?</em></strong> If the product manager has mentioned that a particular AI model will automate a certain flow or it increases a certain metric like add-to-cart in ecommerce, the business owner will need these metrics at any given instance of time. So, real-time analysis has been built into the system with a concept of “robustness” to give more confidence to the business that the model’s performance has been great and the same can be expected in the future due to the robustness index.</p><p>Conceptually, while the tool is built to give smooth ML Ops and an easy way to develop new AI models, the biggest benefit it offers is that it helps in capturing the edge case behavior of the model. The feedback loop is instrumented to get negative samples from real world. Also called hard negative mining. Analyzing the negatives or where the model has failed explains the model in its truest form. You can understand a lot about a model based on how it behaves in edge cases than normal cases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*jQJTifpR-Z9oYOu6S5GjZg.png" /></figure><p>Streamoid’s AI Studio is a tool that handles all the stake holders and brings them to a single page. It has dramatically increased the efficiency of the data science team in Streamoid. People with little to no experience in data science have been able to understand AI models and build them. Till date close to 3000 models have been built analyzed and many of them deployed in live production.</p><p>Contact Streamoid Sales via malini@streamoid.ai or ashutosh@streamoid.ai for early access to the tool. Quite a possibility that it might make managing data sciences teams a lot easier than you think.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99557e1a7475" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>