<?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 Ishi Singhal on Medium]]></title>
        <description><![CDATA[Stories by Ishi Singhal on Medium]]></description>
        <link>https://medium.com/@ishisinghal?source=rss-e33006a3e2b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*rWgKgPAkBZNcDjOA</url>
            <title>Stories by Ishi Singhal on Medium</title>
            <link>https://medium.com/@ishisinghal?source=rss-e33006a3e2b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 05:17:42 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ishisinghal/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[The Protocols making your Agent your Shopper]]></title>
            <link>https://medium.com/@ishisinghal/the-protocols-making-your-agent-your-shopper-901ef23b9015?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/901ef23b9015</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Thu, 26 Mar 2026 08:34:07 GMT</pubDate>
            <atom:updated>2026-03-26T08:34:52.736Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Z7leG8avOublgVUj" /></figure><p>On 18th March, Tempo and Stripe announced the famous Machine Payments Protocol (MPP), a protocol that enables agents to make payments and check out on your behalf. However, this is not the first big step in agentic commerce. This industry has seen a lot of effort and progress from both tech companies and also fintechs, everyone knows they cannot afford to lose out. Soon, large amounts of money will be moving through these agents and this race will decide who gets to cash in on this shift.</p><p>This short read explores the journey before MPP — several protocols from Google, ACP from OpenAI and Stripe and x402 from Coinbase which leads us to understanding MPP and how it is different.</p><h4>Google’s Stack</h4><h4><strong>A2A</strong></h4><p>The first protocol to understand here is Agent2Agent Protocol (A2A). Although this is not strictly an agentic commerce protocol, this is what enables agents to talk to each other which is the prerequisite for them to make payments to each other.</p><p>Why do we need a protocol for this? Imagine that you are building a travel agent and want to use someone’s flight booking agent. Without knowing how the flight agent works, how would you communicate with it? In fact, how would you even know what services that agent offers? Even if you could figure all this out, how will authentication and security be ensured during communication? A2A takes care of all these matters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/1*hCovaL8Z70SdmqSWfAYeZA.png" /><figcaption>Source — <a href="https://a2a-protocol.org/latest/topics/what-is-a2a/#a2a-and-mcp">https://a2a-protocol.org/latest/topics/what-is-a2a/#a2a-and-mcp</a></figcaption></figure><p>How does it do this? For discovery and authentication, your flight booking agent will publish an “Agent Card”. Google describes it as a “digital business card”, it has all details regarding what the agent can do and how to communicate with it safely. For the actual communication, it sends cryptographic JSON messages over HTTP.</p><p>Now that our agent can discover the services of the flight booking agent and knows how to talk to it, how will it make payments for these services? Enter, Agent Payments Protocol, AP2.</p><h4><strong>AP2</strong></h4><p>When your AI agent starts going and making payments on your behalf, several problems arise. Here’s the 3 key problems Google identifies:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*ngCgyqikpoAuy-bf3Ls7Kw.png" /><figcaption>Source — <a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol">Google</a></figcaption></figure><p>AP2 solves these problems through mandates. These are like permission slips from the user. Here’s an example of how a mandate would be generated in two different flows -</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/565/1*z3dZpgquY84iMJfzh0OM8A.png" /></figure><p>This works across payment rails so that you can pay via card, PayPal or even crypto.</p><h4><strong>UCP</strong></h4><p>A2A informs your agent what services are available and how to communicate with the other agent and AP2 provides the infra for payments. But when it comes to actually placing the order, that’s where Universal Commerce Protocol (UCP) comes in. This protocol has many big names associated: Shopify, Walmart, Target, Etsy and many other giants came together with Google to build this. <br>UCP enables the agent to transact with any merchant. To do so, it defines catalogs, checkout sessions, line items, status, etc. In fact, it is defined to even capture nuances like loyalty points and contextual discounts. <br>One caveat: It is only built for Google’s ecosystem — AI search, Gemini, etc. More on this later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/788/1*4HRvj-o36Tf63ewnwWwP6g.png" /><figcaption>Source — <a href="https://developers.googleblog.com/under-the-hood-universal-commerce-protocol-ucp/">Google</a></figcaption></figure><h4><strong>Agentic Commerce Protocol by OpenAI and Stripe</strong></h4><p>ACP is the direct competitor of UCP, it essentially serves the same purpose. For now, one major difference is that ACP can be adopted by any platform whereas UCP only serves Google’s platforms. ACP has been implemented in ChatGPT’s Instant Checkout already.</p><h4><strong>x402 by Coinbase</strong></h4><p>This protocol solves a slightly different problem. Rather than shopping for your boots (which it can also be used for), it’s more suited for microtransactions and services that required APIs traditionally. For example, imagine you made an agent and you need access to the articles from a journalism website. Traditionally, you would need to set up an API key and make an account. x402, on the other hand, enables you to use their services without having to do all this work.</p><p>So what does the payment look like? It leverages the age-old HTTP 402 “Payment Required” status code. All the HTTP clients in the world know how to read it. This makes for a rather simple and elegant solution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L9qJJ8TzBQuCxI_9PAzyJw.png" /><figcaption>Source — <a href="https://www.x402.org/">https://www.x402.org/</a></figcaption></figure><p>And the integration for a merchant is also very simple.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/789/1*Yp7diHnZLyJzBPIYvBYA9g.png" /></figure><p>However, x402 is primarily focused on crypto. Why? Crypto serves micropayments much better than other payment rails. It’s instant and it does not charge-per-transaction. Imagine you want tokens from OpenRouter and your card payment takes a day to go through and charges you the regular fee for a 3 cent transaction.</p><p>In the last month alone, 75 million transactions have been processed by x402.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/755/1*qS9-1khjPVSprVcRrSvKmA.png" /></figure><h4><strong>MPP</strong></h4><p>Stripe launched MPP with a partner: Tempo. This should give you a good guess of which protocol it competes with (Tempo is a blockchain designed for payment through Stablecoin).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/596/1*KIfWYzamQnM6B5z3DlKM9Q.png" /><figcaption>Source — <a href="https://x.com/davewardonline/status/2035144713703928309?s=48">X</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*4oi7e41T1T54qXxQlI1syw.png" /></figure><p>One key difference apart from openness is that while x402 does pay-per-request, MPP allows you to allow a certain spending limit post which microtransactions can happen without interference till you hit that limit. So if your agent is doing hundreds of transactions in a short period of time, MPP probably works better.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/621/1*TAsdP6T6DyAbwqWiHGoR7Q.png" /><figcaption>Source — <a href="http://www.rootdata.com/news/584290">http://www.rootdata.com/news/584290</a></figcaption></figure><p>Also, MPP has embraced fiat currencies relatively more and point that out as a key difference too, but I’m sure it’s a matter of time before x402 catches up.</p><p>Stripe has chosen to work with and support both x402 and MPP. Looks like even the biggest fintech can’t predict who’ll be the winner here ;)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=901ef23b9015" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Model to Agent]]></title>
            <link>https://medium.com/@ishisinghal/from-model-to-agent-a56834e6aa97?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/a56834e6aa97</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Sun, 15 Mar 2026 06:30:45 GMT</pubDate>
            <atom:updated>2026-03-15T06:30:45.844Z</atom:updated>
            <content:encoded><![CDATA[<h4>Harnesses and how they shape performance</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yPb3WODwqWJF_kqiZo5TxA.png" /><figcaption><a href="https://blog.langchain.com/improving-deep-agents-with-harness-engineering/">Source</a></figcaption></figure><h4><strong>What is an Agent Harness?</strong></h4><p>A lot of coding agents use the same models today, but show a massive difference in performance. How is this possible? Because of the harness each one builds. The harness is what decides what goes into the model input, how many tokens does the model use, what tools does it call, what does it do if it encounters an error, basically everything you would need for a model to actually execute actions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7QpVMWjxedxA30qPmncNVg.png" /></figure><p>When a developer uses a Claude Code SDK, they are essentially using their harness and customising it to build their own agent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aw2iUjDytvxIApTrT69dyw.png" /><figcaption><a href="https://www.vtrivedy.com/posts/claude-code-sdk-haas-harness-as-a-service/">Source</a></figcaption></figure><h4><strong>Deep Dive into Claude Code’s Harness</strong></h4><p>Harnesses have numerous components which change from agent to agent. For example, Cursor’s harness focuses heavily on customisation for each model that the user opts for. So when you use GPT versus Opus on Cursor, the harness is completely different. Here, I cover the details of tool calling and context management, specifically with respect to Claude Code.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*oof9fLNdopBaydZYEI3N3w.png" /></figure><h4><strong>Tool Calling</strong></h4><p>This is what gives the agent the ability to execute any action, from editing your files to buying your shoes online. Claude Code divides their primary tools into the following general categories:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wi4d58HRIhxIGr7uXBKfGw.png" /><figcaption>Source — <a href="https://code.claude.com/docs/en/how-claude-code-works">Claude Code Docs</a></figcaption></figure><p>Here is how these tool calls would work when you tell Claude Code to “Fix the tests that are failing”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tOkTzBhVFrS-v_puWJkaOA.png" /></figure><h4>What is Bash?</h4><p>Bash is probably the most frequently used tool by Claude Code. To understand more about it, we have to first understand what a shell is.</p><p>A shell is a program that enables you to communicate with your computer’s OS in text rather than your standard visual interface. For example, if you want to rename 500 files, you could select each one and rename it — which is an extremely inefficient way to do it. Or, you could write a shell command that does this for you.</p><p>Bash (Bourne Again Shell) is just the most popular shell, others being zsh, Fish and PowerShell.</p><p>Why is this a perfect tool for agents?</p><ul><li>Firstly, the Text In, Text Out (TITO) model of Bash perfectly aligns with the nature of LLMs. Not only can the model give instructions in text, but bash also generates a text output describing the action and result. So if a test case fails, Bash can relay this to the agent which can then decide the next steps.</li><li>Had Claude Code not supported Bash, they would have to write custom integrations for many kinds of tools, one for running test cases, one for checking git status, etc. Though this is probably not a big deal for Anthropic, too many tools flood their context and reduce performance. Bash, on the other hand, can enable all these commands through one tool. It basically gives access to all developer CLIs like git, npm, docker, python, etc.</li></ul><p><strong>Memory and Context</strong></p><p>All your interactions with Claude are stored in context windows — the files you’ve uploaded, all your chats, etc. The size of this window however, is limited. This means that once it is filled up, Claude starts forgetting some of the stuff you would want it to remember. For example, in the beginning of the chat you told it to add a comment in each line but since that instruction is no longer in the context window, it stops following it. <br>This is perhaps the biggest challenge with long running agents and Anthropic came up with four main features that help overcome this (Slides are generated from <a href="https://www.kerns.ai/home">Kerns</a>)</p><p>Firstly, Claude summarises all the content in your context window as soon as it starts filling up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XoI4ROfCJ3mRioFSDLLspA.png" /></figure><p>It also divides the task into sub-tasks that are performed by sub-agents. This means that each sub-agent only has the relevant information in its context window, making it a more efficient architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/852/1*bxFceWfpht-nx7ddn6ZdMg.png" /></figure><p>Now you probably have some instructions that you want to be followed across sessions. For example, you only want the agent to use pytest. Having to give this instruction each time you start a new session would be a terrible experience. Hence, Claude Code stores this context in files like Claude.md and Memory.md</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/826/1*bxdso8ltEc35g3KDxgFYfg.png" /></figure><p>Lastly, Anthropic has also engineered its tool call to be the most optimal. So when you tell it to fix the “authentication” bug, it doesn’t scan the whole codebase. Rather it only fetches the lines relevant to the “authentication” function.</p><p>For this, they use GrepTool. This is different from previous approaches that used to index each file so that it could be searched up in the future. Instead, grep does not index anything and only runs a search on the actual files to fetch the most relevant data.</p><p>Link to anthropic blog in case you want to take a deeper dive —</p><p><a href="https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents">Effective harnesses for long-running agents</a></p><p>A link to the article that prompted me to read more about harnesses -</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/Hxlfed14/status/2028116431876116660&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/18ebcb3fa24d7066f5cba2e0dc236bab/href">https://medium.com/media/18ebcb3fa24d7066f5cba2e0dc236bab/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a56834e6aa97" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evaluating Agents]]></title>
            <link>https://medium.com/@ishisinghal/evaluating-agents-25c2d890df06?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/25c2d890df06</guid>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Mon, 02 Mar 2026 16:21:33 GMT</pubDate>
            <atom:updated>2026-03-02T16:21:33.057Z</atom:updated>
            <content:encoded><![CDATA[<h4>Bringing trust and reliability to non-deterministic systems</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i8oPparlpn8KzS4JBX1dmQ.png" /><figcaption>Source — <a href="https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents">Anthropic</a></figcaption></figure><h4><strong>What are evals?</strong></h4><p>An eval (short for evaluations) performs the same function as the name suggests: to test whether the AI system generated a satisfactory output.</p><h4><strong>How evals have evolved over time</strong></h4><p>Benchmark tests and RLHF were basically evals for LLMs like ChatGPT. This was relatively simpler as you were just grading one output. Did the model solve the math problem? Did the model spew out any toxic content? These questions are more straightforward and hence easier to test.</p><p>Compare this with an agent, for example a travel planning and booking bot. The agent is carrying out so many steps here, from understanding your prompt to pulling the right hotel and flight data, and maybe even making the payment on your behalf. This is an extremely complex flow with multiple possible points of failure. One wrong step also compounds in the entire flow, if the agent did not understand your preferences in the first step itself, there is no way it can produce the desired outcome. Consider the following scenarios -</p><ul><li>The agent might not orchestrate the flow well. Maybe it booked the flights without even confirming whether a hotel was available</li><li>It might lose context mid-conversation and displays options that are over your mentioned budget</li><li>API calls might have errors — wrong dates, destinations, etc. Similarly, tool calls are often not accurate</li><li>The agent might tell you in the chat that it has made the reservation but did not actually implement the payment step</li></ul><p>These are just a few examples which, in best cases might negatively affect user experience and in the worst, generate a completely wrong outcome like exceeding the budget.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o3gJEcEVkPLtw5hXVfAqeA.png" /><figcaption>Source — <a href="https://cloud.google.com/blog/topics/developers-practitioners/agent-factory-recap-a-deep-dive-into-agent-evaluation-practical-tooling-and-multi-agent-systems">Google</a></figcaption></figure><p><strong>So why is it so tough to generate automatic tests to judge the agent’s performance?</strong></p><ul><li>Agents are non-deterministic. Give it the same prompt twice and it will give different responses each time.</li><li>Further, the context, prompt, required tool calling, and the expected action are vastly different with an infinite number of possibilities.</li><li>They also have numerous steps until they reach the final output. Without the detailed logs of each step, it is impossible to diagnose a failure.</li><li>The final output is also not the sole north-star metric. For example, an agent processed the user’s refund, showing a satisfactory output, but it never retrieved the order number of the user.</li></ul><p>The evals have to be generated such that the agent never fails. An agent that works 99% of the time will still do 100 wrong bookings for every 10,000 users, this is not at all a desirable outcome.</p><h4><strong>How do we do evals today?</strong></h4><p>Most evals today are written by humans who made the agent. They spend a lot of time crafting these in order to try and cover every possible scenario where it might fail. However, some teams have started using agents to generate some evals as well, for example — an agent might keep making test cases from each time your code failed in production.</p><p>LLM-as-a-judge is the closest approach today to automating writing evals. This approach leverages a different LLM from your agent to evaluate its output. It tests whether the agent gave the desired outcome, spewed out any toxic content, operated within its constraints, etc. However, this is again a non-deterministic test and hence cannot be relied on completely.</p><p>𝜏<strong>-Bench</strong></p><p>Tau-bench is one of the most popular ways to judge whether the agent is successfully executing the task. It is a benchmark created by Sierra AI Research and adopted by the likes of Anthropic for agents. Most other benchmarks do semantic assessment (did the agent say the right thing, did it hallucinate, etc), but 𝜏-Bench focuses on whether the desired change was made in the database.</p><p>For example, instead of testing whether the agent told you that it made your flight booking, it checks in the database if the booking has actually been made and if all the parameters are correct.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gcvyltdgLdZ9wtkIFeeAMw.png" /><figcaption>Source — <a href="https://sierra.ai/uk/blog/benchmarking-ai-agents">Sierra</a></figcaption></figure><p>For agents to be deployed on a large scale and across industries, they need to be reliable, which they are not by their nature itself. This means that one of the biggest bottlenecks to be solved in the coming years would be eval generation. What the ideal future would look like is agents generating these themselves, making it faster and more comprehensive and resulting in truly trustworthy systems.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25c2d890df06" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Evolution of Data Labelling: Sama, Scale AI, Surge AI and Mercor]]></title>
            <link>https://medium.com/@ishisinghal/the-evolution-of-data-labelling-sama-scale-ai-surge-ai-and-mercor-8a8d69514336?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a8d69514336</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[meta]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Sat, 03 Jan 2026 20:21:22 GMT</pubDate>
            <atom:updated>2026-01-04T04:06:42.767Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/178/1*OzYxX00JVkuh0XMPjIg9Yg@2x.jpeg" /></figure><p>10 years ago, data labelling used to look very different from today. It consisted of low-cost, repetitive tasks like labelling each pixel in an image and doing sentiment tagging, which have now been replaced. It has also undergone a massive change in terms of market dynamics in 2025 when Meta acquired Scale AI for $14B. To add to this, regulations only seem to get more stringent with time. In this short read, I hope to cover the fundamentals of this industry and how it has evolved over time.</p><h4>What are typical data labelling tasks today?</h4><ul><li>Reinforcement Learning from Human Feedback (RLHF) — The space has moved towards tasks that require more expertise than before. For example, choosing which output response is better and why, trying to make the model behave undesirably to find vulnerabilities, etc.</li><li>Synthetic Data Curation — Experts write what an ideal response should look like and this data is used for training the model. It could also be programmatic generation of a dataset rather than using a human expert.</li></ul><h4>Which players are entrusted with these tasks?</h4><p>Traditional players like Sama, Appen and Cloudfactory used to be the default options for labs like OpenAI in the low-skill era. These companies would onboard data labellers from countries like Venezuela and Africa where labour was much cheaper and perform simple tasks at scale. When the requirements shifted, they could not keep up. Appen went under financial distress and subsequently got delisted. Sama saw severe public backlash and lawsuits. An investigation by TIME magazine revealed that the workers hired to carry out the data labelling were mentally scarred from the disturbing content they had to filter out. This resulted in the cancellation of their contracts with OpenAI. Other legacy platforms like Lionbridge are trying to adapt but have not seen much success.</p><p><a href="https://techcrunch.com/2023/07/14/workers-that-made-chatgpt-less-harmful-ask-lawmakers-to-stem-alleged-exploitation-by-big-tech/">Workers that made ChatGPT less harmful ask lawmakers to stem alleged exploitation by Big Tech | TechCrunch</a></p><p>To bridge this gap, players like Scale AI and Surge AI emerged. They reportedly saw revenues of $870M and $1B last year respectively. The two would fiercely compete to land the same clients. Scale AI showed dominance in creating a government presence and secured a $249M contract from the US Department of Defense. However, in June 2025, Meta acquired Scale AI. This raised a huge concern for labs like Google Deepmind and Open AI. Their demand for more neutral platforms gave birth to Mercor.</p><p><a href="https://www.reuters.com/business/scale-ais-bigger-rival-surge-ai-seeks-up-1-billion-capital-raise-sources-say-2025-07-01/">Exclusive: Scale AI&#39;s bigger rival Surge AI seeks up to $1 billion capital raise, sources say</a></p><p>Mercor started as a platform for AI hiring and then pivoted into connecting experts to AI labs for data labelling. Scale AI’s acquisition could not have come at a better time for them. They saw their valuation go from $2B in February 2025 to $10B in October 2025, A 5X jump in 8 months.</p><p>Labelbox, now a public company, has a different positioning. Rather than providing services, they provide a software platform for data management and annotation. Hence, they operate on a SaaS model unlike the other players. They also have much higher margins: 65% on an ARR of $180M in Q4 FY25. The platform also enables Reinforcement Learning with Verifiable Rewards (RLVR), an approach expected to complement RLHF in the future.</p><h4>How is this space seeing regulations unfold?</h4><p>The EU AI Act has imposed stringent controls on the AI industry as a whole. One aspect of these regulations is the accountability for training datasets which would cause labelling vendors like Scale AI to ensure much more quality control and documentation. In the case of non-compliance, the fines are heavy, reaching up to 7% of global revenue. Regulatory pressure and public scrutiny are also pushing these companies to ensure more ethical practices, especially in regards to the labellers.</p><p>Data, being arguably the most important differentiator in the AI race, will become an important moat for frontier labs. This results in a huge market being created for players like Scale AI. Will these companies also find themselves to become irrelevant like Sama in the coming years? Or will they define the development of AI? While no one knows what the future holds, it is definitely going to turn out to be a fascinating story.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a8d69514336" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Guide to the Voice AI Stack]]></title>
            <link>https://medium.com/@ishisinghal/a-guide-to-the-voice-ai-stack-569de2bf6b4d?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/569de2bf6b4d</guid>
            <category><![CDATA[voice-ai]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Tue, 30 Dec 2025 14:42:20 GMT</pubDate>
            <atom:updated>2025-12-31T07:45:50.273Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8WzUtuUXg8fJNoc1" /></figure><p>Voice AI has seen dramatic changes in the last couple of years — we went from Siri and Alexa to agents delivering actual economic value. There have been instances where AI has been successfully solving more than 95% of customer support queries, acting as testaments to the great potential this technology holds. In order to understand it better, it is important to first understand the layers that go into a voice AI system. This article maps that landscape, right from voice model provider ElevenLabs to application layers built on top.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Aj7jPquSueZcm1Jf.jpg" /><figcaption>Source — <a href="https://www.bvp.com/atlas/roadmap-voice-ai">Bessemer Venture Partners</a></figcaption></figure><h4><strong>How AI Listens</strong></h4><p>Let us start from the beginning. When you talk to an AI chatbot, it first converts whatever you say to text since it is more easily processable than data in the form of audio. This step is called Speech-to-Text (STT) or Automatic Speech Recognition (ASR). These models need to have high accuracy even in the presence of noisy signals, accents, etc. In order to provide a seamless experience, they also need to have very low latency.</p><p>OpenAI Whisper, Deepgram, AWS Transcribe, etc are amongst the well known ASRs. Whisper is open source and trained on 680,000 hours of multilingual audio and interestingly, about a third of this data is non-English.</p><p>Next comes the LLM that actually processes this text and generates a response. This could be any of GPT, Gemini, Claude, etc.</p><h4><strong>Enable Human Voice for AI</strong></h4><p>Once the response has been generated, it needs to be communicated to the user through voice, the job of the Text-to-Speech (TTS) layer. Here, the most important factor is how natural the speech sounds. It needs to feel like a human is on the other side of the conversation, which of course brings latency into the picture too.</p><p>ElevenLabs has been, for long, the default TTS for many. However, emerging players like Cartesia have been challenging this position. Check out the link below to read more about their model and architecture.</p><p><a href="https://medium.com/@ishisinghal/what-comes-after-transformers-the-100m-bet-on-ssms-bcbd7014a49b">What comes after Transformers: The $100M bet on SSMs</a></p><p>There also exist other players like PlayHT, Google, OpenAI, etc.</p><h4><strong>An Alternate Approach</strong></h4><p>While the three layers of STT, LLM and TTS have traditionally been used, there also exist end-to-end models like Gemini Flash Native Audio. This means that instead of converting the data to text, it would process the audio itself. These models, while looking attractive due to low latency, have not yet reached scalability. Cascading (STT, LLM and TTS) provides you with better cost control, tooling and customisations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/0*G4Xw53M8Kqze7lKI" /></figure><h4><strong>Runtime Layer</strong></h4><p>Now when you’re deploying a voice model, you have STT, LLM, TTS, telephony (the infrastructure that allows you to actually make those phone calls), external tool calls, memory, and other layers that you have to stitch together. This is not a trivial task — imagine a human conversation, when we interrupt, it should first get detected, the TTS stream has to be paused, STT needs to be resumed, and this text needs to be processed keeping the context in mind. This is where tools like Vapi, Retell AI, Voiceflow, etc come in. They do all these integrations and ensure that everything happens with very low latency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uSffmMzxklqZcInwpjX6Sg.png" /><figcaption><a href="https://vapi.ai/">Vapi</a> has completed more than 150M calls, enabling the adoption of voice AI</figcaption></figure><p>Developers can also pick and choose the runtime layer they want to adopt depending on the level of control they would like to have. Below is a quick read on different players in this space and what they offer.</p><p><a href="https://medium.com/@ggarciabernardo/understanding-voice-ai-development-paradigms-7bda6550780c">Understanding Voice AI Development Paradigms</a></p><p>Next comes the layer that allows users to design, customise and control the voice AI agent. Voiceflow helps teams visually design conversational flow, LangChain can provide toolings like RAG, Zapier can enable the agent to trigger workflows, etc.</p><h4><strong>Creating Economic Value</strong></h4><p>Finally, at the top of the stack sits the application layer. Today, a lot of these tools are built specifically for a function or an industry. Consider the case of an AI receptionist at a hospital that needs to identify emergencies, comply with numerous regulations and also show human emotion. Such nuances are difficult to capture in a generic voice bot. Abridge, having raised $300M for medical conversation intelligence, being an instance of this hypothesis.</p><p>On the other hand, platforms like SoundHound which are industry agnostic and offer agents for multiple functions propose something closer to general purpose AI.</p><p>With the rapid pace of innovation, it won’t be long before we start seeing voice AI in many applications beyond customer service. As architectures and models become increasingly capable, it will enable more horizontal players to emerge and get deployed across functions and industries. Most importantly, it will become the de facto way of humans conversing with machines — soon there might come a future when the keyboard, mouse and screen become obsolete — and players like ElevenLabs, Vapi, Abridge, etc are leading this transformation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=569de2bf6b4d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[World Models: Prediction to Imagination]]></title>
            <link>https://medium.com/@ishisinghal/world-models-prediction-to-imagination-f3a95d8c8227?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3a95d8c8227</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[world-models]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Mon, 22 Dec 2025 07:30:06 GMT</pubDate>
            <atom:updated>2025-12-25T06:06:22.134Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*h9kQLkz6GXx1IRpk" /></figure><p>With Yann LeCun just announcing his venture into world models - raising €500M at a €5B valuation - I thought it might be interesting to cover in a quick read what this new paradigm of technology is all about.</p><p>The first time I read about world models it seemed like an abstract notion, a model that can imagine the world and understand its dynamics. So let me try to approach this with an example — when you ask an AI tool to recognise an image or answer your question, it is doing prediction. Even if it is generating an image, it is only predicting what the next pixel will look like.</p><p>Contrast this with a self-driving car. The first difference — it is taking each action on its own and not just responding. Secondly and most importantly, for each brake, each turn and each acceleration, it has to simulate the impact it would have on the external environment before it takes the action. A robot, while picking up a cup of coffee, has to simulate which position would result in the most stability. While playing a game, the model has to imagine numerous chains of steps and decide the most optimised one.</p><p>What if we were using Reinforcement Learning to solve this game? The agent would take step A, get feedback, take step B, get feedback, so on and so forth. This means you need a lot of interactions to train this robot. It also means that when you tweak this game — suppose you make the chess board 10x10 — the robot has no idea what to do now since it has not been trained for this. On the other hand, a world model can simulate these actions without actually taking them and hence would not require many interactions to train. It can also now generalise its learnings and play on this terrible chess board.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ScPGEFzwYgPGiFif.jpg" /><figcaption><a href="https://www.xunhuang.me/blogs/world_model.html">Source</a></figcaption></figure><h4><strong>How do world models work?</strong></h4><ul><li>The first step is to compress all the data from sensors, cameras, etc into a latent space. This is like a summary that only captures what matters to the model — not raw pixels, but object properties, position, velocity, etc.</li><li>Next, the model learns to predict how this state changes with each action. This was done using an RNN.</li><li>Finally, the current latent state and the predictions from the RNN are used to take the best possible action.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/0*a0HMwEz71NAfozl_.png" /></figure><h4><strong>Who is building these models today?</strong></h4><p>World Labs (founded by the veteran, Fei-Fei Li) are pioneers in this space and have picked up a funding round of $230M in 2024 from the likes of Andreessen Horowitz.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w68px9POnXRLUSpTJtNFYQ.png" /></figure><p>They have two important products — Marble helps you create your own 3D worlds from images, videos, text or 3D panoramas and RTFM (Real-Time Frame Model) makes interactive videos.</p><p>Google’s Gemini also creates such environments, whereas some models like Runway GWM focus on creating interactive videos rather than entire environments. Nvidia’s Cosmos platform takes a different approach, creating a playground for the model based on physics engines.</p><p>Additionally, robotics companies like Figure and self-driving cars like Tesla have their own proprietary world models that have started showing promise in the real world.</p><h4>What does the future look like?</h4><p>As AI moves from apps in our phones to our homes and in our workforce, world models play a pivotal role. They come with challenges of their own, being extremely compute intensive, struggling to maintain accuracy and showing inconsistency when you revisit a frame. But the speed of innovation tells us that it will only be a matter of time until they cause the next big shift in AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3a95d8c8227" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GPU, ASIC, NPU and FPGA explained]]></title>
            <link>https://medium.com/@ishisinghal/gpu-asic-npu-and-fpga-explained-d2bd9525c516?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2bd9525c516</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[semicondu]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Sat, 22 Nov 2025 17:39:55 GMT</pubDate>
            <atom:updated>2025-11-22T17:39:55.063Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oUwAgrWXUkR7tWSW" /></figure><p>For a long time, GPUs were the default hardware for AI compute. However, with models becoming larger, inference becoming more expensive and applications becoming increasingly diverse, custom AI chips have started gaining ground. This development is driven not only by efficiency gains, but also by BigTech’s desire to decrease reliance on Nvidia’s expensive and supply constrained GPUs. <br>Today, four types of compute architectures have dominated the discussion — GPUs, ASICs, NPUs and FPGAs, each coming with its own set of tradeoffs.</p><h4><strong>Graphics Processing Units (GPUs)</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*6OI1fYnA-6ckvwda.jpg" /></figure><ul><li>Originally built to render graphics, generating each pixel required a number of computations</li><li>Hence, unlike a CPU that can perform a few complex tasks sequentially, GPUs execute thousands of simple operations at the same time — this is called parallel processing</li><li>This gives very high performance and flexibility, as they are not made for any one task but this also makes them expensive and energy inefficient</li><li>Although Nvidia leads here, its competitor AMD has been seeing major commitments from the likes of OpenAI and Oracle</li></ul><p>More on the story of GPUs here -</p><p><a href="https://medium.com/@ishisinghal/from-gaming-to-global-power-the-gpu-story-a1ab9cbe11a2">From Gaming to Global Power — The GPU story</a></p><h4><strong>Application-Specific Integrated Circuits (ASICs)</strong></h4><ul><li>Google’s famous TPUs are a kind of ASIC — these are custom chips built for a specific type of computational task</li><li>Since the hardware has been tailored for this single workload, it gives much higher efficiency and is much cheaper than GPUs</li><li>However, they are not versatile — once manufactured, they cannot be changed</li><li>Tensor Processing Units (TPUs) are optimized for Google’s TensorFlow platform and have been used to train models like Gemini and Apple AI too</li><li>The famous transformer architecture is said to have been born only because of the existence of TPUs, as those algorithms would be near impossible to run efficiently on GPUs</li></ul><p>More on the transformer architecture here —</p><p><a href="https://medium.com/@ishisinghal/what-comes-after-transformers-the-100m-bet-on-ssms-bcbd7014a49b">What comes after Transformers: The $100M bet on SSMs</a></p><ul><li>In fact, Google also makes another type of ASICs that power YouTube, VCUs or Video Coding Units</li><li>While Google has seen the most success in ASICs, Amazon, Meta and Microsoft have also come up with their own custom AI chips, while OpenAI is set to start building theirs in 2026.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMucuOtJtqnM%3Flist%3DPLDt0J62yU45vkJfCCLvZJQQwVjjntRhks&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMucuOtJtqnM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMucuOtJtqnM%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/3096aadb3e19e8e088a3266504164a76/href">https://medium.com/media/3096aadb3e19e8e088a3266504164a76/href</a></iframe><h4><strong>Neural Processing Units (NPUs)</strong></h4><ul><li>An NPU is integrated into the chip of all your personal devices — they are used for edge processing — on device computation instead of on the cloud</li><li>This has far reaching implications, from ensuring privacy on your phone by running AI models locally, to reducing latency in a humanoid robot</li><li>It reduces time by shortening the distance that needs to be travelled by the data</li><li>These are made by most major semiconductor players — Nvidia, Intel, Qualcomm, AMD, etc. Although Apple does not use the term NPU, their neural accelerators that power the MacBook and iPhones have become an important part of their AI strategy. Samsung also has their own NPUs for their Galaxy phones.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FFviaHPrFhTg%3Flist%3DPLDt0J62yU45vkJfCCLvZJQQwVjjntRhks&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DFviaHPrFhTg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FFviaHPrFhTg%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e42982e582a94eca65307d6e59dbdf4a/href">https://medium.com/media/e42982e582a94eca65307d6e59dbdf4a/href</a></iframe><h4><strong>Field-Programmable Gate Array (FPGAs)</strong></h4><ul><li>FPGAs are basically more versatile ASICs — they can be reprogrammed even after they have been manufactured</li><li>However, this flexibility comes with lower performance and energy efficiency, making it the preferred option mainly for customers that do not wish to design their own ASIC</li><li>AMD, after acquiring Xilinx for $49B became the leader in FPGAs, closely followed by Intel.</li></ul><p>As compute becomes increasingly important in today’s world, everyone is racing to control it. This has brought about a variety of chip architectures that are optimized for different scales, applications and workloads. GPUs offer great flexibility, ASICs offer lowest costs, NPUs offer compute on-device. With the market becoming bigger day-by-day, we will probably continue to see new players and architectures come in and define the next decade of AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d2bd9525c516" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What comes after Transformers: The $100M bet on SSMs]]></title>
            <link>https://medium.com/@ishisinghal/what-comes-after-transformers-the-100m-bet-on-ssms-bcbd7014a49b?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/bcbd7014a49b</guid>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[funding]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Tue, 18 Nov 2025 15:48:22 GMT</pubDate>
            <atom:updated>2025-11-26T13:14:41.574Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jQgIaNogIBDA9mD6" /></figure><p>The future of how we interact with technologies is not through keyboards or screens, it is through voice. But building a model that can converse with you in real time is not that easy and this is exactly what Cartesia wants to change. Having raised $100M, they are redefining what a conversation with AI looks like. What makes them different? Rather than using transformers like everyone else, their model Sonic-3 uses something called State Space Models (SSMs). This signals an important shift away from traditionally used architectures.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SAUPJNaWcO0ZKYnH5hX1dg.png" /></figure><h4><strong>How do transformer models work?</strong></h4><p>This architecture is what most text models like ChatGPT, Gemini, etc use today. It has also been widely adopted in audio generation and image recognition models. How does it work?</p><p>The key is attention, in fact the architecture was introduced in a landmark paper titled “Attention is all you need” <br>So let’s understand attention first. Attention basically determines what parts of a data sequence a model should focus on. Stick with me here, it’ll get simpler. So for example in an LLM, when it’s processing a word, attention helps decide which other words in the input sequence are closely related to that one. Let’s take a look at the example Google AI gave me.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XqxMaK9icgyj3WfvhqeMHg.png" /></figure><p>How does it do this? It basically creates all words into vectors and assesses which vectors are most similar to each other.</p><h4><strong>Enter SSMs</strong></h4><p>State Space Models are something closer to how you and I operate — we don’t remember everything the other person said verbatim like transformers do. And we definitely do not revisit each and every word spoken before when we interact again. To solve for this inefficiency, State Space Models have a running internal state (like a memory store) that gets updated every time it encounters any new information. The famous Mamba model is also built on this architecture.</p><h4><strong>What does the future look like for SSMs?</strong></h4><p>Because they are much more efficient than transformers, SSMs have much lower latency. For natural conversation, this is huge, no one wants to talk to an AI bot that takes forever to reply. <br>On the flip side, most models have already been trained on the transformer architecture, and changing it to SSMs will mean rebuilding the entire infrastructure and retraining the models. This might be the biggest factor holding back their adoption.</p><p>Tranformers, though wildly successful, are bound to get replaced by something more efficient. This means that unless there is another breakthrough architecture in the near future, we might just see State Space Models running the biggest models.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcbd7014a49b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inside the shift from SEO to GEO]]></title>
            <link>https://medium.com/@ishisinghal/inside-the-shift-from-seo-to-geo-7585ddf845d6?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/7585ddf845d6</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[geo]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Mon, 17 Nov 2025 08:09:17 GMT</pubDate>
            <atom:updated>2025-11-17T08:09:17.289Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pX7JHw-CbxJO40lk" /></figure><p>Until a few years back, Google was the default gateway to the internet. In 2022, with the release of ChatGPT, that dominance began to erode. A survey conducted in June 2025 showed that 60% of Americans use AI to search for information, and this number only becomes higher when you look at the younger generations. Click through rates are falling rapidly on Google, since users are increasingly relying on the AI overview provided by the search engine.</p><p>This behavioural shift marks the beginning of a new era — where discovery is mediated not just by search engines like Google and Safari, but rather by AI chatbots like Perplexity and ChatGPT. Consequently, the traditional Search Engine Optimization (SEO) is losing relevance, making way for its successor — Generative Engine Optimisation (GEO). GEO makes sure that your brand, product and content are discoverable on AI generated responses and has become the new essential for online visibility.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_QgwRd0_d3OmRVvYMlOaxw.png" /></figure><h4><strong>How do generative engines work?</strong></h4><p>First let’s understand the search engine model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*aIoJFrUAC51n15sR.jpeg" /></figure><ul><li>First, bots/meta spiders scrape web pages from the internet</li><li>This data is stored in an index — these are massive databases that store only relevant data with meta information like keywords, user engagement and freshness. Any low-quality and duplicate data does not make it to this step.</li><li>When a user enters a search query, Google scans through its index to find results — weighing numerous signals like backlinks, click through history, relevance to the search query, etc. This step is the core of SEO — optimising all these factors in order to display your page on the Search Engine Results Page (SERP).</li></ul><p>Now looking at the generative engine model.</p><ul><li>LLMs are trained on millions of tokens — books, code repositories, web data, etc</li><li>Now when you pass your query a retriever module finds semantically similar content (for up-to-date answers) along with the training data. Here, instead of finding webpages and simply ranking them, LLMs prioritise information gain since they only use portions of text. <br>We will understand this with the help of examples when we deep dive into SEO versus GEO.</li><li>Then the engine fuses the retrieved information with the training data to form a coherent answer and outputs it to the user.</li></ul><p>Here is where the players like Parallel come in. Parallel basically provides APIs that an AI can use to access up-to-date web data. Founded by Ex-Twitter CEO Parag Agrawal, this startup has raised $100M from the likes of Kleiner Perkins, Index Ventures and Khosla Ventures.</p><h4><strong>SEO versus GEO — What changes?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ioH9MzoSP0uegPaW.png" /></figure><ul><li>The biggest shift is the prioritization of knowledge gain in generative engines. For example, if a website mentioned that they do not offer a service, they might still show up on a search engine but a generative engine rewards high-quality, factual, structured information.<br>Also, instead on focusing on backlinks like in traditional SEO, GEO would focus more on building a presence across many sources like LinkedIn, Wikipedia, etc</li><li>While crawlability and site speed matter in SEO, GEO is primarily concerned with whether the content is easy for models to parse and is not ambiguous.</li><li>Success is usually measured in SEO through click through rates and impressions, whereas visibility and citations are tracked for GEO</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7585ddf845d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Gaming to Global Power — The GPU story]]></title>
            <link>https://medium.com/@ishisinghal/from-gaming-to-global-power-the-gpu-story-a1ab9cbe11a2?source=rss-e33006a3e2b------2</link>
            <guid isPermaLink="false">https://medium.com/p/a1ab9cbe11a2</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ishi Singhal]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 10:20:37 GMT</pubDate>
            <atom:updated>2025-11-02T10:20:37.584Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Gaming to Global Power — The GPU story</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UvepbopvSwNI-yMM" /></figure><p>On October 29th, Nvidia became the first company in history to cross $5T in market capitalisation. A chipmaker becoming the most valuable company in the world signalled an important shift in the world — compute is now more important than ever and their Graphics Processing Units (GPUs) proved to be the ultimate source of it. Originally meant to be gaming hardware, they have powered the AI revolution.</p><p>GPUs today control who can train models and participate in the AI race. This quick read explores the technology that powers them and the business empire that Nvidia has built over them.</p><h4><strong>The Origin Story</strong></h4><p>In the late 1990s and early 2000s, CPUs were proving to be insufficient for graphics, they could not render millions of pixels simultaneously. While they could handle a few complex tasks sequentially, this problem required executing thousands of simple tasks at the same time — termed parallel processing. Here is where Nvidia’s GPUs came in. Each core of the GPU (processing unit) was less powerful than one of the CPU, but together they brought about breathtaking and realistic graphics into the world.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*ODQcui8YM_eNGkip.png" /></figure><p>The turning point came about in 2006, when Nvidia released CUDA (Compute Unified Device Architecture) — a platform that would allow developers to use GPUs and their power of parallelism for general-purpose computation. Now not just gaming, everything from physics problems to molecular modelling leverage these.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*7i2kFa2zvht5Tltx.png" /></figure><p>Then came deep learning, which required billions of matrix multiplications and GPUs had been doing this kind of repetitive, parallel operations for years. Thus, the piece of gaming hardware became an essential part of the revolution that was about to unfold.</p><h4><strong>Why did Nvidia win this game?</strong></h4><p>Naturally, as GPUs started to gain traction in deep learning applications, players like AMD attempted to enter the arena too, but no one could match Nvidia. This was because CUDA had formed an ecosystem moat — the software had been evolving for 20 years, forming a full-stack platform with multiple layers designed to make the use of Nvidia GPUs frictionless. Additionally, it had also locked in customers who had built all their infrastructure around Nvidia’s GPUs and had also trained all their models on it.</p><p>AMD missed out on the ecosystem glue, primarily focusing on hardware in their initial years. By the time they tried to replicate it with OpenCL and ROCm, it was too late. Intel had the advantage of owning their fabs but did not realise the potential of GPUs until it was too late. The likes of Google and Microsoft also came up with accelerators, which were chips specifically designed for the math computations needed in neural networks. However, these too could not break the ecosystem lock-in of Nvidia and ended up becoming just another player in the game, still relying heavily on Nvidia’s GPUs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*CMNdWHGadrvWco4_" /></figure><h4><strong>The economics of GPUs</strong></h4><p>The biggest cost (39%) of a $35 billion 1GW data center are the GPUs itself. Most big players like Meta, Open AI, etc use hundreds of thousands of Nvidia GPUs each year, with each H100 chip (also called the Hopper) costing between $25,000 to $40,000.</p><p>The margin Nvidia makes on these GPUs is a staggering 70%, resulting in a $73B profit in FY25. However, experts warn that these margins might soon degrade with the Big Techs working increasingly more on their own AI chips.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FnjyZR0Ceke0%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DnjyZR0Ceke0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FnjyZR0Ceke0%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/101133bb6ce7a167baeebd9b4400e938/href">https://medium.com/media/101133bb6ce7a167baeebd9b4400e938/href</a></iframe><p>GPUs are often referred to as the “new oil of the digital economy”. Countries and corporations are racing to make the best chips, seeking independence from each other. The strategic assets are being used to curb technology in other countries — prime example being the export controls from the US over China.</p><p>Just as the 20th century ran on oil, this decade is defined by compute. Whoever commands it is bound to lead in the technological era and by extension, tomorrow’s world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a1ab9cbe11a2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>