Harshit Tyagi

1 truth EdTech founders learn the hard way → Apps don’t build companies but trust does. People think EdTech is built on apps, features, and dashboards. I learned something very different while building Next Toppers. → Students don’t open an app because it looks good. → Parents don’t pay fees because of animations. They do it because they trust you. → They trust the guidance. → They trust the teacher. → They trust that one wrong suggestion won’t waste an entire year. Technology helps us reach scale. It helps manage classes, tests, and thousands of students at once. But it doesn’t teach. People do. That’s why every decision we take at Next Toppers starts with one question: Is this right for the student sitting on the other side of the screen? Because in education → Tech supports. Education leads. And trust decides everything.

421

15 Comments

𝗢𝗻𝗹𝗶𝗻𝗲 𝗴𝗮𝗺𝗶𝗻𝗴: 𝗦𝗸𝗶𝗹𝗹 𝗼𝗿 𝗷𝘂𝘀𝘁 𝗱𝗶𝗴𝗶𝘁𝗮𝗹-𝗮𝗴𝗲 𝗴𝗮𝗺𝗯𝗹𝗶𝗻𝗴? Dr. Arvind Gupta recently raised an important point in his interview with ET NOW: Digital platforms cross state boundaries, so regulation should come from the Centre—not individual states. I couldn’t agree more with this. When every state sets its own rules, we create confusion, loopholes, and even encourage illegal operators. But here’s where I’d take it a step further While many draw a line between “games of skill” and gambling, I believe that line is fading fast. Even so-called skill-based games can become addictive and financially risky. To me, they’re just gambling in a new form—with better marketing and design. It's time to stop treating them as harmless fun. We need strong, unified central rules that protect users, bring clarity, and support responsible digital growth. What do you think— Are “games of skill” really different, or just dressed-up gambling? #DigitalIndia #OnlineGaming #PublicPolicy #Gambling #DigitalEconomy #Regulation #TechPolicy #IndiaFirst

217

130 Comments

How can India be ready to build foundational level LLM/ AI models? I recorded a 3 hour workshop to demonstrate how to build a completely functional Small Language Model (SLM) from scratch. You can access it on our Youtube channel here: https://lnkd.in/gVUB3hr2 After this workshop, you will be equipped to build any Small Language Model (SLM) of your choice! The video already has a large number of views since it was released yesterday. My only hope and desire is that everyone learns to build SLMs from scratch. Here is what I cover in the workshop: (a) Download a dataset with 1million+ samples (b) Pre-process and tokenize the dataset (c) Divide the dataset into input-target pairs (d) Assemble the SLM architecture: tokenization layer, attention layer, transformer block, output layer and everything in between (e) Pre-train the entire SLM (f) Run inference and generate new text from your trained SLM! This is not a toy project. It's a production-level project with an extensive dataset. These 3 hours will train the next generation of foundational LLM/AI engineers. P.S: Want to build SLMs through our live classes? Explore our live lectures starting soon: https://lnkd.in/gPzeJbms

801

15 Comments

Dhruv Rathee apparently launched an app wrapper giving access to foundational models at ₹999 per month. Yes, it caps at 400k tokens a price many engineers argue is absurd compared to services like OpenRouter or GlobaGPT. But the real point isn’t about whether his wrapper is “useful” or whether he’s squeezing his distribution. The point is - he built that distribution. If people buy it, it’s because his audience sees value, or because they didn’t care enough to explore cheaper, better alternatives. Blindly following him is as foolish as blindly dismissing him. Entrepreneurship was never about building the “best” product in some absolute sense. It’s about leveraging what you have timing, trust, reach and seeing whether the market validates it. Good discovery is the responsibility of every individual. If Rathee profits, it’s not only because he packaged a subpar product as a startup. It’s also because we, as users, outsourced our discernment. Distribution is a moat; outrage is not. The lesson isn’t whether his app is “worth it.” The lesson is that the value of anything is co-created: half by the builder, half by the buyer. ❤️

935

153 Comments

Apparently, some people were terrorized after finding out that AI graded 10,000 handwritten exam papers in India's Jodhpur district. One argument was that since the technology used in this program was made by a for-profit company, their motives must surely be questionable, and another terrified person said that AI was just giving marks without providing any understanding to the teacher. Well well... The fact is that my organization has been a donor to a state government for three years to create a large-scale impact of assessment digitization using AI. Some people ask me: "What benefit is it to you, aren't you a for-profit company?" Well, we are a for-profit with a non-profit mindset (after reading this, no VC will give me any money, but I'm going to take that risk). We believe that investing in education will lead to societal growth, and so we do it because of our values and principles. EdTech companies don't have to be non-profit to create a genuine positive impact in society. Also, Smart Paper by EdOptimize is not just about grading efficiency and mechanizing the human elements of the assessment process. Technology is an extension of human ability. With our tool, we are giving teachers the ability to handle a large amount of handwriting in a systematic way. Our tool is not limited to just marks/points - we are helping people look into the thinking process in handwriting. LLMs are good at analyzing text data and finding patterns in a large number of written responses. But somehow, one terrorized person missed this feature entirely and was hence terrorized. Have you done something in EdTech that terrified other people? Please do share.

71

30 Comments

ChatGPT doesn’t replace thinking. It redefines the landscape of thinking. But it will always produce false information. 💭 “ChatGPT will always give you BS answers — it’s built that way.” I’ve seen this claim a lot lately. ChatGPT doesn’t understand the world; it predicts language patterns. That means it can sound right while being wrong — confidently. But calling that “by design” misses the point. When calculators entered classrooms in the 1960s, teachers said they’d “kill real math.” When priests once mediated truth, people feared chaos if everyone read for themselves. In both cases, we didn’t abandon the tool — we learned to use it wisely. ChatGPT is the same. It’s not an oracle of truth — it’s a language calculator. It handles words; we handle meaning. The future of AI isn’t about replacing thinking — it’s about elevating it.

104

51 Comments

Amid the GPT-5 buzz, we asked a simpler question:  👉 Is it truly better for our test-prep exams - and by how much? ✅ At PhysicsWallah, we prioritise students' experience and learning success over everything else. Because for our students, 2 percentage points isn’t “statistically insignificant” - it could be the difference between IIT Delhi and a backup plan. So we lined up every major AI model, including the freshly launched GPT-5, and ran them on JEE numericals set, representative of actual exam questions. The verdict?  🏆 GPT-5 led the pack in accuracy (92.22%).  ⚡ GPT-o4-Mini (Low) and Gemini 2.5 Flash matches each other closely - but O4-Mini beats them on latency.  📉 GPT-5’s latency was significantly higher than most models (second only to Gemini 2.5 Pro) - SLOW enough to make students drop off even before seeing the output. 💡 What this means: Smarter isn’t always better. The right answer is the one that comes fast, reliably, and at scale — even if it’s not from the “shiny new” model ❓If you were building for millions of users, would you choose…  A) 2 pp more accurate results or  B) 3× faster responses at lower cost?

26

Here is a 2-minute, easy-to-understand explanation of Model Context Protocol (MCP) and how it works with LLMs: Imagine you are chatting with an AI assistant like ChatGPT. You ask: “What’s the weather in Pune?” Now, normally, the AI might just guess or hallucinate the answer. But with Model Context Protocol (MCP), the AI does something much smarter-it reaches out to a real tool (like a weather API) to fetch the actual data. 🧠 How does it know what to do? Before answering, the AI is given a list of tools it can use-almost like a menu of services. Each tool has a clear description, the inputs it needs, and the format of the response. For example, it’s told: There’s a tool called "get_weather" that takes a "city" and returns temperature and condition. 🔄 What happens next? When you ask your question, the AI uses this information and generates a structured request like this: {  "tool": "get_weather",  "parameters": {   "city": "Pune"  } } This message is sent to an MCP server, which runs the actual tool and returns real data like: {  "temperature_celsius": 32,  "condition": "Sunny" } The AI sees this and responds to you with: “The current weather in Pune is 32°C and sunny.” ✅ Why is MCP powerful? Reliable: The tools have fixed input/output formats, so the AI can use them without guessing. Safe: If the AI sends a wrong input, the server rejects it clearly. Traceable: Every tool call and response is logged - you can replay it and verify what happened. Flexible: You can plug in any tool - math engines, databases, internal APIs - and the AI can use them all through the same protocol. ______ Follow me (Sreedath Panat) for daily AI/ML insights.

104

6 Comments

Most underrated skill right now? → asking better questions. Combine that with: → YouTube transcripts → ChatGPT → an obsession with iterating and you can skip 3-4 years of traditional education. Especially in countries like India, where the system is still stuck in the 90s. At 19, young learners can know more about go-to-market than MBA grads. Because they're not waiting for a syllabus. They’re building one in real-time. The formula is simple: pick a domain, binge the top 50 podcasts + videos, dump transcripts into ChatGPT, ask questions, challenge the answers, repeat until you actually understand. You don’t need degrees. You need better curiosity loops. School teaches you how to memorize. This teaches you how to think. And thinking is a monetizable skill now.

24

1 Comment

We have launched a course on “Reasoning LLMs from Scratch” on our YouTube channel. This course will focus heavily on the fundamentals and give you the confidence to understand and also build a reasoning model from scratch. My approach: No fluff. High Depth. 4 lectures have been uploaded in this playlist as of now. Lecture 1: Introduction to the course Link: https://buff.ly/eHunHbi Lecture 2: Chain of Thought Reasoning Lecture Link: https://buff.ly/GbAHXkj Lecture 3: Verifiers, Reward Models and Beam Search Link: https://buff.ly/my3lLAP Lecture 4: Fundamentals of Reinforcement Learning (Premieres this Friday at 10 AM IST) Link: https://buff.ly/Y7f36gq Image Below: Notes of my Reinforcement Learning Lecture.

59

2 Comments

UPI just changed the game and most people don't even know it yet. UPI processes 40+ crore transactions daily. Starting August 1, 2025, new NPCI guidelines reshaped how those massive volumes work: 🔹 Only 50 balance inquiries per app per day 🔹 25 requests to view linked accounts per day 🔹 Autopay executions only during non-peak hours (avoiding 10am-1pm and 5pm-9:30pm) 🔹 Maximum 4 attempts to execute a single auto-debit Why do these changes matter? The system was getting overloaded with this massive daily volume. Too many balance checks, too many retry attempts, peak hour congestions. NPCI is basically saying: "We need to optimize for scale, not convenience." Your UPI experience just got more structured. No more unlimited balance checking. No more peak-hour autopay failures eating up your bandwidth. For fintech companies: Compliance deadline was July 31. Non-compliance means API restrictions, penalties, or suspension of new customer onboarding. For users: Smoother transactions during peak hours, but with some usage limits. These rules are about making that 40+ crore daily transaction volume sustainable long-term. Drop your thoughts below ⬇

53

6 Comments

We’ve built AI tools, automation layers, and custom AI workflows for a lot of Indian Enterprises. And in the beginning, everyone wanted a pilot. One small use case. One safe experiment. But over the past year, something shifted.   The same companies who were nervous about even one AI use case, are now running multiple GenAI deployments in parallel.   The new EY 𝗔𝗜𝗱𝗲𝗮 𝗼𝗳 𝗜𝗻𝗱𝗶𝗮: Outlook 2026 report came out, and it honestly felt like a mirror of what we see on the ground.   𝟰𝟳 𝗽𝗲𝗿𝗰𝗲𝗻𝘁 𝗼𝗳 𝗜𝗻𝗱𝗶𝗮𝗻 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗵𝗮𝘃𝗲 𝗺𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝗚𝗲𝗻𝗔𝗜 𝘂𝘀𝗲 𝗰𝗮𝘀𝗲𝘀 𝗹𝗶𝘃𝗲 𝗶𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻. Only 23 percent are still in pilots.   That’s a huge shift from pilot to performance.   Even the sentiment has changed:   𝟳𝟲 𝗽𝗲𝗿𝗰𝗲𝗻𝘁 𝗼𝗳 𝗹𝗲𝗮𝗱𝗲𝗿𝘀 𝗻𝗼𝘄 𝗯𝗲𝗹𝗶𝗲𝘃𝗲 𝗚𝗲𝗻𝗔𝗜 𝘄𝗶𝗹𝗹 𝗰𝗿𝗲𝗮𝘁𝗲 𝘀𝗶𝗴𝗻𝗶𝗳𝗶𝗰𝗮𝗻𝘁 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀 𝗶𝗺𝗽𝗮𝗰𝘁 and 63 percent feel ready to leverage it.   That level of confidence wasn’t here two years ago.   Our experience has been the same. Right now, every leadership room is asking one question: Where do humans create the most value, and where can AI agents take over the rest?   EY calls this the move from experimentation to autonomy and from our side, we strongly agree with EY’s direction - India is moving fast, but the real advantage will belong to companies that build systems where humans and AI collaborate seamlessly and where AI isn’t a pilot, but a partner.   If you haven’t already, the report is worth reading – link in the comments below. #EnterpriseAI #EY #AgenticAI #AIAdoption #AIinIndia #ShapeTheFutureWithConfidence

240

23 Comments

📲EdTech or 👩‍🏫EdTeach? Most so-called “EdTech” apps in India are actually EdTeach — scaling teachers and video lectures, not technology. That’s why outcomes stayed flat while valuations ballooned. At MemoNeet we’ve taken a very different path. 👉 We’re serious about gamification - making boring revision addictive through streaks, levels, and rewards. 👉 Our next release pushes this even further with adaptive revision and AI-driven memory tracking. 👉 That’s why investors are showing strong interest because we’re not building another online classroom, we’re building real EdTech. And here’s the big validation: OpenAI just appointed Raghav Gupta (ex–Coursera MD Asia Pacific) to lead its education push in India and Asia. The world’s leading AI company believes education is still the largest AI market. Round One of EdTech was video lectures. Round Two will be AI + gamification → tools that adapt to every student and deliver real outcomes. #edtech #openai

34

1 Comment

Seeing a lot of chatter around edtech ESOPs. The thing is, founders must structure ESOP buybacks this way. Buybacks are more about belief than just monetary value. Most long term folks don’t want to sell much. They are just looking out for the company’s intent primarily. And what each buyback solves is something for that : 1. Some sort of a validation for folks that their stocks aren't just paper, but can be actual currency. Pre-emptive signal that their effort can be really valuable one day. 2. Biggest display that the co actually cares for the team. Buybacks come directly from cash reserves, requiring a strong internal pitch to the board. Action > Words. — Easier, one can always align it towards a common company goal - biz performance / new round raised / hitting profitability etc etc Goes good for the biz, good for the team :) — And having done two buybacks at Classplus (and hopefully the third one in 2026) - can easily say that the happiness of seeing a teammate do that house down-payment / buying that car / gifting their family a foreign vacay is at par with any product innovation / funding round / metric hit. Humanity is the most underrated aspect about building a biz :) 

265

12 Comments

A Billion Screens. A Billion Students. But Are We Assessing Them Right? In traditional classrooms, we had visual cues — Raised hands, puzzled looks, whispered doubts. But in online learning? You need new eyes to spot real progress. Here’s what I’ve learned at ScienceWithShobha after thousands of virtual sessions: 🔍 Don’t just track answers. Track questions. A curious student is a growing student. 📈 Look for patterns, not perfection. If they solve 7 of 10 questions with confidence, that’s progress. Don’t wait for 10/10. 🧠 Assess thinking, not memorizing. Ask: “How did you arrive at this answer?” Their logic reveals more than their score. 💬 Make reflection part of the assessment. Every week, ask students: “What’s one thing you understand better now than before?” 🫶 And always: Give feedback that feeds hope. Not “You got it wrong,” But — “You’re close. Let’s try it another way.” Because online learning isn’t about control — It’s about connection. Let’s assess with insight, not just data. Let’s raise learners, not just achievers. 💡 I’m curious — how do you track real learning in your online or hybrid classrooms? Whether you're a teacher, coach, or leader — what signs of growth do you look for beyond grades?Let’s learn from each other. 👇 #ScienceWithShobha #OnlineLearning #AssessmentMatters #EdTechWithHeart #LearningDesign #StudentGrowth #EdLeaders #TeachersOfLinkedIn #LinkedInLearning #EducationReimagined

47

33 Comments