Sudheer Tammini⚡

Writers @SettingBox, trying to decipher hard subjects like Data Science | Data Visualization | Python | Excel | Finance

[Sudheer Tammini⚡ @SudheerNotes]

Here is a simple Free & Open source software that I have developed to export Mutual Funds CAMS statement data on to a nice CSV file. #FOSS
refer below link for details https://t.co/PfG7PNc1E2

[Sudheer Tammini⚡ @SudheerNotes]

@rasbt Congratulations Sebastian, great fan of your writing I must say Build LLMs from scratch is a foundational writing and I'm very sure this book will be a game changer as well.

[Sudheer Tammini⚡ @SudheerNotes]

Whether you're an AI engineer or an executive architecting the future of corporate data systems, understanding this mechanism changes the game.

Read the full step-by-step breakdown & grab the modular code here: https://t.co/g8ij2bRHa0 (4/4)

[Sudheer Tammini⚡ @SudheerNotes]

Most people talk about Enterprise LLMs at a surface level.

But if you want to understand how autonomous analytical pipelines actually process complex contextual data, you have to look at the math underneath.

Let’s talk about Multi-Head Attention (MHA) 👇 (1/4)

[Sudheer Tammini⚡ @SudheerNotes]

In my complete implementation guide, I break down:
- The math behind Q, K, and V matrices 🧠
- Why dividing by $\sqrt{d_k}$ saves your gradients 📉 - Explicit tensor shape tracking at every layer 💻 (3/4)

[Sudheer Tammini⚡ @SudheerNotes]

I was first introduced to Deep Learning via FastAI course and I've no computers degree (I'm an MBA) but the course was so well structured I was hooked immediately and ever since I've been learning, building and training models.

[Ihtesham Ali @ihtesham2005]

A McKinsey consultant with no PhD, no AI background, and no academic position quietly built the most-watched deep learning course on Earth and gave the entire thing away for free.

I opened the first lesson at 1am and could not believe a single human had taught this many people without ever charging a cent for it.

His name is Jeremy Howard. The course is called "Practical Deep Learning for Coders"

For most of his career, Jeremy Howard had no business teaching artificial intelligence to anyone. He spent 8 years as a management consultant at McKinsey and AT Kearney. He started an email company called FastMail in Australia and ran it for years. He built an insurance pricing startup that got acquired by Lexis-Nexis. None of that was AI work. None of it was academic. He did not have a doctorate in computer science or mathematics or anything else. He was a businessman who happened to be very good at writing code.

Then in 2010 and 2011, he became the top ranked competitor in data science contests on Kaggle, beating teams of PhDs from the most credentialed labs in the world. He was eventually made president of the company. And what he saw from that seat was the thing that ended up changing his life.

The best deep learning research on Earth was being done by maybe a few hundred people in five or six elite labs in San Francisco, London, and Toronto. To get into those labs, you needed a PhD from Stanford or MIT, a recommendation from a tenured professor, and access to expensive GPUs that no individual could afford. The practical knowledge of how to actually train these models was almost never written down. It lived inside the heads of a small priesthood, and the priesthood was almost entirely closed to outsiders.

In 2016 Jeremy Howard and Rachel Thomas decided to break that gate.

They founded https://t.co/zIF9XHFhgn with one mission. Take the techniques that the elite labs were using and teach them to anyone in the world who could write a Python loop. No PhD required. No advanced math required. No expensive hardware required. Just a laptop, an internet connection, and the willingness to actually finish the lessons.

The way they did it was the part that almost nobody in academia had ever tried.

Every other deep learning course on the planet started with theory. You learned linear algebra. You learned multivariate calculus. You learned probability theory. You spent 6 months on the foundations before you ever got near a working neural network. By the time most students reached the part where they could actually build something, they had quit.

Howard inverted the entire curriculum. Lesson one of his course is not theory. It is a working image classifier that you train in 15 minutes on a model that can distinguish dog breeds with 99 percent accuracy. You build the thing first. You make it work first. Only then, once you have proven to yourself that you can actually do this, do you start peeling back the layers to understand why it works.

His justification was simple. The reason most people quit learning hard things is not that the material is too difficult. It is that the curriculum is structured to make them feel stupid for as long as possible before they ever get to do anything interesting. Howard refused to do that to his students. He believed that if you could see something working with your own hands on day one, you would have the motivation to fight through the hard math three months later. And he was right.

The course has now been viewed over 6 million times. His students include a Canadian dairy farmer who used the course to build an AI system to monitor the health of his goats. They include a French math teacher and a network of doctors in Africa. They include people who walked into the lessons knowing nothing about AI and walked out building production systems at Google Brain, OpenAI, Adobe, Amazon, and Tesla.

In 2018, Jeremy Howard and a young researcher named Sebastian Ruder published a paper called "Universal Language Model Fine-tuning." It introduced a transfer learning technique for language models that worked so well it cut error rates on text classification by 24 percent on the hardest benchmarks in the field.

That technique, refined and scaled by the labs that came after, became the foundation of how every major language model on Earth is trained today. ChatGPT. Claude. Gemini. The fine-tuning step that makes them useful traces back to the methodology in that paper.

The man who co-wrote it had no PhD. He had been teaching the same ideas to strangers on the internet for free a year before he ever published the paper.

His students were already building with it before the elite labs had even read it.

The best things on the internet are almost never the ones with paywalls in front of them. They are the ones built by people who decided the gate was the problem and then quietly walked around it.

[Sudheer Tammini⚡ @SudheerNotes]

Ever wondered how Self-Attention mechanism works in AI models? here is a detailed post with step by step implementation ==> https://t.co/m9bVcCLOHd
credit @rasbt
#AI #DeepLearning #PyTorch

[Sudheer Tammini⚡ @SudheerNotes]

But now that we have already paid TP in advance and most of us think its one time cost so there is no reason for them to still keep price attractive. So they choose to increase OD premium from second year.

[Sudheer Tammini⚡ @SudheerNotes]

Further update on this. The @ACKOIndia escalation team reached out to me and gathered all the details & after checking internally they have come back saying this is the final premium & they were not able to give me reasons for 7 time premium increase. #ScamAlert

[Sudheer Tammini⚡ @SudheerNotes]

@ACKOIndia two wheeler insurance premium scam - details below:
#1 Last year on Jan 2025 I bought a 350 CC bike and I have opted for ACKO insurance, I've selected comprehensive policy for 1 yr Own Damage + 5 Yrs Third party, I bought directly from Acko.

[Sudheer Tammini⚡ @SudheerNotes]

As I mentioned before, I had my old bike insurance policy from the same vendor when there was no mandatory 5 years requirement. At that time the bulk of the premium has been allocated to TP & small amount for OD to make pricing attractive.

[Sudheer Tammini⚡ @SudheerNotes]

@ACKOSupport Just to make things clear, your colleague didn't say anything new nor was able share any reasoning for the increase. So she just said that's it and pay the premium. I've moved out of Acko and bought a new policy & I'll lodge a compliant with IRDA.

[Sudheer Tammini⚡ @SudheerNotes]

@ACKOIndia two wheeler insurance premium scam - details below:
#1 Last year on Jan 2025 I bought a 350 CC bike and I have opted for ACKO insurance, I've selected comprehensive policy for 1 yr Own Damage + 5 Yrs Third party, I bought directly from Acko.

[Sudheer Tammini⚡ @SudheerNotes]

@ACKOSupport I just responded.

[Sudheer Tammini⚡ @SudheerNotes]

#9 This entire thread is not about the 3k Premium but how they have systematically manipulating customers. In this case I still have a choice to buy OD from other vendors but imagine if you have bought their so called Health Insurance.

[Sudheer Tammini⚡ @SudheerNotes]

#8 But in this case they have already collected TP premium in advance for next 5 years, so now they are taking advantage of OD renewal to bump up premium.