Packt

Editor in Chief

Glance

Product Management Intern

Mela

Content Marketing

Center for Entrepreneurial Leadership(CEL), BITS Goa

the best club in bitsg. without the support of the amazing people in this amazing club, i wouldn't be a writer today

Core

Crew Member

My daily routine: 
6am: wake up
6:05am: check LinkedIn, see notification that “Someone has viewed your LinkedIn profile”
6:06am - 10pm: Stare at ceiling, ponder who it could have been.
10pm: bedtime

(@Chris Bakke)

Google was trying to build ChatGPT in 2012.

Or more specifically, a team inside Google was trying to build an AI for Google Search so that when people typed a question in the search box, the AI would directly give the answer instead of showing ten blue links to the user.

But the AI that they were using had a bad memory. It kept forgetting the user’s question. Answering it was far away.

The team tried different methods to increase the AI’s memory. Didn’t work.

One guy on the team, Jakob Uszkoreit, had an idea: instead of trying to increase this AI’s memory, why don’t we make another type of AI from scratch that has a strong memory?

No one took Jakob’s idea seriously.

But Jakob didn’t give up.

He found a couple of people in Google who found this idea interesting: Illia Polosukhin(AI scientist), Ashish Vaswani from Google Brain, Niki Parmar from Google Search, Llion Jones from Google Research, Łukasz Kaiser and his intern, Aidan Gomez.

Together, they started working to turn Jakob’s idea(new type of AI) into reality.

They called it: transformer.

They quickly built the prototype version of transformer and used it to make a translator.

For strong memory, they designed the transformer to look at texts in one shot instead of word-by-word. So the transformer remembered what it read earlier.

The translator performed quite well. Full speed ahead.

But then they hit a wall.

Making the prototype version of transformer was easy. But they didn’t know how to make it perfect.

The problem was, these people were brilliant and smart, but to build a new type of AI, you need a legend. You need lisan al gaib.

Enter Noam Shazeer.

Noam was a legend at Google. He was one of the earliest employees, built Google’s ad system, was a principal engineer, and had written countless research papers on AI.

One day in 2017, he was walking in Google’s Building 1965 when he heard some people talking about a new type of AI.

He was amazed. He joined Team Transformer immediately.

On his first day, he threw away the code that the team had written.

He came back with his own code for the transformer.

After that, the team perfected the idea of transformer and wrote a research paper about transformers called Attention is All You Need.

Google execs did not understand the paper, but a small company called OpenAI did.

They used transformer to create an AI called ChatGPT, where T stood for transformer.

And the rest is history.

Follow Shreyans Singh for more stories like this.

Shreyans Singh

Blinkit launched PS5 today at 10am today.

2 hrs later when I checked, it was out of stock.

Who is buying PS5 from a grocery app?

Why do they need it in 10 minutes?

Why is Blinkit selling PS5?

Third question is the most interesting.

Blinkit(and all other 10 min apps) started with groceries.

They wanted to disrupt kirana shops.

They selected:

⭐target market(metro cities)

⭐target customer(lazy rich people)

⭐operating model(lots of small dark stores in a city that have a limited stocks)

And it worked! All these apps have grown rapidly, Zepto and Blinkit are close to profitability.

But somewhere along the way, they realized that lazy rich people don’t just want groceries in 10 minutes.

They want everything in 10 minutes: clothes, shoes, beauty, electronics, books, toys, stationery. Everything.

So they said ok.

Slowly, they started adding extra stuff. 

First stationery(remember Blinkit printing service for 20 per page?), then clothes and then iphones and samsung ultras.

And now, PS5’s.

So this is the real reason why Blinkit is selling PS5s. Their customers want it.

(note: I have said “want”, not “need”. No one “needs” PS5 in 10 minutes. But if you offer it to me, even I “want” it in 10 minutes)

In fact, the demand for blinkit to stock everything is so strong that they are looking for bigger dark stores (source: ET).

Obviously, Amazon and Flipkart are threatened by this.

Initially, they weren’t threatened by Blinkit because they thought it was a grocery app and it will compete with kirana shops.

But surpriusngly, it has become an ecommerce app and now its competing with both Amazon and Flipkart.

As a result, Amazon is focusing on 2 hour deliveries. And Flipkart is restarting Flipkart Quick delivery service.

Follow Shreyans Singh for more insights like this.

On 21st February, Google’s AI (Gemini) made a mistake.

Google immediately lost $90 billion from its market cap.

People’s thinking: Because their AI made a mistake, Google doesn’t know how to make AI. RIP Google.

The same thing happened last year in February.

Google’s AI(Bard) made a mistake and Google lost $100 billion from its market cap.

Two important lessons here:

1. February is a bad time for Google.
2. People are massively underestimating Google’s AI power.

Gemini is an incredibly powerful AI.

It has a capacity of 10 lakh tokens, compared to GPT4’s 1 lakh tokens.

Its 20 times cheaper than GPT4.

It beats GPT-4 on 30 out of 32 AI performance tests(benchmarks).

The second version(Gemini 1.5) is even more amazing.

It can analyze hour-long videos, 30,000 lines of code, and make $100 billion losses in market cap.

People were afraid that ChatGPT will disrupt Google Search. 

But people forget that Google Search has faced multiple disruptions over the last 20 years and it won every time.

(List of disruptions: desktop to mobile shift, social media, verticalized search)

To fight the AI disruption, Google launched Search Generative Experience.

It gives an AI generated answer to your google search so that you don’t have to click on ten blue links.

Not just Search, Google is putting AI into all its products: Gmail, Maps, Photos, Docs, Assistant, Cloud, and Pixel smartphones.

Pixel 8 phone uses Gemini for AI features like smart recording, smart reply, etc.

Bloomberg reported that Apple is talking with Google to use Gemini in iPhones.

Samsung signed an exclusive multi-year deal with Google to use Gemini in Samsung phones.

This year, Samsung’s S24 series phones are using Gemini for AI features like Circle To Search, Magic Compose, live-translate phone calls, transcribe voice recordings, summarise web articles, fix handwriting.

Samsung phones are also using Google’s Imagen2 AI for photo editing features.

Still think RIP Google?

Follow Shreyans Singh for more insights like this.

Starting this week, Ola and Uber will disable online payment for rickshaw bookings. Only cash allowed.

They will also stop taking commissions from auto rickshaws.

Instead, they will take daily/weekly subscription fees. 

(Online payment means ola money/uber money/rapido coins/credit card/amazon pay. You can still pay using UPI after the ride ends. If the driver doesn’t use UPI, start running)

And most importantly, they won’t set prices. Rickshaw drivers can charge whatever they want.

Amazing, right? 🙂🙂

Namma Yatri started this amazing concept.

They saw that rickshaw drivers didn’t want to pay commissions for every ride on Uber/Ola.

So they started charging Rs 25 per day.

Drivers happy. People happy.

And Namma Yatri happy.

By charging fixed fees instead of commissions, they escaped the 5% GST that Uber Ola Rapido were paying.

How?

Because on 15th September last year, the Karnataka Authority for Advance Rulings said that CGST Section 9(5) is not applicable to Namma Yatri.

Namma Yatri is only connecting customers with rickshaws.

They are not providing the service.

Namma Yatri is not driving the rickshaw. Namma Yatri is not setting the prices. Namma Yatri is not collecting the payments.

So they don’t have to pay 5% GST.

Rapido copied this idea in February. And now Ola Uber have also copied it.

Follow Shreyans Singh for more insights like this.

Pavan Davuluri, an IIT Madras alumnus and working at Microsoft for the last 23 years. He has just been appointed as the new head of Windows and Surface divisions at Microsoft👏👏

All big tech companies are building gen AI: Google Gemini, Amazon Q, etc.

But Apple is doing the opposite.

Yesterday, Bloomberg reported that Apple is talking to Google about putting Gemini in iPhone.

(Gemini is a generative AI model- it generates text, pics, etc)

Inside iPhone, Gemini will power the new iPhone 16’s AI features.

This is not new.

In January, Samsung released S24 with a lot of AI features.

It was powered by Gemini.

(Looks like Google has figured out a business model for Gemini)

Last year, Apple was also testing an AI chatbot called Ajax.

Looks like it didn’t work.

What I don’t understand is: if Apple’s AI failed and Apple still wanted to give AI on iPhone, why did it go to a direct competitor? 

Why not a less competing company, like Amazon or Microsoft or OpenAI?

I think the answer is: Apple thinks that generative AI will be a very small part of AI.

Most probably, they think that generative AI features like “writing emails” or “summarizing websites” won’t be very profitable.

That’s why they are ok with outsourcing it from a direct competitor.

That’s why, last year Tim Cook promised that amazing new AI features will come to Apple’s platforms.

Most probably, they are integrating AI into iOS so that the iPhone becomes smarter.

Something like, “hey siri order pizza from dominos” and the AI will open dominos app, add pizza to cart, select the address, and do the payment.

High hopes from Apple this year🤞

Follow Shreyans Singh for more insights like this.

I’m bored of “First day at Google” posts. But this “First day at Google” post by Fei-Fei Li, who joined Google as Chief Scientist of AI is awesome:

📌“I had decided to take a job as chief scientist of AI at Google Cloud in 2017” 

📌“Everything I saw was bigger, faster, sleeker, and more sophisticated than what I was used to.”

📌“The abundance of food alone was staggering"

📌"The breakrooms were stocked with more snacks, beverages, and professional-grade espresso hardware than anything I’d ever seen at Stanford or Princeton, and virtually every Google building had such a room on every floor. And all this before I even made my way into the cafeterias.”

📌“After so many years spent fuming over the temperamental projectors, meetings at Google were like something out of science fiction.”

📌“Cutting-edge telepresence was built into every room, and everything was activated with a single tap on a touchscreen.”

📌“Then there was the talent—the sheer, awe-inspiring depth of it." 

📌"I couldn’t help but blush remembering the two grueling years it took to attract three collaborators to help build ambient intelligence for hospitals.”

📌“Here, a 15-person team, ready to work, was waiting for me on my first day."

📌"And that was just the start—within only 18 months, we’d grow to 20 times that size. Ph.D.s with sterling credentials seemed to be everywhere”

📌“Finally, there were the data—the commodity on which Google’s entire brand was based.”

📌“I was surrounded by data—data in categories I hadn’t even imagined before: from agriculture businesses seeking to better understand plants and soil, from media-industry customers eager to organize their content libraries, from manufacturers working to reduce product defects, and so much more.”

Follow Shreyans Singh for more.

(These are some interesting points from her article. The entire article is more interesting. Link below)

Did you notice that on Zepto, you have to first add the item to the cart. But on Amazon, you can order directly using the Buy Now button. Why?

I wonder why these 2 apps show such different behavior for the same order.

Why does Zepto force me to go the cart? And why doesn't Amazon send me to cart?

One theory is that Zepto wants people to put more stuff in their cart. They want to increase the average order value of customers.

10 minute delivery is expensive. Zepto can't afford to do it for single items. They will lose a lot of money per order like that. 

That's why they want us to order lots of stuff. It will make delivery profitable for them.

But this theory doesn't answer Amazon's side of the story.

Why is Amazon not increasing the average order value/cart value?

Surely they are also losing money by delivering single items like this.

Are they eating the losses? Or do they have an amazing delivery network, better than Zepto, that profitably delivers single items?

A few years ago, Nvidia took a big risk.

All other chips in the market used to process AI at high precision(FP16 and FP32).

But Nvidia decided that, their new AI chip (called Hopper) will process AI in low precision(FP8).

This was a big risk.

What if FP8 precision was too low to accurately process AI?

But the risk paid off.

Hopper was able to accurately process AI and it took Nvidia to a trillion dollar valuation.

Now, Hopper is old. Nvidia has launched a new AI chip called Blackwell.

And this time, they have taken an even bigger risk.

Blackwell is going to process AI at not FP32, not FP16, not FP8, but FP4 precision! (with FP6 for backup)

But Blackwell doesn’t stop here.

Compared to Hopper, Blackwell is 4 times more powerful in training AI, 30 times more powerful in AI inference and it uses 25 times less energy.

How?

The answer is boring: Blackwell has two GPUs. Hopper had a single GPU.

In the Blackwell chip, Nvidia has put two reticle-sized GPU chips, each with 104 billion transistors, and combined them with a superfast 10TB per second connection.

(The superfast connection is important because these two GPUs have to work as a single GPU. Nvidia doesn’t want the software to see two GPUs).

Nvidia goal for Blackwell is big. 

They want companies to use it to train AI models with trillion parameters.

Nowadays, companies are making AI using a technique called Mixture of Experts. (example: GPT4, Gemini)

In this technique, you take many small AI models that are expert in one topic, and mix them to make one big AI model.

Hence: Mixture of Experts.

Nvidia has added a special feature in Blackwell just for Mixture of Experts: model parallelism. 

In this feature, Blackwell will train all the smaller AI models parallelly. So we can see which AI is the expert at giving what kind of answer.

Amazing, right?

I haven’t told you the most amazing part: a single Blackwell chip will cost between 25-35 lakhs. 

To train a 1.8 trillion parameter AI model, you need 2000 Blackwells.

So your total cost is just 600 crores! Amazing, right?

Follow Shreyans Singh for more amazing stuff like this.

24 hours since Zomato launched Pure Veg Mode.

One feature(green tshirt) has already been rolled back.

Green tshirt was a bad idea anyways. It directly implied that red tshirt was nonveg delivery.

But otherwise, the Pure Veg Mode is a very good idea.

Not because of my personal preference.

But because there are lots of pure veg customers in India.

Because of their personal beliefs, they don’t want to eat, touch, see, or smell nonveg food.

Its our job to respect their beliefs.

So far, we have been doing a good job.

We have red and green labels on food packets.

We have pure veg restaurants.

Pure veg mode on Zomato is simply an extension of this.

When a cloud kitchen makes both veg and nonveg food in the same kitchen, it is offending the beliefs of pure veg customers.

When the delivery person carries veg and nonveg food in the same bag and food spills, it is offending the beliefs of pure veg customers

When restaurant mistakenly puts nonveg in a veg order, it is offending the beliefs of pure veg customers.

Offending their beliefs is not cool. Our job is to respect their beliefs.

That’s why Pure Veg mode is needed.

It will give assurance to pure veg customers that their food has not been made in the same kitchen with non veg food.

Pure veg food won’t be carried in the same delivery bag as veg food.

Pure veg restaurant cannot accidentally put nonveg stuff in their order.

These were my unbiased, balanced thoughts on this topic. Follow Shreyans Singh for more unbiased, balanced stuff. 

This Shark Tank startup came to WeWork yesterday. They were promoting their smart cooker. It is a cooker plus a screen attached. It costs 23k. Thoughts?

If I ask ChatGPT “who will win IPL”, it doesn’t know.

But if I show Jay_Shah_Script_IPL_2024 . pdf to ChatGPT, it will give me the correct answer: GT

This is called RAG.

RAG is a very popular technique in AI nowadays.

AI models like ChatGPT have a very bad habit.

Hallucination.

They give wrong answers. They invent facts. They make shit up.

RAG solves this problem.

We connect AI models like ChatGPT and Google Gemini to a document filled with authentic facts so that the AI uses the facts to give correct answers, and not invent new facts.

So if you take ChatGPT to board exam, your answers will be wrong and you’ll fail.

But if you connect ChatGPT to NCERT books pdf, you will get full marks in board exam.

That’s the power of RAG.

I’m telling you about RAG because last week, Google took the next step in RAG.

They connected Google Search to Gemini.

So if you are a company and you want to add Google’s AI in your app, but you are afraid that it will invent a new fact about your company and say something bad, you can use this feature.

(Google calls this feature “Grounding”, but from what I’ve understood, its the same as RAG).

So before answering any user’s question, Gemini will first verify the answer through Google Search.

(if you want, it will even cite the websites from which it has taken the answer)

The result? Your AI app won’t invent new facts. The answers will be based on facts.

One more point to Google in the AI competition.

Follow Shreyans Singh for more insights like this.

Because of low viewers, Breaking Bad was about to be killed after three seasons. But Netflix saved it. Here's how:

In 2008, a TV series was launched, about a chemistry teacher who makes drugs.

It was called Breaking Bad.

The first season didn’t go well. It barely had a million viewers.

Same for 2nd, 3rd. season. The show was about to be cancelled.

But then they signed a deal with Netflix.

They put all 3 seasons on Netflix, just before the 4th season.

Then something happened.

Because when the 4th season came on TV, there were 2.9 million viewers.

And when the 5th season came out, there were 10.3 million viewers.

So what exactly happened 2 sentences ago?

On TV, you launch new episodes one by one. Every day, or every week.

The logic is that, you want people to wait for new episodes, to build suspense and create anticipation and hype.

Netflix changed this.

They put all episodes from S1,2,3 at once. No need to wait for a week.

And they encouraged people to watch all episodes immediately. Sit in front of a TV or laptop for hours and watch all episodes of Breaking Bad, before S4 arrives.

This concept is now known as binge watching.

So millions of people binged previous breaking bad seasons on netflix to get ready for the next season.

The hype before the finale episode in 2013 was crazy. Enormous binge watch parties were arranged in America. Hundreds and thousands of people sitting in front of giant screens, binge-watching breaking bad 24 hours before the finale dropped.

The result? 10.3 million people watched the finale, making it one of the most popular finales of all time.

Follow Shreyans Singh for more stories like this.

LinkedIn today announced that it is taking the goal of supporting creators one step further by acquiring creator economy platform OnlyFans.

The acquisition brings many synergies to the LinkedIn platform.

For example, LinkedIn already has the feature where they serve informative content with a complimentary selfie.

OnlyFans will strategically synergize this.

LinkedIn says they are planning a new feature where informative content will have complimentary HD content from onlyfans so that users can properly assimilate the information and insights.

LinkedIn will also expand their creator program: existing linkedin creators will be trained to create HD content. And existing onlyfans creators will be trained to write informative posts.

(linkedin says the disable button for the creator program is still in progress, so temporarily, it is compulsory for all creators)

This move will make linkedin the first super app in the social media space, disrupting even instagram and youtube.

Meta CEO Mark Zuckerberg and Twitter CEO Elon Musk both expressed shock at this news. Both their companies are reporting a massive fall in user numbers as people rush to LinkedIn.

More news as we get it, stay tuned.

Follow Shreyans Singh to get the news instantly.

Amazon is shutting down their AI technology for shopping inside stores. Because there was no AI. It was actually 1000 Indians manually scanning the camera footage to note down what items you bought and then they sent the bill.
Follow Shreyans Singh 

Self driving cars have started food delivery in America. 

These photos are of a city called Phoenix, where Uber Eats has partnered with Google's self driving car company called Waymo. 

(yes Google has a car company. It is a backup plan. In case ChatGPT disrupts them, Google will focus on selling cars)

The car arrives at your location, you have to go outside, do a phone verification and the car's trunk will open. Take out your order and enjoy.

Although I know it is still a pipe dream in India, but I hope Swiggy Zomato also think about cool innovations like this instead of green delivery t-shirts.

Follow Shreyans Singh for more.

The secret to Nvidia’s success in AI, according to Nvidia chief scientist, Bill Dally👇

⚡Instead of using really precise numbers in their AI calculations, Nvidia found they could use less precise ones without sacrificing accuracy. 

This means they could make their chips faster, smaller, and more efficient. 

For example, they started using a less precise number format called FP16 instead of FP32 for representing neural network parameters like weights and activations.

Nvidia H100, can even utilize 8-bit numbers for certain tasks.

⚡Nvidia redesigned their chips to handle big calculations in one go instead of breaking them down into smaller steps. 

This made a huge difference because fetching and decoding instructions took up a lot of energy. 

By doing things in a more complex but efficient way, they saved a lot of time and power. 

For example, using complex instructions like integer matrix multiply and accumulate (IMMA) helped minimize energy costs compared to simpler instructions.

⚡While Moore’s law helped Nvidia somewhat, they focused more on using the most advanced manufacturing technology available to consistently improve its GPUs, leading to incremental performance gains.

⚡After training, some parts of a neural network (the brain of an AI system) aren't really needed. 

Nvidia figured out how to remove these unnecessary parts to make the computations faster and use less energy. 

They introduced a method called structured sparsity, which helps make the calculations more efficient by cutting down on the number of things the chip needs to consider.

Follow Shreyans Singh for more insights like this.

nice weekend with friends in kodaikanal❤️

Taplio

Shreyans Singh's Linkedin Analytics

Shreyans Singh

What is Shreyans talking about?

Shreyans Singh's Best Posts (last 30 days)

Want to drive more opportunities from LinkedIn?

Taplio

.css-1jgreyo{background-image:linear-gradient(284deg, rgba(39,43,72,1) 0%, rgba(33,47,183,1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Shreyans Singh's Linkedin Analytics

Shreyans Singh

What is Shreyans talking about?

Shreyans Singh's Best Posts .css-qqfgvy{font-size:var(--chakra-fontSizes-md);}(last 30 days)

Want to drive more opportunities from LinkedIn?

Shreyans Singh's Linkedin Analytics

Shreyans Singh's Best Posts (last 30 days)