2025-01-29 08:00:00
There are three ways to fund an ambitious project like a foundational AI model:
Free Cash Flow from an Existing Business: Xerox PARC was a hub of innovation in the 1970s, pioneering cutting-edge technologies (though it did not commercialise them). Many companies, including Apple, took inspiration (stole) from PARC. It attracted some of the brightest minds in computer science, actively recruiting from top universities. Today, DeepSeek has emerged from a quantitative hedge fund, leveraging its financial resources to develop AI models. Some speculate that DeepSeek is backed by the Chinese government, but assuming it has sufficient GPU resources and follows the past research from OpenAI’s, its progress is plausible. While Google, Microsoft, and others are investing heavily in AI (e.g., Gemini and Phi), India lacks trillion-dollar companies capable of independently funding foundational AI research, ruling out this option.
Venture Capital (VC) funding: Venture capital works well in the US, where investors take higher risks and operate under a portfolio strategy, rarely investing more than 10% of their fund in a single company. Large-scale AI projects require mega funds to co-invest—OpenAI and Anthropic are prime examples.
Since Indian VCs are more risk-averse, they tend to invest in startups that have a path to a liquidity event, chances to raise follow-on capital from bigger US funds (as well as sovereign funds). US funds and sovereign funds won’t invest in Indian foundational models because they can easily deploy in market leaders like OpenAI and Anthropic. They have global TAM, lead in usage, and have the willingness and ability to build better frontier models. This is why Sarvam is the only model that has a chance because they have come up with their own model. They have one co-founder who was the Chief Project Manager of Aadhaar and another who was teaching GPU programming at IIT Madras in 2018. They are probably the only ones who have a chance to make it as a VC-backed Indian foundational AI company. Even then, when I meet VCs, a lot of them question if their TAM is big enough (Indic languages focus vs English where global players can win). I would bet they will be the ones working with Government of India, like Anduril and OpenAI in the US. Krutrim, I will ignore. You can’t compete with OpenAI and Deepseek when it is not your major focus.
What will it take for us to get there?
It will take time, but we have to start somewhere.
Someone asked the following question on Twitter: “This is probably ignorant since I dont really understand the working on the indian chaebol but is there a reason Adani/Ambani/Tata/Infosys couldnt foot half the bill - with the gov doing the other half - to train a frontier model?”
My response to that was: “Good point. A joint project between one of these companies and the Indian Government can happen because they are close to the government and the Government wants to work together on this. However, for most service companies, it makes more sense to improve existing free AI models and use them for their clients. Making products has never been their strength. If they were good at making products, they would have already built products to compete worldwide using their extra money. Risk taking is also limited in these service companies.”
2025-01-28 08:00:00
Most US government contractors make a lot of money. They are on cost-plus contracts: Cost + Margin.
This means OpenAI through ChatGPT Gov could have made a lot of money by working with the Government over the next few years.
However, Deepseek arriving now is bad timing for them.
Imagine finishing deals with Biden and Kamala, then Trump and Musk win, with Musk wanting to reduce Government spending, and then Deepseek also arrives with their training costs. Not a great week for OpenAI especially when they would have wanted to celebrate their ChatGPT Gov launch.
P.S I learned about US defense contractors and their margins by listening to many hours of 20VC podcast episodes about US American Dynamism startups. Take a look at how well US defense contractor stocks have performed in the last ten years to verify how good of a business defense contracting is.
2025-01-27 08:00:00
“Oh, Deepseek is pushing their app instead of focusing on their web and API. It is a sneaky way to collect more user information.” - X dot com.
There is a reason to push the app.
The most important design principle is: Out of sight, out of mind.
For example, I pay for both ChatGPT and Claude.
ChatGPT is a native Mac app. While Claude’s desktop app is a very buggy Electron app that crashes and freezes frequently, so I use Claude only on the web and phone.
As a result, I almost never use Canvas (Artifacts) on ChatGPT because it is only available on the web version. The user experience is also poor. You have to select it, unlike Claude where it appears automatically in the right sidebar.
This shows why Claude has Projects even on the mobile app, even if people might not create projects on mobile. When you open the sidebar on the mobile app, you see Projects listed.
My ChatGPT usage is far higher, even though Sonnet is a better model for coding or writing because I work on my laptop. I just open ChatGPT more often because it is easily accessible in my Mac’s dock.
The only place I use Sonnet more than the latest GPT models is in Cursor, where it is selected by default.
Likewise, I would probably even use Google’s AI Studio more, but they don’t have a native app.
Not everything is a conspiracy. Sometimes it is product management 101.
You need to reduce the time and steps to get value.
2025-01-26 08:00:00
The last decade has demonstrated how the United States, as the world’s leading power, can exercise significant economic control against nations it thinks of as competition. They removed Russia from international payment rails. Similar pressure has been applied to China through various measures: new tariffs, product bans, and other restrictions, marking an escalation of tensions between the two nations. The new Cold War.
China is also not far behind when it comes to exerting global influence. See what they are doing in Africa. And now that India is trying to catch up in manufacturing they are doing this. As AI becomes increasingly crucial over the coming decades, nations must consider their technological sovereignty. Owning and controlling foundational AI models becomes a necessity.
The key is developing AI models trained on a nation’s own data. See Sarvam’s approach. You need to align your model to your values and perspectives. I am sure the Chinese and US models have been fine-tuned to give different answers to particularly sensitive topics. See the screenshots below. Yes, the question is controversial, but both ChatGPT and Claude give a balanced answer while Deepseek refuses to respond.
Nations need to invest in developing their own AI models that align with their national interests and strategic objectives.
Yes, Deepseek is amazing, and I am happy how open source is catching up. I love using Advanced Voice on ChatGPT and Claude for my random side projects.
But if I was Modi, I would start a Manhattan Project scale project in AI. It is far far important to not do so.
2025-01-25 08:00:00
“Stupid question why would users on ride hailing apps be high intent and on food delivery apps be medium intent”
Someone asked me the above on x dot com. Here is my response:
Yesterday I had to go to Koramangala. I don’t have a car. So I opened Namma and booked an auto. I could not decide to go two hours later in case I got a better deal than that time. I could not go to Marathalli because it was probably showing a cheaper ride. Yes, I could open Rapido if I saw the price on Namma being too high, or there were supply issues. But my destination will not change. My timing will not change. I cannot decide not to go if there is a higher price when it is busy.
Compare this with food. How many times have you decided not to order because you felt the price was too high and you were not getting good value for money. There is tiredness from making choices. Oh, I selected chai, but it does not meet the minimum order value, so now I need to add more items. Do I add a samosa? Maybe I don’t want to eat a samosa. Maybe I will skip snacks and then order dinner instead. I do not have any offers. I see some random packaging charge and now I feel the total is much higher than the menu price. Maybe I just went to the place myself yesterday, and the items are marked up 100%, and I feel I can write a tweet about food delivery being absolutely the worst model that does not serve anyone and get a few hundred likes instead. And so on.
From ride-hailing to food ordering to ordering a t-shirt that I don’t need, but if I get a good deal on the shopping websites I will buy, there is a range. If the purpose is strong, it is easier to get people to buy.
You can compare how many people actually buy from ride-hailing versus food delivery process (I can’t share both due to obvious reasons) versus shopping websites (2%) and realize how much purpose matters.
2025-01-24 08:00:00
“Genuine q: do many people see/interact with the banner on an app’s homepage? Doesn’t seem like a great use of prime real-estate. For me whenever I open an app it’s a muscle-memory scroll/search to where the useful stuff is.”
I saw someone post the above on x dot com. Here are my thoughts having worked on on demand marketplaces for 6 years now:
A significant number of users in low to medium intent products are “promo hunters”.
Remove all banners around affordability and price perception and you will start seeing behavior change. Their frequency of ordering drops.
Start polling these users a few months later and they will say “competitor x has far better price”, while when you compare final price on checkout for both they will be more or less the same. It is the discounting/ promos that drive that perception.
Giving a good deal is important for conversion.
Even more important: Making users feel like they are getting a good deal.
Users want to feel that they are getting a good deal from your platform.