How to become an AI superpower?

Artificial intelligence is already transforming our lives, and things have just begun. The rate at which the technology is improving is hard to keep up with. Before this decade is up we will be living in a very different world. The whispers about an impending super intelligence or AGI (artificial general intelligence) have started to get shriller and urgent calls for aligning AI with human values has become louder.?

The cutting edge of AI is concentrated in a handful of companies (OpenAI, DeepMind, Anthropic etc.) with most of them located in the US west coast with some presence in London. OpenAI is famously backed by Microsoft , and DeepMind is part of Google . If we look closely at the investors of the other major players we will see commonalities. (And as usual, we have no idea what is going on in China)

If the goal is to build a safe super intelligence which aligns with human values, the question arises, whose values? Humans have very different value systems even within the same country, the same city, the same house.?

Let us say that a super intelligence is created in one of these labs. Who can say with surety that they will only be used for “good” and not “evil”. Depends on your definition of good and evil doesn’t it? You would think that some things are obvious and universal, like human rights and equality. But you would be wrong. There are entire populations whose value systems have very different definitions of human rights and equality.

Let me retrace a parallel (which many have already made) - the harnessing of nuclear energy and the creation of the atomic bomb vs a breakthrough in artificial intelligence and the creation of a super intelligence.?

In both cases there is the potential for profound impact across all walks of society, while ethical and moral implications abound. There is a race between powerful entities to be first and then dictate the direction of the regulation and control that will surely follow. In both cases there is the faint whiff of smoke in the air from impending doom.

I would argue, however, that the creation of a super intelligence is a much more pivotal moment in history. Nuclear weapons gave humanity the power to destroy the world and the responsibility to not do so. The creation of a super intelligence however would abdicate both that power and that responsibility into a non-human intelligence. The atom bomb never had to make policy decisions.?

Regardless of whether AGI happens soon or not, I think everyone in the field agrees that AI capability will improve further, and it will transform almost every human endeavor and industry. It is in this context, and as an Indian and a global citizen that I make the following statement.

India must prioritize becoming an AI superpower within this decade.

Why India? India is home to 1.3 billion humans, more than any other nation at any other time in history, and in all its flawed diversity it is arguably the most complete representation of humanity we have today within a single country. And while the country of India is just 75 years old, the idea of India has been around for millennia. Indian systems of thought, reasoning, philosophy and logic are in many ways unique and complementary. And unlike the big brother to the east, India is a democracy.?

But what is the definition of an AI superpower? Is it enough to have the most number of trained data scientists? Is it enough to have a million "AI" startups? Is it enough to teach courses in AI at universities? The answer to all of the above is a clear and emphatic "NO".?

OK, so if the goal is clear, how do we go about achieving it? I will attempt to lay down a roadmap of sorts of what needs to be done to meet the stated goal above. I will focus on the practicalities of achieving the goal, and will try not to veer off into the philosophical and existential morass that any conversation about super intelligence quickly descends into.

1. Act Now

There is no time to lose. There is no time to grow gradually and build the systems and infrastructure over decades to create an ecosystem the traditional way. There is no time for a planning commission and a series of 5 year plans.?

To follow the analogy from earlier about the discovery of nuclear energy and the building of the atom bomb, we are in 1940 and in a few years time some entity (most likely a private company, by the look of things) will have access to the most powerful object humanity has built. In the case of the nuclear race, the US got there first, establishing them as the global superpower for generations to come. It took the USSR a few more years to get there and the other powers even longer, by the time India went nuclear in the 70s, 3 decades had passed and the have’s were preaching caution and austerity to the have-nots.?

Now consider AI and the speculation about AGI - Sam Altman of OpenAI has already been on a world tour calling for increased regulation of AI. The closer we get to super intelligence, the louder the clamor for a comprehensive training ban treaty will get.?

So, while the government has to by definition think longer term, about the nation's place in the world and the future of its citizens, and while it might seem sensible to form planning commissions and think about growing a population of data scientists and AI engineers. Given the timeline we find ourselves in, it could all be a case of too little too late.?

India needs to act now.

2. Allocate an AGI budget

AI is already transforming entire industries, and a super intelligence could change the rules of the game entirely.?

All the money, resources, time and R&D you spend on defense, healthcare, science and other fields may be of little use if someone else has a super intelligence that enables them to be 100x better or faster or cheaper than you.?

It is the time to take a bet, a bet that if AGI or AI super intelligence arrives in the next couple of decades then India will be ready, and at the forefront of it. This will require significant political will, a leadership with the capacity to forecast the future with an open mind and very deep pockets.?

If the bet pays off, then any amount of money and resources spent will seem like a bargain.?

If the bet does NOT pay off and after a decade or two, we merely have more powerful versions of today’s AI, that is still a very valid use of time and money.

3. Leverage the diaspora

The Indian diaspora is famously embedded right into the heart of the technology industry. The CEOs of Google, Microsoft, IBM, Adobe and so many more companies were born in India. And Indians form a significant chunk of engineers and researchers who are building the cutting edge of AI. Just that they are doing it outside India, for non-Indian companies.

To precipitate a small reversal in this brain drain, and to entice at least some of the highly skilled diaspora to work for India - is not easy. The Indian government would need to take radical steps.?

One idea is to create an Indian AI service, analogous to the Indian Civil service. The hiring process should be similar to a top tier international AI startup or research lab. The pay should be lucrative, it does not need to match what is being offered in Silicon Valley as cost of living and quality of life conditions in the rest of the world vary wildly. And for most Indians, at least those with a patriotic inclination, the prestige of nation building will be a powerful motivation and will likely compensate in part for some of the lost salary. There are also thousands of non-Indians in Europe and other parts of the world who are waiting for a grand AI project they can contribute to but are currently being excluded based on their geography, which brings us to…?

4. Think and act internationally

To succeed India will have to look beyond nationalism and parochialism. Talent is available all over the world, this much is obvious to any pragmatic rationalist. The US became the beacon of technological progress over the last 100 years by selling the American dream and welcoming those who wanted to change their fortunes by changing the world.

But the pragmatic rationalist should also understand that despite all of its advantages, it is not rational to expect experts from all over the world to flock to India. But - they don’t “physically” need to. We live in a distributed world today, and despite tech CEOs wanting everyone to work from the office for building “culture” and sustaining innovation through water cooler conversations. The truth is that a lot of people would rather work from wherever they like to.

This is an opportunity. Not everyone will be able to or want to relocate to India to work on AI. But they don’t have to. They can work from wherever they want to.

5. Harness the velocity of a startup

Why do startups with a handful of people disrupt huge corporations that employ hundreds of thousands of people??

Every extra layer of communication incurs a cost - a cost in terms of round trip time delay for each decision to go up the chain and down the chain, and an even bigger cost in terms of information loss at each step.?

To get things done fast, the team must be condensed and empowered.?

Define clear, actionable goals with a fixed timeline, allocate a budget and plan for periodic checkpoints to ensure the project is on track. But beyond this there should be autonomy in terms of how the project is run. If you make sure you hire those with the right skills and motivation, and that they fully align with the vision - then you should show the courage to trust them to get the job done.

6. Create the largest, most diverse training corpus ever

India is home to a millennia old culture, thousands of languages and the largest epics known to man. If ever there was a nation who could put together the most complete training corpus known to man it is this.?

The majority of the most powerful AI models today have been trained on text available on the internet which is mostly English. And yes the Chinese AI have been trained primarily in Chinese. And while these models do understand other languages I would not call them truly multilingual.?

Could multilingualism in AI unlock cognitive benefits? If yes, then there might be new frontiers to cross and worlds to discover.?

There is an immense wealth of language available in an unorganized mess of written text, oral tradition and video and audio recordings. The task of collecting this data and organizing it in a useful form is gargantuan. But India has taken on many ambitious projects and made them successful in the past.?

Once the data is available new techniques will need to be discovered to embed all this information optimally. The current tokenization mechanisms are notoriously expensive to encode non-Latin alphabets.

7. Run the largest RLHF exercise the world has ever seen

What is RLHF? Huggingface has a nice explanation on their blog .

Wouldn't it be great if we use human feedback for generated text as a measure of performance or go even one step further and use that feedback as a loss to optimize the model? That's the idea of Reinforcement Learning from Human Feedback (RLHF); use methods from reinforcement learning to directly optimize a language model with human feedback. RLHF has enabled language models to begin to align a model trained on a general corpus of text data to that of complex human values.

RLHF's most recent success was its use in ChatGPT .?

RLHF when done well can lead to significantly better performance and quality in terms of generated responses.

This is where India can leverage its huge English speaking, technically educated, computer literate population. To create a higher quality, more diverse, more complete RLHF than anything that has been done till date.

8. Seed an Indian Cloud

If all your AI innovation happens on AWS (Amazon Web Services) or Microsoft Azure or Google Cloud, then the question arises of how independent you truly are. Geo political decisions taken in a foreign country could lead to you using access to all of your “compute” instantly. The bigger concern is that all of this is reliant on GPUs primarily made for Nvidia by TSMC (Taiwan Semiconductor Manufacturing Company Limited) – please note the giant dragon shaped shadow being cast on this discussion by China. And if the global AI supply chain is starting to look like its held together by sticky tape and post it notes, then wait till you hear that the machines TSMC uses to build the chips are all produced by a single company ASML

This is the pitfall of the cloud. It feels robust, until it isn’t. It feels reliable, until it isn’t. It feels secure, until it isn’t.?

India needs a strategy to counter these threats, and cannot afford to be reliant on compute resources that can be taken away at the whim of foreign governments. But this is not going to be easy or quick, building self reliance will be a painful, expensive and lengthy process.?

What can be done in the short term? Well here’s an idea. Create a national private cloud, with geo redundant data centers spread across the nation. Purchase the necessary GPUs from Nvidia to do this.?

There is no alternate Indian cloud provider but building a dedicated GPU cluster and training environment is definitely do-able and possible.?

Sounds impossible? I'd suggest reading the story of how Google Cloud has been catching up to AWS despite its late start. And yes, the guy who leads Google Cloud, Thomas Kurian , was born and grew up in India (See point #3).

9. Tap into the immense potential of India's tech ecosystem?

Let's start with academia, focusing only on engineering students for now.?

The total engineering seats in institutes of higher education across India is slightly over 2 million. But lets ignore this massive number for now, not everyone makes it through 4 years of engineering ready to graduate at the end of it. Just considering the premier IIT (Indian Institute of Technology) and NIT (National Institute of technology) colleges, there are more than 30000 new graduates every year. To this let us add the cream of the crop of the remaining universities and round up the number to 50000 every year.?50000 highly skilled engineers who can help with all the different aspects of this giant project, the majority of whom would jump at the chance to work for the Indian AI service.

India also has a burgeoning startup scene, and of course the multi billion dollar IT majors who need to be decisive to avoid becoming the dinosaurs of this AI impact story. Why would these companies contribute financially, through findings or taxation or through other resources for such a project? What's in it for them?

Consider, for example, the Indian AI project can create an AI model equivalent to GPT-5, then all the Indian companies and universities and the government could have priority, free, unfettered access to it. Think of the possibilities. Think of the savings from not having to pay the OpenAIs and Google's and Microsoft's of the world.?

10. Focus on multidisciplinary innovation?

So far I've been referring to transformer based large language models that can be conversed with, such as ChatGPT, but of course this is just one kind of AI.?

I do not think we will achieve super intelligence by throwing more data and compute at this particular architecture.?

There are other path breaking AI models such as Deepmind's AlphaFold and AlphaCode that are revolutionizing entire fields of science and solving problems long thought impossible.

And there are yet many more undiscovered breakthroughs which could shift the balance of power further. The best way to arrive at these breakthroughs is to try to solve the really hard problems that matter immensely across different domains. Problems in different spheres such as agriculture, healthcare, tackling poverty and pollution will all require unique and possibly bespoke solutions. And to find gold you need to attempt to solve problems that appear intractable. You need to become familiar with defeat. The problems have to be immensely meaningful to attract the extraordinary problem solvers necessary.

The new research and technology that arises out of this smelting pot could yield the key to unlocking super intelligence.

***

Of course there is a lot more that can be done to achieve this goal but I think that the 10 point plan laid out here is as good a place to start as any. And each of these actions will bring about huge benefits for the people, the businesses and the economy of the nation regardless of whether super intelligence is achieved or not.

I'm certain that the Indian government, technology companies, universities and various other industry organizations are thinking of this. And that they have plans in motion. But I do not think that the urgency and gravity of the matter is yet clear and evident to the decision makers. The AI revolution is not going to be like the industrial revolution or the computing revolution. The stakes are higher. And you don't want to catch the televised rerun of the revolution. You want to be in the room where it happens.?

I would love to hear feedback and criticisms of this roadmap. The more we discuss and debate, the more chances there are of elevating this subject to the attention of the decision makers. So, please share this widely if you agree in spirit with what is expressed here. And if you don't agree then please vehemently disagree.?

The fence is worn with use, and the time to sit on it has long passed us by.?