Blog

Geoffrey Hinton, the Nobel laureate and Godfather of AI is 77 now, and is glad that his fear of AI overtaking the world in future will not come true in his lifetime. He foresees the potential dangers of AI.

AI advances could be compared to a tiger cub being reared by human beings, without realizing that once fully grown the cub as a full-fledged tiger can kill you.

There is a 10-20 per cent chance of AI systems eventually seizing control but it cannot be predicted precisely.

The emergence of AI agents which perform the tasks autonomously also is a matter of great concern.

Superintelligence would be 20 years away, but the rapid AI advances can shorten this time span to 10 years or even less. They are all after the next shiny thing.

Hinton resigned from Google in 2023. He is now Professor Emeritus, University of Toronto.

Google and DeepMind merged in 2023. DeepMind’s scientist Murry Shanahan is not able to resolve a dilemma he faces. He describes LLMs as ‘exotic mind-like entities’. AI in future does require a new vocabulary.

What is the logic behind this diction? Our vocabulary falls short while describing emerging digital intelligence. It is mind-like but not ‘mind’ as we have humans have. To hedge the bets, we have hyphenated it — mind-like. LLMs do use language the way we do. Still, they exist without physical embodiment. A new concept of self-hood emerges. It is eerie and unfamiliar.

Muiray’s observations carry weight. DeepMind founded in 2016 is in the forefront of AI research. AphaGo has scored a victory against a world Champion. AlphaFold predicted protein structures. DeepMind blurred the line between machine and mind. The merged entity Google DeepMind has done valuable AI research. It has trained neural networks to master chess, Go and Shogi. It has trained networks to solve puzzles of molecular biology.

Murray is a professor of Cognitive Robotics at Imperial College London and a senior scientist of DeepMind. He has studied the intersection of AI, cognitive science and philosophy. He advised in the making of Ex Machina (2014 film) to navigate such nuances.

Defining intelligence in entities that are not human is challenging. He calls LLMs exotic. It shows his deep understanding. The machines are evolving rapidly, but we lag behind in understanding them. As AI gets integrated into our lives, it will be prudent to forge a new conceptual framework of intelligence, self-hood and consciousness.

As it is, we are creating minds we can barely define, let alone understand them. Humanity is on the threshold of a philosophical frontier. We are searching the diction for companions we have created from code.

AI has witnessed many hypes since the 1950s. The term was coined in MIT in 1956 by John McCarthy. There is steady progression for the last 15 years when databases became data warehouses. The descriptive data analysis and presentations were in static forms. These adopted machine learning techniques to move to predictive and prescriptive analytics. This represents the first wave of AI and algorithmic models. Algorithmic decision-making using big data created prediction models and applications such as customer service and supply chains. Further, the weather forecasting models and traffic navigation systems facilitated our lives.

Google’s pathbreaking 2017 paper titled Attention’s Is All You Need proposed a new Transformer architecture. It facilitated the computer’s understanding of human communication. The ‘attention mechanism’ focused on AI attention on the most relevant parts of a text. It paved the way for LLMs which can take any input token like a sentence and predict the next token. It gave birth to a revolution called generative AI. ChatGPT and other large compute-based LLMs ruled AI till DeepSeek, a Chinese model came up which did justice to the job more economically.

India has assigned the task of building its own muti-modal, multilingual, AI model to Sarvam, a Bangalore-based startup with the backing of IIT, Madras.

Sarvam, a Bangalore-based startup, has been assigned the task for building an LLM for India. Sarvam in Sanskrit means ‘all’, and it implies that the proposed LLM will encompass everybody.

The startup was founded by Vivek Raghavan and Pratyush Kumar in July 2023. They had in mind to create an LLM that can be used by a billion people. The arrival of DeepSeek, the economical Chinese model that was competent enough, fueled the ambitions of Sarvam founders.

Previously, Sarvam founders worked for AI4Bharat, a research initiative by IIT Madras to develop open-source Indian language AI. Apart from NLP, the model should have speech recognition capability, transliteration capability, converting texts of Indian scripts to English and speech synthesis ability.

AI4Bharat was a collaboration between IIT Madras and EkStep of Nandan Nilkeni. Raghavan worked here as chief AI evangelist.

At the same time, Raghavan worked as a knowledge partner for National Language Translation Mission — Bhashini. Raghavan facilitated the development of Bhasaverse, an app to do speech-to-speech translation across 11 Indian languages and text-to text translation in all 22 Indian languages. The work ultimately led to Bhashini app. Raghawan had also spent more than a decade simultaneously at Aadhar.

Kumar too worked for AI4Bharat and joined as an adjunct faculty at IIT Madras in 2021. Kumar had a doctorate from ETH Zurich and was an IIT Bombay alumnus. He had a research stint at IBM and Microsoft. He was a co-founder of Sarvam AI that proposed to build an indigenous LLM from scratch. The initial funding came from Peak XV and Lightspeed Venture.

The building of LLM is resource-intensive process. They initially worked with open-source models and fine-tuned them on Indian datasets. Later, they developed voice agents. They launched Sarvam-1, a 2-billion parameter model, trained on four trillion tokens in October 2024. It supports apart from English 10 Indian languages.

The government has now assigned them a task to build a 70-billion parameter AI model optimized for voice reasoning and fluency in 22 Indian languages. More parameters enable the model to learn more complex patterns. It also requires more data and computational power. The government will provide Sarvam access to 4096 Nvidia H100 GPUs for six months.

Sarvam is also working with NITI Aayog to develop a pilot for Enterprise Reasoning Engine (ERE) on the National Data and Analytical Portal (NDAP) to enhance portal’s data accessibility and usage.

Sarvam has also deployed AI solutions to enhance user experience of Aadhar services, including voice-based interactions.

The next six months are critical for this pair of Raghavan and Kumar as they try to develop a foundation model. It makes a defining moment for India since an Indian startup is not tinkening at the edges of AI but owning the full stack.

AI has deeply influenced the fields such as graphic design, movie making and computer programming. However, AI has potential to influence the hard sciences too.

David Silver, DeepMind researcher, has interestingly predicted the role of AI in the future of mathematics. AI can solve math puzzles.

The Clay Mathematics Institute in 2000 offered a million-dollar prize for seven different mathematical problems. The time frame was a quarter of a century. However, just one mathematician accepted the challenge. Perhaps, the prize could go to AI.

AI has scored impressively at Math Olympiad. Initially, AI was not good at math. Previous LLMs struggled with basic arithmetic and could not count the r’s in strawberry. Even today LLMs struggle to multiply large numbers. LLMs do not have a mathematical model. They go through previous mathematical operations in their training data.

LLMs, however, are improving in math. GPT O3-has scored well on the AIME — a feeder exam to Olympiad. AI could soon win a medal at Olympiad.

AI, according to Silver, DeepMind, will revolutionize math. It can percolate down the downstream fields such as physics. Math is considered the purest science. The ripple effect could facilitate all scientific progress in near future.

A Bangalore-based startup Sarvam has been chosen by the government to build India’s first LLM after a scrutiny of the 67 applications. The government will provide it compute resources. It will get incentivized by Rs. 10000 crore IndiaAI Mission. Sarvam model will be having reasoning capability, will respond to voice, and will be fluent in Indian languages. It will get access to 4000 GPUs for six months for the company to build and train its model.

It will not be an open-source model but could be fine-tuned to Indian languages. It will have 70 billion parameters, and many innovative features in engineering and programming. It will be in a position to compete with some of the best models in the world.

There will be three variants– Sarvam Large for advanced reasoning and generation, Sarvam Small for real time interactive applications and Sarvam-Edge for on-device tasks.

It will be optimized in India using local infrastructure and talent.

Sarvam’s goal is to build multi-modal, multi-scale foundational models from scratch.

The development occurs after the arrival of the Chinese DeepSeek model.

Those Indian companies that provide GPU support are Jio, Hiranandani-backed Yotta, Tata Communications, E2E Networks, NxtGen Datacentre, CMS Computers, Ctrls Datacenters, Locuz Enterprise Solutions, Orient Technologies and Vensysco Technologies.

There are certain challenges. India has to market and monetize a closed-source model in competitive global space. Sarvam’s proprietary approach aims at strategic autonomy and enterprise appeal. There is monetizing potential through subscriptions or application programming interface (API) access. However, the global market is very competitive. OpenAI was losing money on ChatGPT Pro subscription due to unexpectedly high usage outpacing the $200-per-month pricing that was set. This underscores the challenge Sarvam faces.

Indian users of OpenAI are the second-largest bas. They overwhelmingly prefer foreign models. There are issues of transparency, biases and data privacy, especially in sensitive sections like healthcare and finance.

Space investigation was initiated in India by Vikram Sarabhai way back in 1968. He expected ISRO to work for the benefit of common people and help in providing solutions to the country’s problems. He never fantasized exploration of the celestial bodies or manned missions. Sarabhai considered space technology as a tool to fulfill development requirements. He was aware of the information about our critical resources and communications.

Sarabhai passed away in 1971 and his successors at ISRO were Prof. MGK Menon, Satish Dhawan, VR Rao. They continued to work on his vision. ISRO built capabilities in remote sensing, communication, broadcasting, meteorology, earth observation, satellite technologies. UR Rao left in 1994. By that time Sarabhai’s vision has been fulfilled.

India expended its space programme. Later. ISRO has turned into a space exploration agency. ISRO’s Kasturirangan (1994-2003) undertook the moon mission. India faced international controls which became worse after the 1998 nuclear tests. India was denied the crucial cryogenic technology. India created in-house capabilities.

As we know, there are many accidental discoveries in science. Penicillin was discovered accidently when Alexander Fleming noticed an inhibition zone where there were no microbes on an agar agar plate, since something carried by wind had fallen over it. Even Microwave oven is an accidental discovery. Modern LLMs or large language models might fall in the same territory.

OpenAI employee who built GPT-1, the version of ChatGPT, did not quite understand how it worked and why it worked. But he was impressed by the results. Later, teams worked on GPT-1, and made it bigger and better.

Finally, GPT-3 was released as ChatGPT in November 2022. This event triggered an AI revolution. This is a common thing in the history of technology. The initial lack of understanding did not deter the team to work on it. The empirical approach paved the way for further scientific enquiry. Its genesis is in empirical result first.

Sometimes the technological progress is haphazard. It can emerge out of ‘messing around’.

An environment of experimentation and embracing unexpected results must be fostered. The initial success could be based on observation but later the focus shifts to scaling laws and scientific investigation. It is an iterative process. It is a crucial lesson for future. There should be both intuitive experimentation and rigorous scientific analysis to drive tectological breakthroughs.

Once upon a time, out-of-home (OOH) advertising was synonymous with static hoardings and colourful tinplates. Formats like billboards, unipoles and gantries continue to deliver strong recall in metros and tier II and tier III cities. Since then the advancing technology has disrupted this medium.

Brooke Bond’s Taj Mahal tea. It displayed Megh Santoor billboard at Silver Cannes Lions 2024 in the outdoor category. The raindrops fell on the huge santoor and created a symph, resonating with the raga of rains. In Lucknow and Kolkata, an empty plate was transformed into a platter of pakodas in a Fortune hoarding. In both the instances, technology has transformed the out-of-home into digital out-of-home (DOOH).

The OOH market stood at Rs.5920 crore. Out of this, digital commands a share of 12 per cent at Rs.700 crore. The DOOH is likely to grow at a CAGR of 24 per cent. It will then all out for 17 per cent of all outdoor advertising revenue by 2027.

The urban centers are driving the growth of digital transformation. But it is spreading to many more centers. The top six metros account for 80 per cent of digital screens.

There are advantages of digital outdoor. The creative can be changed quickly. It has the ability play video and animations. The off-line messages can quickly transition to online. And the significant game changer is programmatic DOOH. It makes advertising effective and facilitates targeted messages. The messages could be location-based and real-time. They are contextually relevant.

Advertisers can leverage micro-moments using weather, time, traffic density or audience data to service the most relevant massages.

Though the medium of DOOH has grown in India, it has yet to see the kind of growth it has in markets such as the US, UK, China and others. Globally, DOOH accounts for 56 per cent of OOH advertising.

There are several challenges the outdoor media faces — there is fragmentation of the medium and its ownership, lack of standardization in terms of pricing, measurement and quality of displays, inability of advertisers to plan and execute large scale OOH and DOOH campaigns.

There are local municipal authorities that regulate OOH. There is an issue of measurement — there is no unified measurement and real-time audience tracking.

Nvidia will not be able to sell its customized AI chips in China. Both the company and the markets have been caught by surprise. The chips have been turned into a bargaining point. Export controls are ineffective since it stimulates the domestic supply chain.

Of course, China desperately needs AI chips. Export controls could hold back China in the AI race, Access to chips and computing power has been at the core of Silicon Valley’s lead over China in AI but that gap is closing fast. Chip restrictions are porous. These are ineffective without international co-operation.

China is not far behind in AI software and DeepSeek has proved it. America’s key lead now remains in hardware, and that too in advanced chips. Huawei and Semiconductor Manufacturing International Corp are working to produce domestic alternative to Nvidia’s processors. Even the local players are catching up.

Chip war has pre-dated the trade war. The US may lose both. These chips should not have been used as pawns.