Author: Shabbir Chunawalla

At the historic AI Action Summit in Paris, the US and UK did not sign the declaration on ‘inclusive and sustainable AI.’

America was sure that the world would have to rely on American Big Tech industry for AI, until the arrival of DeepSeek from China. It was believed that AI innovation require hefty investment, and Trump’s first announcement was a Stargate programme involving an investment of $500 billion. What China did was to make AI innovation at low cost. In fact, it is a first step towards democratizing the spread of AI applications and capabilities. Such democratisation benefits more people from AI capabilities.

So far, there was dominance of proprietary AI models. DeepSeek’s debut marked seismic change. It lowered the entry barrier by being cost-effective. It fosters inclusive AI.

Many other nations in Europe, Asia and Australia would follow suit in introducing low-cost models. Less developed countries will have other options to purchase AI services.

It is a paradox that China, a nation devoid of democracy, has stimulated democratisation of AI.

Many nations advocate their sovereignty-first AI. Sovereign AI refers to investments by nations to gain control over vital technology like AI. Post DeepSeek era, the concept has been reignited, although the investment requirements are lower.

To sum up, ChatGPT and DeepSeek are change points in the history of civilization.

Since the arrival of ChatGPT in late 2022, there has been an AI boom. The whole market has become an oligopoly with a few players such as Google, OpenAI, Meta and Amazon.com. In fact, the real competition is between Google and OpenAI, while Meta and Amazon are marginal players.

Many startups have been swallowed up by the tech giants — Inflection AI by Microsoft, Adept by Amazon and Character AI by Google.

Google and OpenAI’s duopoly have been challenged by Perplexity AI – an answer engine with 15 million regular users. Over the previous year 2024, its valuation at the beginning of 2025 rose to $9 billion. It is headed by Arvind Srivastav who has entrenched himself deeply into Silicon Valley’s ecosystem.

Google’s monopoly of search engine and advertising is formidable. Almost 90 per cent of searches online are conducted on Google. Microsoft’s Bing with a 4 per cent share has hardly dented Google’s monopoly. Yahoo’s usage too has declined. Neeva, a startup of a former Google employee has closed down in 2020 after trying hard to dent Google’s share.

Perplexity has, however, survived and has raised funds. Its answers are clean and comprehensive. Google carried the ad clutter. Perplexity is making money through its $20 a month subscription model. Its annual revenue is $30 million. Perplexity is also entering Google’s lucrative business of advertising. Google’s 80 per cent searches are non-monetizable. Many of these searches do not directly generate ad revenue. These are maintained to command leadership. Perplexity’s informational queries will lead to purchase decision.

Perplexity’s Srinivas is a diplomatic head with a string of Silicon Valley relationships. His company was visited by Elon Musk. He has got a lean team of 147 employees. His investors are well-chosen — Amazon founder Jeff Bezos, AI scientist Yann LeCun and AI investor Nat FriedMan.

Srinivas has withstood the cut-throat competition of Silicon Valley as he has got a good grasp of Silicon Valley’s political economy. He has interned with Google when Google Brain was putting together the transformer architecture. He understands the technical foundations of the present day AI and the complex web of industry’s relationships. He maintains cordial relations with rivals and diffuses tensions with the publishers. He knows how to navigate the landscape of AI.

Google’s search business stands firm, but it now faces a credible outsider who knows how people find information online.

LAMs are considered to be taking us closer to AGI, since these models not only respond but also act upon instructions.

So far, we have impressed upon you the utilization of large language models or LLMs. LLMs power our chatbots and give us the generative AI. Microsoft researchers have created a Large Action. Model (LAM) which can operate Windows on its own. It is a significant advancement in AI. These systems execute complex tasks based on human instructions. Users take user requests into real actions. The action could extend to operation of a software that controls robots.

LAM has been trained to work with Microsoft Office. It got prominence in the early 2024 when Rabbit AI device was launched. Rabbit AI could interact with mobile apps without the need of the user. LAM model can understand inputs in the form text, voice or images. These requests could be converted into detailed step-by-step plans. LAMs in short not just understand but act too.

These models interact with both the physical and digital world. It can create a PowerPoint presentation. ( It opens the app, creates slides, formats them as per preferences). LAMs understand essentially intent and generate action accordingly. It is a dynamic adaptation.

An LAM is built by using data — task-plan data and task-action data. Task action data spells out doable steps. These models undergo training of supervised fine-tuning. These models are tested in controlled environments before deployment. They can be integrated to AI agent systems. It is a big leap. LAMs can be useful to the disabled and can automate workflows. LAMs on further evolution can become a standard AI system for all sectors.

As we have observed already, anti-obesity drugs such as Wegovy and Ozempic (made by Novo Nordisk) and tirzepatide (made by Eli Lilly) have made waves the world over. They are set to enter the Indian market in 2025.

Both the companies intend to price their drugs competitively for the Indian market.

Novo Nordisk planned to bring its anti-obeslty drugs to India in 2026, but could prepone it to match the Eli Lilly’s move to bring its product a year earlier, say 2025.

Many celebrities too have used these weight losing drugs, though they generally deny taking these medicines.

Domestic companies in India too are planning to introduce anti-obesity drugs. Sun Pharma is working on utreglutide. Cipla intends to introduce a generic version of Wegovy. Dr Reddy’s Laboratories too would like to introduce a generic version of Wegovy. Biocon is working on GLP-1 therapies and has received approval for liraglutide. Natco is developing a generic version of Ozempic.

The interest in this area could be attributed to WHO’s assessment that one in eight people in the world lived with obesity and obesity is the main cause of diabetes (44 percent).

In India, according to a Lancet study, 44 million women and 25 million women over 20 years of age are clinically obese.

People from developing countries try to obtain anti-obesity drugs through grey market or travel to foreign countries to access them. This could be called ‘Ozempic tourism.’

Health insurance policies in India exclude obesity treatments. In the absence of insurance, these drugs will be prohibitively expensive — almost $1000 a month. In future insurance may cover the anti-obesity treatment as obesity is recognized as a chronic condition with several co-morbidities.

The existing anti-obesity drug choice in India is poor — patients take orlisat, sibutramine, etc. These are not effective. These have lot of side effects. Therefore, the new GLP-1 pathway is a better choice.

Those who would like to work in ML and AI field should build their math foundation. First, they should learn linear algebra. Here they should learn matrices and vectors. Matrices are grids of numbers. Vectors are lists. Data is often stored this way. There are operations like addition, multiplication and dot products.

Next, we should learn Determinats and Inverses. Determinants tell you whether a matrix can be inverted. It is used in optimization problems and solving systems of equations.

One more area — eigen values and eigenvectors which make us understand the variance in data. They are foundation of Principal Component Analysis. It helps us reduce dimensionality in data sets.

Lastly, one must learn Matrix Decomposition such as Singular Value Decomposition (SVD) used in dimensionality reduction and data compression. Let us turn to basic calculus. It is core to our understanding how models learn from data. First, learn derivatives and gradients. Derivatives measure how things change. Gradients are multidimensional derivatives. These are power optimization algorithms such as gradient descent. Models adjust their parameters to minimize error with the help of gradient descent.

You know the Chain Rule is central to neural networks. It enables backpropagation to work. It is a process to figure out how much each weight in the network contributes to the overall error. This makes the model learn more effectively.

Lastly, learn optimization basics — concepts such as local and global minima, saddle points and convexity.

To be competent in this field, you should learn statistics and probability.

Data is understood through statistics and probability. You should know Distributions (normal, binomial and uniform). You should also know variance and covariance to know the spread of the data and their correlation respectively. Bayes Theorem is a tool for probabilistic reasoning. It is foundation for algorithms such as Naive Bayes and Bayesian optimisation. You should also understand Maximum Likelihood Estimation (MLE) to estimate model parameters by detecting values that maximize likelihood of our data. It is basic to logistic regression.

Finally, you should know sampling and probability.

The resources for Linear Algebra and Calculus are 3 blue 1 brown’s Essence of Linear Algebra and Essence of Calculus series. To get a grasp of statistics and probability, you follow the videos of StatQuest. There is Mathematics for ML from Imperial college, London on Coursera. DeepLearning. AI has released a Math for ML specialization on Coursera.

The manga math books — Manga Guide to Calculus, Algebra and Statistics are good help.

The Mathematics for ML e-book by Deisenroth and colleagues is available free.

There is a debate whether India should develop LLMs or should be satisfied with SLMs. However, China has launched a low-cost DeepSeek, and that has a lesson for India — India cannot be sidelined as far as the development of foundational models.

In past, India put up substitutes for WhatsApp and Twitter by introducing Hike and Koo, but they could not succeed against their global counterparts.

Indian and Chinese economies cannot be compared, since China runs an authontarian regime and blocks platforms such as Google, Facebook and Twitter. China can protect the domestic counterparts such as WeChat and Weibo. India on the other hand is an open economy. Any model India builds will face competition from both the US and the Chinese models.

Indian foundation model should be globally competitive. It has to remain sustainable.

As India is a big market for AI, it is expected to adopt AI. However, the paradox is that users favour the foreign models rather than indigenous models.

India’s AI mission has committed Rs.10000 crore. It proposes to build local foundational models. The real test is in execution. Indian models must be on par with the global models from OpenAI, Google or DeepSeek.

Indian IT companies are service-oriented and not product-oriented.

Indian companies are considering to switch to open-source models which allow them to build customized solutions avoiding prohibitive training costs. Open source declines API costs making AI more accessible to startups.

The government has announced AI safety institution. The India AI mission will satisfy the funding needs. India should develop foundational models and also be a part of AI ecosystem.

Future AI aspirations will depend on access to computing power, talent and market adoption.

France intends to invest $112.5 billion in AI over coming years since Europe cannot afford to miss another technological revolution. Of course, the US leads the AI revolution with its Big Tech and Nvidia chips, and China is the number 2 contender. Europe, the old continent, seems to be already overtaken by events.

All is not lost. Much depends on using and diffusing AI so as to get the benefits of enhanced productivity. Researchers say what matters is the implementation of technology, and not necessarily the pioneer advantage. Academic institutions can reverse the backslide by augmenting and employing the inventions. The US did this in the early 20th century by overtaking the leaders in machine tools such as Britain and Germany.

Europe already enjoys in incidentally an advantage of talented manpower and reputed institutions. European companies can act as agents of change

On the stock market, there is the digital divine where US investments lead the European investments in automation. In retail, Walmart’s capital expenditure is increasing, whereas that of Carrefour is declining. It is this kind of challenge that Europe will have to resolve.

Ever since 1995, US productivity growth has outpaced that of Euro area, and may continue to do so. That should not deter the European leaders to lose a chance to revive the European economy. AI laggards can win the long game of adoption and innovation, since it is a Marathon.

The first AI Summit – 2023 was held in the UK and led to a non-binding pledge by 28 nations to tackle AI risks. The AI summit, Paris will start on Monday, the 10th of February, 2025. It will be a gathering of heads of state, top government officials, tech stalwarts and researchers for two days. The summit is co-hosted by French President Macron and Indian PM Modi. The Summit aims to harness AI’s potential so that it benefits everyone, while containing the various risks AI poses.

More than two years after ChatGPT’s debut, generative AJ continues to make astounding advances at breakneck speed. It is a transformative technology which affects many aspects of life by generating high quality of text, images or video or carry out complex tasks.

The participants will hammer out pledges on guiding the development of rapidly advancing technology.

The Summit comes at a fresh inflection point as China’s budget friendly chatbot DeepSeek shakes up the industry.

President Trump wants to make America the ‘world capital of AI’ by leveraging its oil and gas reserves. He has revoked former President Biden’s executive order for AI guardrails.

European Commission President Ursula is attending along with officials and CEOs from 80 countries including German Vice Chancellor, Canadian PM, OpenAI’s CEO, Microsoft President, Google CEO. Elon Musk has been invited. Liang Wenfeng, founder of DeepSeek too has been invited.

Thus, the Summit will focus on geopolitics of AI.

AI has been evolving at an astonishing space and is quickly moving from being emerging technologies to a transformative business force. In 2025, the major focus in AI development will be agentic AI and custom AI.

The rise of muti-agent AI systems is one of the most exciting developments. Custom AI aligns with the unique DNA of the business. It provides competitive advantage by solving niche problems while aligning with the business priorities. Custom AI allows organizations to extract maximum value from their data, while aligning with business goals.

Generative AI is entering transformative phase of agentification, evolving from task specific tools to specialized, interconnected AI agents.

In 2025, we may witness the emergence of superagents coordinating interactions among multiple AI systems to enhance collaboration, efficiency and reliability.

India’s advertising industry reached in 2024 Rs.101,084 (6.3 per cent growth). In 2025, it is expected to reach Rs.107,664 crore and in 2026, it will cross Rs.115,000 crore (CAGR 6.87 per cent).

India’s digital advertising outpaced all other media to reach a market size of Rs.49,251 crore in 2024 (21.1 per cent growth). It currently accounts for 49 per cent of the ad market. Digital advertising is likely to reach Rs.59,200 crore in 2025 (20.2 per cent growth).

TV and print account for 28 per cent and 17 per cent growth of advertising pie respectively.

The top five categories in terms of ad spending are FMCG (31 per cent of the total), e-commerce (15 per cent), consumer durables (7 per cent), automobiles (6 per cent) and government (5 per cent).

Traditional media such as TV, radio and print will continue to grow, although at slower pace.

(Dentsu Digital Advertising Report)