Blog

WhatsApp, the messaging platform has over 1.5 billion daily users. Since the time of its launch, and post its acquisition by Facebook, it has never been keen on advertising.

In June 2025, WhatsApp announced a set of new features. One such feature is rolling out ads in Status. It allows users to find new businesses while browsing through the tab.

The users can support their favourite channel by subscribing to receive exclusive updates for a monthly fee. They can easily start a WhatsApp chat with them about a product or service.

The new feature stays true to its commitment as users who do not use Status update will not see any ads.

The age of slow creativity is over. There should be a combination of emotionally resonant message and technological savvy. Traditional formats must be broken, and tech should be fused with culture.

Indian advertising is at creative crossroads. It shows it is good at craft, storytelling and is soulful. However, the future demands speed, system thinking and platform-native ideas. The advertising landscape is being reshaped at great speed.

AI is not a novelty. It has emerged as a new normal. The very grammar of advertising has changed, Advertising ideas should live across realities — digital, physical and virtual.

Are we Indians creatively complacent? Maybe, our work is emotionally rich, but is it technologically fluent?

Brands these days build eco-systems, not just ads. Tech savvy players rule the roost.

2025. Creativity is not just restricted to imagination. It demands integration.

India displays sparks but these are isolated. Indian ads still follow a set formula. India has rich story-telling tradition and talented creative minds. The need is to be bold and take the risk. India still lacks tech-forward ambition.

Future demands not louder but smarter campaigns. Just nostalgia cannot carry us into the future. Soul is okay, but speed too matters. Decide whether you have to lag or leap.

The annual Cannes festival has made sure its definition of creativity keeps pace with the world.

AI can create music, but a live performance confirms what AI cannot do. Soon AI will not be talked about. Just like electricity, it will fade into the background. What matters is what we make with AI.

Disney and Universal Studios from Hollywood have filed a copyright violation suit against Midjourney, an AI image generator in Los Angelas (LA) Federal Court. Hollywood has been drawn into the battle over generative AI. The law suit contends that Midjourney ‘helped itself to countless ‘ copyrighted works to train it’s software. They pirated the material from their libraries to generate and distribute ‘endless unauthorised copies ‘ of their famed characters (Dart Vader from Star Wars and Minions from Despicable Me). People can create images and videos that blatantly incorporate and copy their famous characters. Midjourney is the quintessential copy free-rider and a bottomless pit of plagiarism. The studios sent a ‘cease and desist’ notice and Midjourney has not responded.

The suit asks for damages (but does not specify the sum). The studios want court to stop offering it’s forthcoming video service without appropriate copyright measures.

LLMs such as GPTs are considered a breakthrough event as they represent a significant leap in how machines understand and generate human language. We can appreciate their transformative influence by considering following points.

1. Multi-tasking : LLMS can perform a wide variety of tasks: coding, translation, question-answers, tutoring and summarization without task-specific programming. Previous models required rules for each task or there were separate models for different tasks. LLMs broke this barrier.

2. Understanding and Generation of Language: These models produce text that is coherent, context-aware and that could not be distinguished from human writing. They can grasp nuances, tone and even reasoning. They are highly fluent.

3. Non-linear scaling: As models become larger, say with billions to trillions of parameters, their abilities improve non-linearly. It is the scaling effect.

4. Few-shot and Zero-shot Learning: Without explicit training, LLMs can complete certain tasks (few shot) or none at all (Zero shot). It becomes possible because of emergent generalization, rare in previous ML systems.

5. Fine tuning: A model need not be trained for all possible tasks. A foundation model can be fine-tuned by separately training for specific tasks, say healthcare.

6. Democratization and Productivity: AI systems make several skills available to non-technical users. They boost productivity of manpower.

7. Beyond Language: LLMs are basically designed for text, but also display reasoning, planning and some symbolic logic. It hints towards a journey to AGI.

8. Innovation: LLMs accelerate R&D in the fields of drug discovery, protein identification and software engineering.

LLMs are breakthroughs by a combination of competence and generalization capability. They emerged suddenly with massive capability jumps. They have widespread influence across several industries. They suggest the beginning of more advanced ML.

In short, the timeline is shown below.

Pre-LLM Era (Before 2017)

Word2vec (2013, Seq2Seq (2014) and Attention Mechanism (2015)

Transformer Era

Transformer model (2017)

Early LLMs and Unsupervised Pretraining

BERT (2018), GPT-2 (2019)

Scaling Breakthrough

GPT-3 (2020)

Mainstream LLMs

ChatGPT (2022)

GitHub Co-pilot (2022)

BLOOM (2022)

Multi-modal and Agentic Intelligence

GPT-4 (2023)

Claude (2023)

Auto-GPT(2023)

LLMs as Platforms

GPT-4 Turbo (2024)

Gemini 1.5 (2024)

Sora (2024-25)

Different shareholders’ groups such as the National Legal & Policy Center (NLPC), Shareholders Association for Research and Education (SHARE), Inspire Investing raise issues about Google’s potential AI risks. Mostly the concerns are about the privacy rights, responsible investing and censor of speech. The wider shareholders of Google vote down these proposals. NLPC wants Google to clarity whether it is stealing people’s data to train its AI systems. SHARE wants a human rights impact study of AI-driven advertising. Google counters by saying that it regularly publishes AI Responsibility reports. These reports spell out their policies, practices and processes while developing advanced AI systems.

It is a fact that the reports are company generated and have not been scrutinized by any independent regulators or researchers.

Silicon Valley is known for its whitewashing by releasing glossy reports. Facebook has released a report on hateful conduct. Uber Technologies have released a report on safety statistics. These are not audited by any third party. There are no laws on content moderation, algorithm design and model design. The whole thing is pretty opaque.

A key ingredient to AI models is their traning data. It is kept secret by AI companies including OpenAI. If independent researchers get access to such data, they can better scrutinize for security flaws, bias or copyright violations. It is kept under wraps by the corporates under the pretext of trade secrets.

Google’s investors can bring about a real change by pushing for an external oversight prior to deployment of AI systems. Facebook took the initiative in this direction by hiring E&Y to audit its transparency reports. However, it could not go much further.

Google’s activists’ shareholders lack the clout to influence the financial setting at AGM. If their concerns are pushed further till they reach the regulatory ears, political channels such as law makers, something on the lines of FDA can be set up for AI to compel companies to meet certain standards before releasing their technology to the public. Till then, the status quo prevails.

The first Large Language Model especially suited for natural language processing (NLP) and which is transformer-based trained on massive data is GPT-1 by OpenAI in 2018. GPT stands for Pre-trained Transformer. GPT-1 had 117 million parameters. It was pretrained on large text corpus followed by fine-tuning on specific tasks. The transformer architecture was introduced on the basis of Vaswani et al paper in 2017 ‘Attention is All You Need.’

The earlier NLP moder were Eliza (1966) statistical NLP models such as Hidden Markov models, n-gram models, Word2Vec (2013) representing word meanings. Early deep learning models were ELMO (2018) that used LSTM and BERT — bidirectional transformer (2018) by Google. However, they were trained differently from GPT.

GPT is the first true LLM used for NLP.

IBM Infrastructure, India has five businesses — mainframes, power services, storage, cloud and customer support. Mainframes! What these quint machines are doing here! In this client-server era and cloud era, we hardly take notice of mainframes.

When all your banking apps running on public cloud, to get the balances or do money transfers, there is 70 per cent chance that the app sends a message to mainframe to get the information or complete the transaction. You will be surprised that 70-80 per cent of the world’s transaction data still resides on mainframes. It is more or less true for credit card transactions.

Mainframes are still used by banks, airlines, hospitals, government and even retail for their most critical applications. Mainframes have great processing capacity, security, uptime and resilience.

IBM which pioneered mainframes have made them compatible with cloud, since enterprises have moved to hybrid-cloud strategy.

These days AI has been built into mainframes. To begin with, the mainframe chip used AI for fraud detection in real time. The latest mainframe version brings in Gen AI capabilities. It helps in managing the entire computing system. Some support services get automated. A lot of hardware and software work has been happening with mainframes. There is no more talk of moving from mainframes to cloud.

The latest version of mainframe stack is z17 developed in India. A big chunk of hardware — CPU that goes into mainframes — is built in India.

Google cloud service will be utilized by OpenAI for its compute needs. The deal has been finalized in May 2025. As we are aware, deployment of AI models requires massive computing capacity. So far OpenAI depended upon its financial backer Microsoft for its cloud services. OpenAI would like to diversify its compute resources. It also intends to avail of Stargate data center. It will strengthen OpenAI’s computing capacity for running its AI models and training them. Earlier in 2025, OpenAI partnered with SoftBank Oracle for Stargate infrastructure programme. OpenAI has signed deals with CareWide for more compute.

The move to tie up with Google will reduce OpenAI’s dependence on Microsoft whose Azure cloud served ChatGPT until January 2025. Microsoft and OpenAI are in negotiations to revise their investment.

Google also wants to make available its chips called TPUs for external use which earned it clients such as Apple, Anthropic and Safe Superintelligence.

Cloud business accounts for 12% of Google’s 2024 revenue.

There is a dog in the ___. The options could be kennel, courtyard, street. How the token will be predicted by AI?

First, we will embed a vector, then form a weight matrix and then do a matrix multiplication.

Words are first converted into vectors. Here we have a vocabulary of 5 words:

[dog, in, the, kennel, courtyard. street]

Let us assign each word a 3D vector.

Word Embedding (3D Vector)

dog 1.0, 0.5, 0.2

in 0.3, 0.8, 0.1

the 0.2, 0.4, 0.9

These are taken from pre-trained embedding matrix., learned from large text data.

In step 2, we shall combine input embeddings

The sentence is ‘dog in the’. We will average the embeddings for simplicity (Alternatively, we can concatenate).

dog [1.0, 0.5, 0.2]

in [0.3, 0.8, 0.1]

the [0.2, 0.4, 0.9]

Average [1+0.3+0.2]/3, (0.5+0.8+0.4)/3, [0.2+0.1+09]/3

= [0.5, 0.566,0.4] Input vector**x**

In step 3, we use weight matrix W to transform the input vector into logits (scores for each word in the vocabulary)

W= [0.2 0.6 -0.3]

[-0.1 0.9 0.4]

[0.7 0.2 0.5]

[0.3 0.1 0.2]

After multiplying the input vector

X [0.5, 0.566, 0.4] with W raised to T

Each score

1.kennel

0.2x 0.5+0.6+ 0.566+(-0.3) x 0.4=**0.62

2.courtyard

0.1x 0.5+0.9 x 0.566 +0.4x 0.4 = 0.05 + 0.51 + 0.16

= ** 0.62**

3 street

0.7 x 0.5 +(-0.2) x 0.566 +0.5×0.4

=0.35-0113 + 0.2

= **0.437 **

These logits are converted to probabilities in next step ( Softmax)

Softmax (xi) e raised to xi / Summation j e raised 2 j

Use logits [ 0.32, 0.62, 0.437 ]

Compute exponentials

e x 0.32 = 1.377

e x 0.62 =1.858

ex 0.437 = 1.548

Sum = 4.783

Probabilities

kennel 1.377/4.783 = 28.8%

courtyard 1.858/ 4.783 = 38.9 %

streel 1.548/4.783 = 32.3%

The highest probability is that of courtyard (38.9%). AI chooses courtyard.

There is a dog in the courtyard.

To recapitulate,

1 embed words into vectors

2 merge the input vectors

3 multiply to get scores for each output token

4 Softmax

5 pick the most likely word.

Parellelization in AI chips (say Nvidia chips) denotes their ability to perform many calculations simultaneously, rather than one after the other, as traditional CPUs do. It accelerates data processing in AI and ML tasks.

First of all, parellelization is a great help in training neural networks. It further helps in executing AI tasks. There are massive matrices, and repetitive mathematical expressions ( additions and multiplications). There are independent computations on large datasets.

These tasks are broken into smaller pieces and are executed at the same time.

In such chips, there are thousands of mini-processors or cores. Each core can handle a small task. This many tasks are processed in parallel.

A CPU processes one task at a time, while an AI chips (GPU) can process thousands of operations simultaneously.

In training neural network, there are matrix multiplications on data batches, e.g. 1000 images. Each image can be processed independently. The task is split across many cores and each core processes one image or part of that image. It results into faster training and better scalability.

Matrices in neural networks are made of floating point numbers. These numbers represent data and model parameters. Input data of images is represented as a matter of pixel values ( say 28*28 matrix of numbers ranging from 0 or black to 255 –white. On feeding the model, it is flattened into a 784xj vector.) Neural networks learn by adjusting weights, stored in weight matrices. Input of 784 image pixels are connected to 128 neurons in the next layer. The weight matrix is 128*784. Result Z is 128*1 vector. Each value is the output of a neuron in the hidden layer. It is followed by a non-linear activation function such as ReLU or sigmoid.

In short, matrix multiplication combines input matrices and weight matrices to produce activations and predictions.