Drones have the potential of being used for deliveries, especially deliveries of products of e-commerce and quick commerce sites. By 2027, may be one third of the deliveries of all q-commerce retailers in major cities could be done through drones.

There could be sky pods in residential complexes. The drone carries the parcels of up to 10 kg there, and on landing a delivery executive stationed there collects the package and distributes the parcels to customer’s home. This makes the last mile logistics faster, sustainable, efficient and cost effective. It reduces carbon emissions substantially.

The civil aviation ministry is expected to formulate suitable policy to facilitate the drone deliveries.

Operating costs for a drone delivery service are 40-70 per cent lower than a vehicle delivery. Companies can tie up with startups in this area.

In the beginning of 2023, OpenAI’s internal messaging system was hacked and the hacker stole the details of the design of the firm’s AI technology. In fact, the hacker could get access to the discussions of the employees while they talked about the latest technologies being deployed. However, the hacker could not get into the systems where the company houses and builds its AI.

The disclosure was made to the employees during a meeting at the San Fransico office of the company in April 2023. Even the Board was informed. The news was not shared publicly as no information about customers or partners had been stolen. It was not treated as a national security threat since the hacker was a private individual and had no connections to any foreign government.

There is a lurking fear of foreign adversaries such as China who could steal AI technology and who could ultimately harm US national security. However, it is a moot point whether OpenAI’s security is strong enough to protect it against the theft of key secrets if foreign actors were to infiltrate the company.

OpenAI aims to build ultimately AGI equipping the machine to match or even surpass the human brain. This commitment is to build safe AGI. The Chinese threat is not unreasonable. Micrsoft’s systems were previously hacked by the Chinese hackers to launch a wide-ranging attack on federal government’s networks.

Both under the federal and Californian law, OpenAI cannot prevent people from working at the company on the ground of a particular nationality. Such blanket banning of foreign talent could significantly hamper the progress of AI in the US. The technology requires the best and brightest minds to work on it.

Apart from OpenAI, there are other organizations working in AI tech, e.g. Meta which shares its design with the rest of the world as open-source software. According to them, the dangers are slim, and code sharing enables the researchers across the industry to identify and fix the problems.

OpenAI and Nvidia have realized that there is a need to focus on India specific plan to build AI systems so as to facilitate the work of Indian entrepreneurs, startups, industry and consumers.

OpenAI would like to make affordable access to ChatGPT — the access should be affordable to developers as well as consumers. It also proposes to support Indian languages. GPT-4o is less expensive than its previous version GPT-4 and is faster too. It is getting more accessible too.

Nvidia is talking with startups and developers to assess the compute solutions they are looking for. India cannot depend solely on Western-centric model and the models should adapt to Indian culture and sensitivities. The models should be multi-modal too.

India will require large-scale, multi-lingual datasets, and synthetic data. India has 23 official languages and 10500 unique dialects across 123 unique dialects. The largest model now available deals with 100 languages.

GPT-4o has improved tokenization which cuts the number of tokens and saves costs.

OpenAI and Nvidia plan to go glocal.

Mustafa Suleyman, AI chief, Microsoft, believes in free use of online content to train AI models, and rules out consent or compensation to do so. He expressed this opinion while speaking to Aspen Ideas Festival. He feels that since the 1990s, the social contract of online content on the open web permits its reproduction as it constitutes fair use. The online content is freeware.

His comments have come in the midst of the copyright suits against OpenAI.

Suleyman points out some constraints. There are websites or publishers or news agencies who explicitly declare that scraping or crawling them (for any reason except indexing) is not allowed. Here comes the grey area.

Of course, what Suleyman interprets as fair use is flawed according to the critics. This is especially so when the content is used for commercial AI training. It is a profit-driven activity. It falls outside the realm of fair use doctrine.

AI is here to revolutionize the organizational work and operations. It has to allay the fears of content creators.

Cookies, as we know, are the little text files that track users’ habits. These are vital to online advertising business. Google makes money through this source — in the first half of 2024, 76 per cent of its revenue of $165 billion came from advertising. Cookies facilitate the working of this money-printing machine.

Advertising industry cannot sustain cookies at the cost of user privacy. Cookies days are, therefore. numbered. There should be a suitable replacement of cookies which meets the requirements of all stakeholders, and especially the regulators.

In 2020, Google declared its intention to depreciate third-party cookie support within its browser on the lines of Apple and Mozilla. Third-party cookies suck up behavioral data that is sold to data brokers to be used for various purposes, some of which are not consumer friendly.

Google moved 1 per cent of its Chrome user base into a pilot programme. The cookies of this group were disabled. Instead, tools from Goole’s Privacy Sandbox targeted the ads. In effect, it replaced third-party cookie’s function by some workarounds which were built into the browser. A user’s visit to a site with ads prompted Chrome to declare that this user prefers cricket and butter. That facilitated the uploading of relevant ads. Nothing was revealed about the user. There is no anonymity with third-party cookies. Of course, this is a positive development. However, it sets the standard in the most widely used browser. It allows Google to exercise greater control over online advertising. The problem lies in the prompting system. It encourages nudge in any direction. It prompts to ‘dark pattern’ uses. It is coerced consent. Such a consent would be rejected by the regulators. It is thus a very complex problem. One can be cynical about Google’s plan to phase out cookies. The solution would not satisfy all interested parties.

The transformer arrived in 2017. It represented a new architecture of sequence transduction models. A sequence model transforms an input sequence to an output sequence. The input sequence may consist of words, characters, tokens, bytes, numbers, phonemes or could be multi-modal.

Prior to the advent of transformers, sequence models were based on RNNs, LSTM, gated recurrent units (GRUs) and CNNs. To account for context, they contained some form of an attention mechanism.

Transformer relies entirely on the attention mechanism, and does away with recurrence and convolutions.

Attention is used to focus on different parts of input sequence at each step of generating output.

Transformer allowed parallelization (without sequential processing). It resulted into faster training without losing long-term dependencies.

The important components of a transformer are tokenization, embedding layer, attention mechanism, encoder and decoder. Each token is represented by an embedding capturing some kind of meaning. It is represented as a vector. An appropriate embedding dimension is determined — it corresponds to the size of the vector representation of each token.

The dimensions of the embedding matrix, for vocabulary size V and embedding dimension D becomes VxD, making it a high-dimensional vector.

Positional encodings are added to these embeddings (since the transformer does not have a built-in-sense of the order of tokens).

Attention Mechanism

Self-attention mechanism is that where each token in a sequence computes attention scores with every other token in a sequence to understand relationships between all tokens irrespective of where they are placed.

Attention scores result into a new set of representations for each token. They are used in next layer of processing. During training, the weight matrices are updated through backpropagation. The model than accounts better for relationships between tokens.

Multi-head attention is just an extension of self-attention. Different attention scores are computed. The results are concatenated and transformed. The consequent representation enhances a model’s ability to capture complex relationships between tokens.

Input embeddings with positional encodings are fed into encoders. Input embeddings are 6 layers — each layer has 2 sub-layers; these are multi-head attention and feed forward networks.

The output of an encoder is a sequence of vectors which are contextualized representations of inputs (after accounting for attention scored). These are fed to the decoder.

Output embeddings with positional encodings are fed into decoder. It contains 6 layers. The output embeddings go through masked multi-head attention. It means embedding from subsequent positions in a sequence are ignored when computing attention scores. The reason being that we generate the current token (in position i) and should therefore ignore all output tokens at positions after i. In addition, output embeddings are offset to the right by one position.

The second multi-head attention layer in decoder takes in contextualized representation of inputs before being passed into the feed-forward network. It ensures output representation captures the full context of input tokens and prior outputs.

We want to figure out what the next token is using contextualized target representations.

The linear layer projects the sequence of vectors into a logits vector (of the same length as the model’s vocabulary).

The linear contains weight matrix which when multiplied with decoder outputs and added with a bias vector produces a logits vector of the size 1xL.

Each cell is the score of a unique token. SoftMax layer than normalizes this vector so that the entire vector sums to one. Each cell then represents the probabilities of each token. The highest probability token is chosen and that is the predicted token.

While training the model, the predicted token probabilities and actual token probabilities are compared. We calculate loss function for each token prediction and average this loss over the entire target sequence. This loss is backpropagated over all model’s parameters to calculate gradients and model’s parameters are updated.

The GPT architecture was introduced by OpenAI in 2018. GPT’s do not contain an encoder stack in their architecture. It has been designed to focus on generative capabilities. It is trained on a large corpus of text. There is unsupervised learning of relationships between all words and tokens.

As reported in the media, Infosys, India has developed 250 apps of generative AI for being used in enterprises.

Tech Mahindra is using generative AI to code faster. One such area that has adopted generative AI-based coding is pair programming. This is a software development technique in which two programmers work simultaneously at one workstation. The driver programmer writes the code, and the observer or navigator reviews each line of code. They switch the roles frequently. It is a part of agile software development. AI can be leveraged in pair programming. AI can provide code suggestions and autocompletion. It speeds up the process. AI tools can detect real-time errors. AI can review the code and assure quality. AI can generate documentation for the code. AI can provide feedback. AI become a collaborative tool.

Another area where AI is leveraged is operations or production. AI can be infused into the technology stack.

AI is being infused into applications and business processes. It can facilitate onboarding process of manpower.

IT companies are hiring AI engineers and prompt engineers. The existing staff too is being upskilled in AI and generative AI.

IT companies focus on the BFSI sector (banking, financial services and insurance). They have developed new use cases in financial services. There are models of fraud detection, know-your-business (KYB).

There should be enterprise-wide reinvention across three pillars — data and technology, processes and people and a responsible AI framework.

Accenture has developed a range of generative AI use cases.

While navigating the shift, there should be continuous learning and adaptation and reskilling of the workforce. Only 5 per cent of organizations are actively reskilling their workforce on generative AI.

Silicon Valley is known for its creative surge over the years. In the 1960s and 1970s, there was a flourishing hippie movement here. It was a counterculture. Hippies embraced music, art and alternative lifestyles. There was a spirit of rebellion and experimentation. There was an urge to induce a social change.

Years rolled by. Silicon Valley took a great leap into technological prowess. The creative energy of counterculture was channelized into entrepreneurial spirit. It has defined the valley. The precursor was a PC revolution in the 20th century. That led to internet revolution in the 1990s. The whole valley has become a leader in technology.

Big Tech dominates the valley now. There are companies such as Facebook, Google and Apple. They introduced innovation and fostered economic power. The creativity in fine arts has sneaked into AI, biotechnology and other cutting-edge fields. This shift has brought forth new challenges. There are issues of privacy, corporate monopolies and social upheaval.

All said and done, so far Valley’s creativity has adapted to the times. It has travelled from cultural to technological innovation, still maintaining its role as a pioneer in new ideas and pathbreaking developments.

Hexaware has transformed it into an AI-first company. There is varying level of AI adoption among clients. Hexaware’s transformation focuses on using AI internally. It has developed a secure version of ChatGPT called Tensai GPT made available to all employees. It ensures data confidentiality. The manpower thus can leverage AI for decision making and for support functions.

AI has been adopted at scale. The technical staff is AI-enabled.

AI has been integrated with the existing services of AI — AI manages data estates for clients. Hexaware helps their clients to manage their data estate.

Hexaware also focuses on developing new AI solutions to address real-world business problems, e.g. AI copilot for insurance industry aimed at optimizing the work of underwriters.

Hexaware facilitates successful implementation in use cases where AI augments human capabilities.

Weekly injections of Ozempic made by Danish company Novo Nordisk are used for type II diabetes. The active ingredient includes semaglutide, which is not available yet as a generic. The medicine is sold under the brand name Ozempic and Rybelsus as anti-diabetics. Rybelsus is available in India in tablet form. The medicine controls blood sugar, lowers cardiac risks and stroke risks.

At the same time, Ozempic suppresses appetite. It increases the time of food exiting stomach. It induces a feeling of fullness. It mimics GLP-1, a natural hormone in the gut. It increases insulin secretion. It decreases the release of glucagon. This works to control the blood sugar. By mimicking GLP-1, it affects hypothalamus and reduces hunger. It reduces the motility of the food in the stomach. The user experiences feeling of fulness for a longer period.

Ozempic used as a weight-loss preparation. It is called off-label use. FDA approves it for diabetes type 2. Another variant Wegovy is used for obesity. There is increased demand for the molecule in recent years. Elon Musk famously tweeted about his weight loss: Fasting + Ozempic/ Wegovy + No tasty food near me.

Ozempic is not available in India. It has not been approved too. Its oral form Rybelsus is not as effective. In India, the medicines are imported unofficially or are bought in gray market. Some medicines sneak in from neighbouring markets, e.g. Incepta Pharmaceuticals from Bangla Desh supplying Fitara. In India, Sun Pharma is developing its own weight-loss formulation, while Cipla and Dr. Reddy’s are working on generic versions.

Ozempic’s side effects are mild-constipation, diarrhea, nausea, pancreas inflammation, low sugar, gall bladder disease, diabetic retinopathy. If medicine is withdrawn, the blood sugar control is lost and there are weight issues. There is a psychological fall out of drug usage and withdrawal.

Based on clinical data, there are 8 million plus prescriptions for Ozempic in the USA in 2021. Rybelsus sales in India has shot up to reach Rs.364 crore (Pharmatrack). Semaglutide’s market share in anti-obesity market in India is 66 per cent.