Author: Shabbir Chunawalla

Ever since we are dealing with mathematics and computers, we have come across a term — an algorithm. Algorithm is basically a rule that automates how a piece of data is handled. It brings classical logic to computing — If-then-else type of issues. If ‘z’ happens, do ‘x’. Else do ‘y’. After all, a computer programme is an agglomeration of several such algorithms. All these algorithms are strung though a logical sequence. It facilitates a certain result. These are simple operations on datasets. Computers perform such algorithmic functions at high speed.

Of late, algorithm has acquired a new meaning. There are algorithms that facilitate targeted advertising. The content offered on social media is based on the algorithms that predict the viewers’ likes. Radiological images are interpreted on the basis of algorithms. Generative AI works on predictions based on such algorithms. Some startups are funded on the basis of how they can beat the algorithm of large engines such as YouTube or Instagram.

Thus, algorithm’s lexical meaning has changed. We believe such algorithms can be changed by the organizations who create them. Gemini’s image generating capabilities have been criticized. Even text messages generated by the LLMs have attracted criticism. Can the firm fix this? In the field of AI, though the algorithm may have been initiated by a firm, the algorithm has the capability to learn and change itself. Those who own it can fiddle with it but they cannot completely control it. The functioning of these algorithms becomes a black box.

There are attempts to audit and modify these black boxes. However, these are not enough to control their going off the track and produce unacceptable results. There are limitations of such audits.

There should be complete access to the AI systems — access to its inner workings. There should be information about its development and deployment. There should be tests to see whether black box outputs consistently pass the discrimination tests. Thus, even when the logic inside the black box is irrelevant if the black box provides acceptable results, it is okay. Algorithmic results must be subjected to rigorous testing and audit.

Microsoft CEO, Satya Nadella, candidly admitted that had it not been for a Microsoft tie-up with OpenAI, Google would have been a default winner in the AI space. He called Google a competent player which has both the talent and the compute. They have everything to be successful — data, silicon, products and distribution.

Of course, Microsoft’s tie-up with OpenAI has made the field more competitive. If Microsoft partners well and innovates well, it can bring some competition to Google.

In fact, according to Nadella, Microsoft has been dabbling with AI since long. They did some serious work around speech in 1995. They recruited folks from CMU, and they have been working at Ai in various forms since long.

OpenAI partnership happened as they had a different approach. Microsoft is on the lookout for partners with whom they can innovate, and OpenAI team qualifies here. However, it all was a shot in the dark. It is the technology — you are in and hope for the best. This is the conventional wisdom. Thus a tech company has to backed long before it becomes conventional wisdom.

Microsoft is building AI infrastructure. It should be the best. Azure should be best both for training or for inference. They plan to tie-up with Nvidia and AMD. And with others. Azure should serve the needs of OpenAI, Mistral and Phi. They are focusing on small language model too.

AI models are being trained by vast amounts of data, and therefore are likely to show algorithmic bias. The model may falter while describing a political person’s leanings — whether dictatorial, fascist or democratic. The answers cannot be person specific and could sync with the ideological and political predilections of users. The issue is over-dependence on AI model far factually correct answers. It is forgotten that objective fact is a mirage. Carr, a historian, says facts are akin to fish swimming in a vast ocean. What historian catches, partly by chance, depends on the part of ocean he chooses to fish in, and the tackle he chooses to use. These two factors are being determined by the kind of fish he wants to catch. By and large, the historian gets the kind of facts he wants.

Thus, as Carr points out, historical facts are never objective. Of course, public opinion is influenced by the selection and arrangements of facts. As the popular maxim goes, facts speak for themselves. No, facts speak only when the historian calls on them.

A written work on social subjects would be judged on after examining the background of the writer. It facilitates the understanding of biases. In AI models, all of us expect 100 percent factual answers. The answers depend on the kind of dataset used to train the model.

AI content should be taken with a pinch of salt.

AI Con USA will be held between June 2 and June 7, 2024, at Caesars Palace in Las Vegas, Nevada in person and online. It will bring some of the brightest minds in this field together. It will be deliberate on the future of AI and ML and how to leverage them to transform business processes. It will provide a dynamic environment for knowledge exchange, collaboration and inspiration.

The keynote speakers will be from leading organizations such as Microsoft and DoorDash. There will be training classes on MLOps, GitHub Copilot, ML and AI. There will be tutorials covering topics such as generative AI, data generation with AI models, prompt engineering, image classification using LLMs data analysis and ML with Jupyter Notebooks and so on. There will be concurrent sessions from an array of speakers. There will be an Expo with leading solution providers..

As we know, Walt Disney and Reliance media business has merged. It becomes a conglomerate of 108 plus channels and two big OTT apps — Jio Cinema and Hotstar. It owns two film studios. Such a combination reduces the bargaining power of the media buying agencies. It gives tremendous negotiating power to the newly formed entity.

In the area of sports, a virtual monopoly has been created. It has 75-80 per cent sports properties under its belt. It owns IPL’s TV as well as digital rights. It owns rights of ICC cricket tournaments, both TV and digital. It owns rights of domestic cricket events of BCCI. In addition, it owns rights of kabadding league and Wimbledon.

Sports advertising (2022-23) touched revenues of Rs. 7100 crores (TV as well as digital). It benefits marketers by giving increased reach, optimum content costs and efficient operations. At the same time, it creates a monopoly that takes away the pricing leverage.

In 2023, the digital and TV rights were unbundled. It was an opportunity for brands in fintech, retail, e-commerce and edtech sectors. These brands moved towards digital advertising and legacy advertisers continued with TV. Since both digital and TV rights are now under one roof it has converted into a sellers’ market.

In advertising too, the supply-demand dynamics is at work. There was cautious spending in 2023 IPL. There was less ad money in the market and startups were not spending liberally. With many buyers, there is no significant pricing power. The merged entity will not be able to draw advertisers from social media and Amazon. There will be some shift from TV and print.

Print will have reinforced its inherent credibility. There is an opportunity to integrate print and digital audiences.

Alex Robert, prompt engineer, Anthropic tweeted about his experience while testing Claude Opus 3, a new LLM that was launched in March 2024. Alex reported that the LLM demonstrated a type of ‘metareognition’ or self-awareness during its needle-in-the-haystack evaluation.

In AI, metarecognition refers to the capability of the model to monitor its own internal processes. It is akin to self-awareness. However, this is anthropomorphizing (There is no self here). ML experts are of the opinion that AI models do not possess a form of self-awareness similar to humans.

The test was to measure the model’s recall ability. Here a target sentence (needle) is inserted into a large block of text or documents (the haystack). The AI model is asked to find the needle. The information is to be pulled from the large processing memory — context window. Here it consisted of 2 lac tokens (word fragments).

While being tested, Opus apparently suspected that it is being subjected to an evaluation. It was asked to find a sentence about pizza toppings. Opus spotted the sentence and also recognized that it was out of place considering the other topics discussed in the documents.

In response, Opus stated. ‘The most delicious pizza topping combination is figs, prosciutto, and goat cheese, as determined by the International Pizza Connoisseurs Association.’ The other material is programming related, and the sentence is out of place. It is unrelated. It may have been inserted as a joke or to test whether I am attentive.

Albert called such metawareness surprising and feels that there is a need for deeper evaluations of LLMs to assess their true capabilities and limitations.

Opus found the needle but went a step further and said the needle is out of place in this haystack. It recognized that it has been inserted to test its attention abilities.

In March, 2024, the Government of India launched its AI mission. An allocation of Rs.10372 crore was approved by the Union Cabinet. The idea is to make available compute capacity (consisting of GPU chips) under public-private partnership. At the same time, there should be funding of startups and setting up of innovation centers. All this will create AI sovereign infrastructure.

Compute power will be made available to innovators, startups, students and educational institutes. Guidelines will be framed for the scheme. There would be proper selection of startups.

Initially, the government will invest in 10000 plus GPU chips. These will be used to train neural networks. There is a proposal to set up a three-tier infrastructure consisting of 24,500 GPUs at 17 centers.

Nvidia is the supplier of GPUs and commands 88 per cent of market share. Due to high demand, there is a waiting time of 12-18 months in getting GPUs from the company.

The focus will not be restricted to generative AI, but will be also on real-life use cases in healthcare, agriculture, governance, language translation etc.

There will be focus on AI skilling. The government has signed on MoU with IBM in AI skilling. Nvidia is also co-operating with the government by working with Indian companies such as Reliance, Tata and Infosys to help develop sovereign AI infrastructure. There will be more courses of AI in colleges and universities for skill development. There will be a focus on the development of indigenous foundational models and indigenous tools.

Datasets will be created of anonymised data. The models will use anonymised data for training. Misuse of public data will be avoided by allowing only those companies which are building trusted AI models to access publicly available data.

We live in a fast-paced world, and in an over-communicated society. It is difficult to extract the relevant information. Here extractive summarization comes to our rescue. It selects key sentences in a document and presents a snapshot of the relevant points. It facilitates the understanding of bulky documents without reading each and every word.

In this write-up, we shall examine basics of extractive summarization with LLMs. The model uses BERT or Bidirectional Encoder Representations from Transformers.

Extractive summarization is a part of natural language processing (NLP) and text analysis. There is selection of key sentences or phrases from the original text. It is presented as a concise summary. It involves the sifting of the text to know the crucial elements, ideas or arguments.

In abstractive summarization, new sentences are carved out. On the other hand, extractive summarization is faithful to the original text. There is no paraphrasing or alteration. As far as possible the original wording and structure is maintained. It is a useful technique where accuracy is a desired a goal. Even the intent of the author is not disturbed.

It is used to summarize articles, research papers and reports. It shows high fidelity to the original, since paraphrasing may introduce a bias.

The procedure to do extractive summarization may have the following components.

Parsing of the text where the text is split up into sentences and phrases — the basic units. The text is dissected to understand its structure and parts.

Feature extraction analyzes algorithmically features or characteristics indicating their significance in the overall text. Repetitions and frequency of usage of words and phrases are common features. These may be central to the theme.

Sentences are scored based on their content showing their perceived importance. The higher the score, the more the significance.

Ultimately, the highest scoring sentences are selected and aggregated.

1. AGI: We are striving towards Artificial General Intelligence (AGS). AI will be able to accomplish jobs when told to do them. AGI will not be required to be told how to do the jobs. It does not mean sentient AI systems, but to get AI systems who work more intelligently and autonomously. Here we will have to ensure that AGI is aligned to the goals of society — goals that are in the best interest of society.

2. Quantum AI: Ten years hence, quantum computing and AI combination will lead to an era of supercomputing and faster scientific discovery.

3. AI-Powered Superhumans: By 2034, all of us can get the potential to emerge as superhuman beings with enhanced cognitive abilities. There is better learning and better recall of information. It will make decision making a powerful tool. AI-powered prosthetics can make us strong. We can have visual acuity till we reach advanced old age. We can have enhanced emotional intelligence and empathy.

4. Robotics: When powered by AI, robotics will be of great assistance Robotics will be influencing homes, offices and factories. Robots will be able to do tasks that human beings cannot possibly do or are reluctant to do. They can prove to be good companions.

5. AI in Public Administration: Here infrastructure, public affairs and justice systems will get the benefit of AI.

With respect to LLMs, we talk about the parameters. An LLM consists of billions of parameters. Actually, these are the weights (numericals) assigned to the connections between the nodes of neural network architecture. These weights determine the strength and influence of each connection. It ultimately shapes the LLM’s understanding of language and its ability to generate text. It makes an LLM learn patterns and relationships within the language.

Hyperparameters are the settings that control the training process. These do not affect the LLMs internal network structure. These influence the way an LLM learns from the training data. To illustrate, hyperparameters are the learning rate, batch size and the number of training epochs.

Parameters are adjustable weights. They are adjusted during training based on training data. They affect the processing of information in the network.