Blog

Prompt engineers are employed to train the large language models (LLM) of generative AI to deliver more accurate and relevant responses to questions people are likely to ask.

They can be utilised to generate code suggestions and identify coding errors and bugs. They can used for improving code performance. Prompt engineers design and develop prompts for ChatGPT. They use methods such as prefix tuning or prompt tuning. It works for pretrained LLM models. Prompt engineers use Chain-of-Thought (CoT) prompting to improve the reasoning ability of LLM Models, especially when the problem is complex and multi-step. Here the prompt is broken into a series of intermediate steps.

Prompt engineers must have creativity and must show language precision. He need not be a hardcore computer engineer. He can be drawn from humantities stream. Prompt engineer must be able to imitate human thought. Of course, an understanding of programming and AI and ML and NLP often helps. Tech skills must be combined with the understanding of user needs. A prompt engineer must have better linguistic skills.

There are short-term courses available for learning prompt engineering. Practice makes a prompt engineer perfect.

A prompt engineer cannot afford to be too literal or too fuzzy. He should combine reality and imagination. Essentially, it is an exercise of precision thinking. You should be at generative AI terminal for hours and figure out what prompt generates what response.

Big Tech such as Facebook, Google and OpenAI are making their claims in the area of AI and generative AI. Apple so far kept silent. Its CEO now declares that Apple is doing intensive research in this area, including generative AI. It is reflected in the R&D expenditure the company is doing — this year so far it has invested around $23 billion in R&D. Apple may put in the market an LLM on the lines of ChatGPT, and may call it AppleGPT. All eagerly wait for AI-assisted Apple products.

The Drugs and Cosmetic Act has Schedule M, which has been revised recently, and which spells out Good Manufacturing Practices (GMP). The idea behind revision of the GMP is to bring it on par with World Health Organisation’s (WHO) standards.

India wants to implement these standards in a phased manner. Of course, those companies which export drugs to other countries such as the US have to be compulsorily WHO-GMP compliant. However, so far, domestic companies did not comply with these, and out of 10500 drug manufacturing units, only a fifth were WHO-GMP compliant.

The government has set now a mandatory deadline for its implementation. Those companies with a turnover of Rs.250 crore will have to comply within next six months and those with a turnover of less than RS.250 crore will have to do so within a year.

India is known as a pharmacy to the world. However, there are issues about the drug quality and safety standards. The US FDA raises Official Action Indications for Indian drugs. There are at times import alerts and the concerned drug makers are banned from supplying to the US. There are instanced of drug safety failure in domestic market too.

Mere making the compliance will address only part of the problem. Specific changes are necessary in manufacturing processes. There should be full-fledged quality control system including product quality review. There should be stability tests and validation of equipment. Eye preparations and injections must be sterile. Testing laboratories should be inspected.

There is fragmented structure of drug regulation. Under Central Drugs Control Organisation (CDSCO), there are 37 nodal agencies across India. There is regulatory chaos and regulatory silos. The drug regulation is federal. Ease of business is promoted at the cost quality. There is a demand for centralised regulation and licensing. The enforcement gaps must be fixed.

Cadbury has released a new campaign. It allows consumers to make personalised birthday songs for their loved ones using AI. It has been conceptualised by Ogilvy and Wavemaker and tech partners Gen.ai and Uberduck. The brand has a new microsite ADSMN Interactive where users can share little interesting details about their loved ones to generate a customised birthday song for them.

Cadbury also won Grand Prix at Cannes (for creative effectiveness) for the Cadbury celebrations ‘Shah Rukh Khan My Ad’ campaign.

The recent birthday song campaign generated 30000 original birthday songs created within the first four days of the campaign.

Let us understand the natural immune system of the bacteria. These are infected by a virus and become bacteriophase. Small pieces of virus are captured by the bacteria and these are inserted into their own DNA. A pattern is formed called CRISPR arrays. It enables bacteria to remember viruses or closely related viruses.

When there is another attack, bacteria produce RNA segments from CRISPR arrays which recognise and attach to specific regions of the DNA of viruses.

Bacteria then use Cas9 or similar enzyme to cut DNA apart, thus disabling the virus.

In short, bacteria’s DNA gallery stores the genetic code of any virus that attracts it. The bacteria checks this gallery and when it finds a match, it activates a CRISPR-associated enzyne called Cas9 to cut it and disable it.

Editing Tool

CRISPR has become a highly precise gene editing tool. It is used to remove, add or alter section of the DNA sequence. It is a genetic manipulation tool.

The immune system, we have discussed, has been used to edit DNA.

How CRISPR Works

A small piece of RNA with short guide sequence is created. It attaches to specific target sequence in the DNA of the cell. The guide RNA attaches to CaS9 enzyme. When introduced into the cells, guide RNA recognises the intended DNA sequence. CaS9 enzyme cuts the DNA at targeted location. It snips the DNA. The cells then repair the broken DNA. Scientists can insert replacement DNA into the cell to stitch the broken cells together. The existing segment has been replaced by a customised DNA sequence. The genetic material can take additions or deletions or alterations.

Thus this is a mirror work of bacteria’s immune system. Bacteria have a natural editing system to have an immune defence. It has been adapted to develop CRISPR.

Genome editing is a group of technologies that give scientists an ability to change an organism’s DNA.

CRISPR is faster, more accurate, more efficient approach to genome editing.

Uses

CRISPR is used to prevent and treat diseases. It is more useful in single cell diseases, and promises to be useful in complex diseases.

Genome editing is limited to stomatic cells, and is isolated to certain tissues. It could be used to edit early embryos or germline cells or eggs and sperms. However, this affects progeny. It is a moot point whether it could be used to produce designer babies, say more intelligent and beautiful. It is illegal in the US and elsewhere. CRISPR can be used to produce mice where 80 per cent, or 100 per cent are females. Mosquitos can be made to produce male mosquitos and thus the disease producing, female mosquitos could be avoided.

How Used

CRISPR injection is administered into the eye directly. It consists of a non-pathogenic virus called AAV carrying CaS9 protein and its guide RNA.

Virus are used in gene therapy and editing as they have the ability to get into the cells.

History of CRISPR-CaS9

Crispr-CaS brings find and replace function in text files to the complex task of altering DNA.

It was launched in August 2012. US scientist Jennifer with her European collaborator Emmanuelle published a paper on CRISPR-CaS9. In January 2013 MIT and Harvard researchers adapted CRISPR-CaS9 to edit genes in mice and human cells. In August 2013, CRISPR was used to engineer plant genomes. In March 2015, scientists in the US and China explore CRISPR on stem cells to engineer human organs from genetically altered pigs. In February 2016, the UK allows scientists to alter human embryos using CRISPR. In June 2016, the US allows CRISPR to alter T cells for treating cancer patients. In May 2017, CRISPR is used by the US and Chinese scientists for HIV treatment. It was also shown that cardiac ailment can be treated using CISPR. In March 2018, genie editing of two human embryos was done by a Chinese scientist to disable a genie that allows HIV to enter the cells. In March 2020, for childhood blindness, a patient was treated using CRISPR to alter photoreceptor genes in the US. In September 2020, a Covid test was created using CRISPR called Feluda. In October 2020, Jennifer and Emmanuelle hot a Nobel for their work on CRISPR. In October 2021, there are trials of CRISPR-based cancer treatment.

In August 2023, the Digital Data Protection Bill (DPDP), 2023 has been tabled in the Lok Sabha. The first draft of the Bill was composed way back in 2018. This is the fourth reiteration. The Bill provides a legislative framework to protect the personal data of the data principals ( owners of data), and spells out their rights and duties. At the same time, the Bill also spells out the rights and duties of data fiduciaries (who collect the personal data), of data processors (who process the data), and of consent managers ( who act as intermediaries between data principals and fiduciaries).

We all know that the privacy (of personal data) is a fundamental right and digital data includes personal data. The draft was not for public perusal.

There is no insistence on local data storage. Social media has been declared as data fiduciaries. The data is to be collected for specific purposes with informed consent. It provides for correction or updating or erasing of personal data when the specific purpose for which it was collected is served, or when the principal withdraws consent.

Breaches of personal data must be notified immediately. Or else, the fiduciary is liable to fine.

Significant data fiduciaries collect volumes of sensitive data. They will have appoint independent data protection auditors based in India who will conduct data audits.

The central government and its instruments have been given right to collect data for broad purposes. It is an issue for concern.

The Bill proposes a Data Protection Boards (DPB) as the regulatory body, the members of which will be appointed by the Centre. Ideally, such a board should be independent. There will be an appellate tribunal to take up the cases when a person is aggrieved by an order/decision of the Board.

The onus of data breach will lie with the companies.

Startups will be exempt from the onerous provisions of the Bill, but are still subject to penalty for data breaches. The exemption for startups will be till the time they are developing a new product. On commercialisation of the product, they will be subject to provisions applicable to established firms.

The main aims of the Bill are data minimisation, purpose limitation and storage limitation. Data minimisation means entities can only collect what minimal is absolutely required. Purpose limitation means that data can be used only for the purpose for which it is collected. Storage limitation means once the services are delivered, the data must be deleted.

The government also gets power to block any intermediary or other firms in case of frequent data breaches and violations of the provisions of the Bill. The blockage will be on the recommendations of the DPB.

The government gets the right to exempt any agencies from the provisions of the Bill. This is a lean principle-based draft, an outcome of wide consultation (20000 submissions, dozens of discussions, balancing of various interests).

Some of its provisions are less prescriptive than standards in the EU’s GDPR. Big Tech firms face checks on monetisation of data.

These days, as we know, large language models (LLMs) are very popular in the field of artificial intelligence, AI. However, as we know, these require intensive computational resources to operate. Therefore researchers at Computer Science and AI Lab at MIT worked on an adaptable and efficient solution, especially useful in robotics and autonomous cars. They have come out with Liquid Neural Networks (LNNs). The research team was led by Rus. They were inspired by biological neurons found in small worms (such as C.Elegans) which perform complex tasks with just 302 neurons. These LNNs differ from traditional models as they do not use expensive computational hardware. The neurons are stabilized during training.

Actually, these models make use of dynamically adjustable differential equations. These adapt to novel situations after training. Typical traditional networks do not have this ability.

In essence, we enhance the representation learning capacity of a neuron. To begin with, the neuron stability is increased in training. Then non-linearities are introduced over synaptic inputs. It enhances the expressivity of the model, both during training and inference.

LNNs make use of a wiring architecture different from traditional models. There are lateral and recurrent connections within the same layer. Therefore, the maths and wiring both contribute to learning on continuous-time basis and adjusting behaviour dynamically.

The adaptation is in response to the inputs seen.

These are compact models. A classic model has around 1 lac artificial neurons, and 50 lac parameters to do a task of keeping a car in its lane. Rus and her colleagues were able to train an LNN to do the same task with just 19 neurons.

This reduction in size has its advantages. It runs on small computers. (used in robots and edge devices).

As it uses fewer neurons, the results are easily interpretable. It is easy to decipher how it has reached the decision. Just 19 neurons involved makes it possible to draw a decision tree corresponding to the firing patterns and the flow in the system. Is it possible to do so if we have 1 lac neurons?

The models also make us understand causal relationships. Traditional models struggle to learn this, and move over to spurious patterns unrelated to the problem. LNNs have a firm grasp of causal relationships.

Their generalisations are better for unseen situations. They focus on the task and not the context.

Attention maps drawn for LNNs show their focus on the task, say the road while driving or target object while detecting object. These adapt to the task, even when the context changes.

These models can be trained in one environment, but work fine in a different environment without further training.

LNNs are designed for continuous data streams — including video streams, temperature measurements and so on. There should be a time series of data. They do not work well for static data.

In robotics and autonomous cars, data is continuously fed to ML models. So LNNs work well.

LNNs have been tested for a single robot system, and Rus and her MIT collaborators would like to test them for multi-robot system.

AI has been in existence for the last several years. It makes an impact our work style and living style. The newly introduced generative AI on which ChatGPT is based is no exception. Traditional AI is cognitive technology. The same is used for training the machines. Generative AI generates output was not in existence. The training too is different. If a man reads non-stop to understand the accumulated information, he will take 2000 years to do so. Such a body of knowledge has been assimilated by the GPT4 model based on generative AI.

A teacher teaches essay writing to students in two ways. The first way is to make students do an outline, collect information on the topic, be particular about writing style, be attentive to the opening, middle and the end of the essay, make use of references. When all this is mastered, the students will start writing the essays. The second way is to make available to the students different types of essays to read. The students form an opinion on essay writing by reading them. The new essay will be generated by the students thereafter. This second way is similar to the way generative AI works.

Generative AI has affected artists and designers. New art work can be created using this technology very quickly. It has the capability to automate the designing of layouts, logo and visual elements. It will affect the graphic design field.

Generative AI affects the content writing and literature. It could be used to create articles, stories and other written material. It could generate an entire book. It can affect literary pursuits and journalism. It can summarise big books, legal judgments, medical journals very quickly. It will affect legal firms, creative agencies and marketing agencies.

In the field of customer service, generative AI affects call centre workers in BPOs and KPOs. Generative AI works in Github Co-pilot, and can create websites, mobile apps and computer programmers.

Generative AI facilitates the working of civil constructions and architecture. Repetitive work will be most affected as these are likely to get automated.

Technology displaces old jobs and creates new ones. On the one hand, it adversely affects existing jobs, and on the other hand, there are new opportunities. There should be focus on good prompts to generative AI to get the desired output. AI models can generate unsuitable outputs. AI models do not have an understanding of intangible truth in the models, the sense of distinguishing the right and wrong and common sense. At times, the outputs are fallacious, biased and false. Human oversight is, therefore, necessary.

According to McKinsey Globe Institute, AI and automation can displace 800 million jobs by 2030, At the same time, there will be new employment opportunities. Almost 275 million people will have to cope with the change in their work. The manpower will have to acquire new skills. Manpower can become outdated in 5 years. You cannot continue lifelong depending on existing skills.

Of late, Barbie movie produced by Mattel films earned in millions by brand collaborations — almost 100 plus. A wide range of products were Barbie-branded. The company itself, Mattel, released an entire range of Barbie merchandise — sweat shirts, bucket hats and hoodies.

In India, we do not see licensed merchandise ecosystem for movie characters except a Chhota Bheem here or a Bal Ganesh there.

Licensing depends heavily on brand building. There cannot be any licensing without building brands.

We in India have plenty of characters which can be associated with the brands. To begin with, we can start with the movie market — there are iconic characters like Bahubali and Pushpa which are amenable to merchandising opportunities.

There is an issue here — counterfeiting. Consumers opt for counterfeit merchandise which is available at a fraction of the price of originals. There is lack of awareness, especially in people from smaller towns and rural areas. In addition, licensed merchandise is costly, because of the high cost of licensing. It comes in the way of adoption of this merchandise. There should be consumer education, and strong anti-counterfeit measures.

In future, licensed merchandise will be facilitated by Web 3.0 and metaverse. New channels will be available for Indian characters.

An LLM model uses a neural network to learn the probability of each word that could follow a given sequence of words. The neural network has been trained on a massive dataset of text. It learns to identify patterns in the way the words are used together, e.g. the article ‘an’ comes before a noun with a vowel sound, the article ‘the’ is followed by a noun, and the word ‘to’ is followed by a verb. An LLM has to predict the next word. It examines the current sequence of words, and outputs probability distribution over all possible words that could follow. The word with the highest probability is the most likely next word.

To illustrate, let us take a sequence of words:

‘People wished her a happy__. The output of probability distribution is as follows.

birthday 0.8

Deepavali 0.9

Holi 0.7

X’mas 0.4

The word ‘Deepavali’ has the highest probability. An so it is the most likely next word.

The process continues till the response is complete.

The model actually uses vectors of the words to learn the associations and patterns. The accuracy of the prediction depends on the size and quality of the dataset.