QR code stands for Quick Response code. It makes the user instantly connected to a website, video or app store. It guides the customers to get the information a company wants to provide. It makes a customer travel from the stage of awareness to action or purchase. It is versatile, low-cost and accessible tool. It integrates offline and online journeys and experiences. Scan-thought can be tracked.

In the West, QR codes were widely used in retailing. Of late, in India too we find QR codes at retail outlets.

The QR code was invented in 1994 by Japanese firm Denso Wave for labelling automobile parts.

QR code payments are contactless and convenient. QR code vaccine certificates in pandemic were a great help. QR, in fact, has become mainstream technology.

The excessive use of QR codes can result in a blind spot. Let people seek out the QR codes to make them effective.

Towards the end of May, 2023, WhatsApp announced its message edit feature. It allows users to edit their sent messages within a 15 minute window. It gives greater control over their chat with an ability to correct the mistakes or make changes if necessary.

It is a welcome relief to 2 billion plus users. A user can rectify mistakes or typos, can modify the tone of the message or can add context to it. It brings flexibility and convenience in everyday chats.

It is being introduced globally. It is valid for both individual and group chats.

To edit, the user has to long press on the sent message and select ‘edit’ option from the menu that appears.

The edited messages will display an ‘edited’ label next to the messages. It will alert the recipients that there have been changes in the message. However, the edit history will not be shown. This preserves the privacy and integrity of a personal chat.

So far, the option available to the users was only the deletion of the message, and send a separate corrected message. The competitors such as Telegram and Signal already provided ‘edit’ function.

Reels or short videos are highly popular both on Facebook and Instagram. Facebook is doubling its efforts to monetise these videos, and stay ahead of the competitors such as YouTube Shorts, Moj, MX TakaTak. Several new features are being added, and some old features are being dropped.

Facebook is testing its programme Ads on Reels so as to compensate the creators solely by the performance of the reels and not by ad earnings. The same programme will be rolled out on Instagram too.

Reels Play bonus has been phased out in 2023. It rewarded the creators for video goals including performance metrics, post frequency and thematic content.

Facebook proposes a gift scheme. Here the users can purchase gift stars for their favourite creators. Creators will receive a payout of $0.01 for every star received.

Facebook ties up with the campaigns of many brands. It introduced. MadeonReels inviting brands to submit their creative briefs and participate. Selected brands will receive support on Reels at Facebook expense. It will enhance the visibility of Reels.

The Reels were started less than three years ago. Today 2 billion plus Reels are shared everyday (up from 1 billion just six months back).

The short form video advertising market is expanding. It keeps advertisers interested as these are shorter, crisper and snappier messages as compared to TV commercials. Besides, these are made with smaller budgets, and can be produced faster. TV commercials show budgetary extravagance.

There are challenges. One is the content creation with impact on a continuous basis. There are diminishing attention spans, say only the first three seconds in digital landscape. There has to be a fit between brand ethos and the content creator. There is competition from other platforms.

Facebook, despite all this, has retained its hold over this segment — 140 billion daily views across Instagram and Facebook.

Backpropagation is used in feed forward neural networks. In a feed forward neural network, connections between the nodes do not form a cycle. Rather they move in the forward direction. The word feed forward indicates that information travels in the forward direction. It first enters the nodes, moves to the hidden layers, and comes out through the output nodes. In fact, this is the first and simplest type of artificial neural network.

Businesses, of late, are being data driven. Application Programming Interface (APIs) are developed to connect the frontend to the backend of any application. Such APL can be susceptible to unauthorised access if those do not properly authenticate and authorise the user. Businesses open their APIs to other service providers.

Uber and Ola are aggregators for cabs, and use the Google Maps API. When online payments are made by UPI, the payment gateways such as RazorPay queries an API (asking for access). While signing for a website using Facebook or social media account, the social media’s API ensures authentication. Data transfer happening so seamlessly makes life easier.

API security is important since there is an app to app traffic on account of APIs, and IoT devices. App to app traffic is more than app to human traffic. APIs which are vulnerable and are hacked can reveal financial, medical and personal data to public. Since API is a mini-webapp, the security measures for a full-fledged webapp may not be taken. In a webapp, when user interface (UI) is modified, it is not visible. Testing is, therefore, completed instantly. API is not visible to the end-users, and hence it is released hurriedly without security testing.

Organisations want their platforms to add APIs swiftly to enhance functionality .API adoption is quick. There is a footprint of connecting apps. Such practices encourage API breaches. There should be continuous assessment of API security.

APIs are universal attack vector. They also provide more attack surface across all vectors.

Organisations are not aware how many APIs they have and what those APIs are doing and where they are located. Managing these is complex, especially when these multiple APIs keep changing and evolving.

AI is transformative. It is becoming ubiquitous and influential. Therefore, the issue of its regulation arises. Can it be regulated like banking and healthcare? However, AR is unique and evolving rapidly. AI systems learn and adapt. Rigid regulation can stifle it. It can hinder progress. There cannot be prescriptive regulation here.

AI has to experiment and innovate. It has got to be better trained. Restrictive regulations could cripple AI. The pace of innovation is the driving force for the success of AI. Boundaries get pushed. New apps are being explored. We do need cutting edge algorithms. AI should have guardrails, rather than strict regulations. Guardrails should ensure ethical and responsible development of AI.

Guardrails provide guidelines and principles to encourage AI developers to put a premium on fairness, transparency and accountability. Of course, there are ethical challenges. There is the issue of bias in AI’s algorithms. There is an issue of reskilling the workforce.

The bias emerges from complex and opaque algorithms which regulators do not comprehend.

Organisations must rationalise AI-driven decisions. They should disclose their data usage practices. The assessments should lead to identification and rectification of biases. There should be regular audits.

Guardrails should nurture a culture of responsible AI development. Ethical considerations should be at the forefront.

‘One-size fits all’ approach of regulations is not suitable for AI since it is applied to diverse sectors. It is better to have flexible guardrails.

AI is a global phenomenon. GDPA of Europe imposes strict data protection rules. It safeguards individuals. However, these become challenging for AI development. Here for robust development of AI, data sharing is necessary. China has a top-down approach. AI guardrails can allow adaptation to local and global contexts.

AI guardrails could be collaborative. There could be inputs from tech experts, policy makers, ethicists and the general public.

There cannot be red-tapism in AI as it will be a death blow to AI.

There should be a balance between regulation and innovation. There should be a flexible approach.

Gradient descent is an optimisation algorithm. It is used in training ML models and neural networks. Its aim is to minimize costs of the model by adjusting the weights of the neurons.

To begin with, random weights are assigned. The input data is fed forward to generate an output. This is compared to the desired output so as to calculate an error. This error is propagated backwards through the network using backpropagation. The weights of each neurons are modified based on their contribution to the error. The steps are repeated till the cost function reaches a minimum value.

There are different types of gradient descent — batch or stochastic or mini-gradient descent.

Stochastic gradient descent (SGD) updates weights of the neurons after each training example is processed. It is faster, but less accurate.

Batch gradient descent updates weights of neurons after processing all training examples. It is slower, but more accurate.

Mini-batch gradient updates weights of neurons after processing a small batch of training examples. It is a compromise between stochastic and batch gradient descent.

There are other descents such as Adam, Adagrad, RMSprop, Adadelta and Nadam.

In SGD weights are adjusted after each training example is processed.

Update rule:

w=w — learning rate * gradient

Where w=weight, learning rate is a hyperparameter that controls the step size of the update, and gradient is the gradient of loss function with respect to the weight. It tells us how much we need to change the weight to reduce the loss.

Backpropagation refers to backward propagation of errors. It is used for calculating derivatives in deep feedforward neural networks. It constitutes an important part of supervised learning algorithms for training feedforward neural networks, e.g. stochastic gradient descent(SGD). After all, backpropagation is an algorithm. It makes neural network learn from its mistakes. To do so, weights of the neurons are adjusted, the error is propagated backwards through the network, and weights of the neurons are adjusted depending on how these contributed to the error.

Working Overview

Input data is fed forward. It produces an output. This output is compared to the desired output to compute the error. This error is propagated backwards through the network to calculate how much each neuron contributed to the error. The weights of each neuron are adjusted based on how much they contributed to the error. These steps are repeated till an acceptable level of accuracy is attained.

In 1980’s, Hinton, a University of Toronto professor, began his work on neural networks. These networks were trained by data, rather than programming them conventionally. This was an attempt to give computers intelligence. Hinton received the Turing Award in 2018 for his work on neural networks. The other two scientists who shared this prize with Hinton were LeCun and Bengio. LeCun is with Facebook and Bengio with University of Montreal.

Hinton joined Google in 2013. In fact, Google acquired his company DNN Research which tried to commercialize deep learning ideas.

Researchers at Google devised a new neural network called Transformer which gave birth to models such as GPT4 and PaLM. Hinton co-invented backpropagation which is a fundamental algorithm for training neural networks. He also contributed in developing Boltzmann Machines which are probabilistic generative models. These machines learn to represent complex data distributions.

Though neural networks are inspired by biological networks in the brain, human brain is more advanced as it learns from less data. Still we have much to learn about the complex relationship between AI and human cognition.

Hinton feels AI systems understand similar world views, as they cannot appreciate different interpretations of the same physical reality. Hinton is hopeful about machines being appreciative of different perspectives.

It is a moot point whether coding will be relevant when AI is advancing so rapidly.

Hinton is all praise for Microsoft-backed OpenAI for launching ChatGPT, and is also appreciative of the cautious approach of Google.

Generative AI chatbots such as ChatGPT of OpenAI and Bard of Google have led to a debate about the capabilities of these, and their drawbacks. It should be noted that both Sunder Pichai, Google CEO and Sam Altaman, OpenAI CEO have expressed reservations about the generative AI’s unhindered development.

In fact, Pichai observed that Google Bard had unexpectedly learnt Bengali language by itself. This self-learning raises questions about the nature of AI and our incomplete understanding of it. Even Pichai referred to a black box in AI. Some aspects of AI remained esoteric, e.g. Bard’s unanticipated mastery of Bengali. All this was referred to in fiction about AI, but that fiction is reality in 2023.

However, Google’s researcher Margaret Mitchell countered her CEO’s claim, and said PALM which is forerunner of Bard had been trained to understand Bengali. However, Mitchell was dismissed in 2021. There are ethical issues here.

Another Google engineer Blake claimed that its chatbot has become sentient. Blake too was fired.

It is ambiguous. We still do not know the true capabilities of AI.

AI systems show emergent behaviour and exceed expectations. It is necessary to understand the implications of such advances.