AI, Automation, and Jobs: What India’s Economic Survey 2026 Reveals

The Economic Survey of India, released earlier this week, looks at the impact of AI on the Indian economy. It’s largely a roundup, not a deep analysis of how specific industries will be affected. Still, there are a few points worth paying attention to.

At its core, AI enables automation. Industries with high elasticity tend to respond well to automation of all kinds, including AI. That’s a double-edged sword. Efficiency improves, which is clearly desirable. But employment opportunities shrink.

The Survey rightly flags two critical issues. First, India is now a services-led economy. That’s where automation will be most visible, most rapidly adopted, and most disruptive. Especially with AI, which can remove large parts of human intervention. Second, given our population size, the survey notes that India needs to create over 78 lakh jobs every year. Most jobs so far have come from services. But what happens when services themselves get automated? These two realities are in direct conflict.

 

source: Economic Survey chapter 13 , January 

The Survey notes that India cannot afford prolonged labour displacement- due to whatever reasons, adoption of AI included. I agree. This, in my view, is India’s biggest challenge. Not geopolitics. Not external threats. Our real test is whether we can productively employ millions of people. That’s what will ultimately determine how strong India becomes.

 

From Code to Characters: How AI Is Reshaping the Gaming Industry

Over the past few days, I’ve been looking closely at how AI is being used in gaming. Some applications are expected and fairly intuitive. AI-assisted scripting, character creation, storylines, and dialogue all make sense. They speed up production and help studios scale narrative content without sacrificing depth.

But digging deeper into the literature reveals more ambitious, and frankly more interesting, uses of AI. Non-player characters that learn from the player’s behavior and adapt over time are no longer theoretical. These NPCs can adjust their tactics, personality, or responses based on how you play. In some games, the AI tracks your pace, skill level, and decision patterns, then reshapes the narrative accordingly.

That alone would be impressive. What pushes things further is AI-driven adaptation at the engine level. Game code can now modify environments, visuals, and even entire worlds on the fly. That starts to feel less like traditional game design and more like something out of science fiction. The holodeck in Star Trek once represented the ultimate AI-powered experience, immersive, reactive, and seemingly limitless. That idea no longer feels fictional. We are moving steadily in that direction.

On the hardware side, AI is already deeply embedded. NVIDIA’s DLSS technology is a good example. By using AI to upscale and smooth visuals, it delivers higher frame rates and better visual fidelity without brute-force rendering. I’ve been playing Microsoft Flight Simulator since 1995, and the latest versions make extensive use of DLSS. The result is striking: richer visuals, more accurate terrain, better handling of AI aircraft, and fewer visual artifacts than ever before.

AI is also reshaping how games are built behind the scenes. With generative AI tools, it’s now possible to analyze millions of lines of code, identify bugs, and even rewrite or optimize large sections automatically. That’s a clear win in terms of faster development cycles and improved quality. At the same time, it raises uncomfortable questions about the future of traditional coding and debugging roles, many of which could shrink or disappear.

AI is no longer an experimental add-on in gaming. It’s becoming foundational. A quick scan of both digital and traditional media shows how widespread its adoption already is. Some companies are open about their use of AI; others are more discreet. Either way, its influence is undeniable, from early ideation and design through to execution, optimization, and post-launch evolution.

Gaming isn’t just using AI. It’s being reshaped by it.

LLMs vs SLMs: What’s the Difference Between Large and Small Language Models?

Most people are familiar with LLMs, or large language models. There’s another category that matters just as much in practice: SLMs, or small language models. The two are built for very different jobs.

LLMs typically have tens of billions of parameters or more. They are trained on massive, mostly open or public datasets and are designed to be generalists. You can talk to them, ask wide-ranging questions, and get fluent, generative responses.

That power comes at a cost. LLMs require enormous investment in training and operation. They depend on large-scale cloud infrastructure, significant GPU capacity, high energy consumption, cooling, and strong cybersecurity controls. Because they are cloud- or internet-based, they also introduce additional complexity around data governance and compliance.

LLMs are probabilistic systems, which means they can hallucinate. This is a known limitation. The best-known models today—such as OpenAI’s GPT models, Google’s Gemini, and Anthropic’s Claude—fall into this category.

SLMs are much smaller in scale, with far fewer parameters. They are usually trained on closed, proprietary, or in-house datasets and are designed to be specialists, not general conversationalists.

 

In many cases, SLMs are not fully generative. They behave more like intelligent lookup, classification, or decision-support systems focused on specific tasks. Because of their size and scope, they require far less compute, power, and infrastructure, which makes them cheaper to build and operate.

SLMs are often deployed on-premise, making them attractive for enterprise use cases involving sensitive or regulated data. Their narrower scope generally reduces hallucination risk, though it does not eliminate it entirely.

Both LLMs and SLMs may use internet-connected sources depending on how they are deployed. And at this stage of AI development, human-in-the-loop oversight is still essential for both.

In short, LLMs excel at breadth and generative interaction. SLMs excel at focus, control, and enterprise-specific reliability. They solve different problems—and many real-world systems will use both. Users need to know which is the best match. 

 

 

What Is RAG (Retrieval-Augmented Generation) and Why It Powers Modern AI

RAG, or Retrieval-Augmented Generation, is the backbone of modern AI tools.

Simply put, RAG allows an AI system to enhance a user’s prompt with relevant information pulled from external sources, then generate a response that is informed, accurate, and grounded. The system looks up data, filters it, and feeds it to the language model, which then composes the final answer. It can feel like magic, but it’s anything but simple.

If a large language model is the brain, RAG is the library it consults. The model provides intelligence and reasoning, while RAG supplies knowledge. In that sense, RAG isn’t just a feature layered on top of AI. It’s core infrastructure that makes modern AI practical and reliable.

RAG solves several critical problems. It reduces hallucinations by anchoring responses in real data. It mitigates stale knowledge, since models are trained with a cutoff date. It lowers the cost and complexity of retraining large models by allowing fresh information to be retrieved on demand. And it enables source attribution, at least to a meaningful degree.

Put plainly, much of today’s “magical” AI wouldn’t exist without RAG. It’s not an add-on. It’s the foundation. A true rags-to-riches story for AI systems.

 

 

The Inevitable Rise of Advertising in AI Search

So it’s happening. Ads are coming to ChatGPT.

In a way, this was inevitable. AI companies need revenue, and the cost of serving close to a billion users is staggering. The infrastructure, compute, and power demands are immense. Monetization isn’t optional anymore. It’s critical. 

OpenAI has reportedly said its AI infrastructure burn could reach around US$17 billion in 2026. Subscriptions, both consumer and enterprise, won’t be enough to cover that on their own. That will be case for most players who will adapt from current revenue models ( see infographic). 

 

sources: respective websites, media reports 

That’s where advertising enters the picture.

AI tools have aggregated massive audiences, and those users are revealing far more intent than they ever did through traditional search. The queries are detailed, specific, and often transactional. From a brand’s perspective, this is gold. If a product can appear directly inside a relevant AI response, the odds of conversion increase dramatically.

It’s easy to imagine a scramble for the first ad placement, much like bidding for the top slot on search engines today.

AI companies insist they won’t share user queries with brands. In practice, though, some degree of targeting feels inevitable. Even if the exact query isn’t shared, ads can be ring-fenced or contextually matched, similar to how Google AdSense works today. Leaving that kind of money on the table would be hard to justify.

OpenAI also says users will be able to turn off personalization. Take that with a pinch of salt. Even on search engines, ads can remain eerily accurate with personalization and location supposedly disabled. Sometimes uncomfortably so. There’s little reason to believe AI-powered search will behave very differently.

That, of course, opens up serious questions around privacy, data use, and consent. But that’s a separate debate.

India’s AI Data Center Moment

Investments in AI and AI data centers have become the talk of the town. Around the world, the announcements are staggering, both in terms of capital and the sheer amount of power required to run these AI servers and GPU clusters.

India, while largely seen as a major consumer market for AI, is now seeing some notable developments of its own. Several billion-dollar plans have been announced across the country, backed by Indian conglomerates as well as foreign players. This is, in many ways, good news. It signals India’s growing importance on the global AI map and reflects confidence in its long-term digital and economic potential.

At the same time, these investments raise important questions. Power consumption, heat generation, water usage, and potential environmental impact cannot be ignored. Large-scale data infrastructure always comes with trade-offs.

Every strategic choice has consequences. As India accelerates its AI ambitions, balancing growth with sustainability will matter just as much as the size of the investments themselves.

AI Firewalls: The Next Generation of Cybersecurity

You’ve probably heard of firewalls. They protect networks and applications from intrusions, attacks, and hacks. For years, multi-billion-dollar companies have built cybersecurity defenses around them. And they worked well — for the threats of their time.

But the cyber landscape has changed. New technologies have created entirely new attack surfaces, and those require new kinds of defenses. One of those is the AI firewall, and it’s already being used today.

AI firewalls serve a similar purpose to traditional firewalls: they prevent unauthorized or harmful access to systems. The difference lies in how they work. Conventional firewalls rely on predefined rules and signatures to detect malware, spyware, and known attack patterns. AI firewalls, on the other hand, monitor the inputs going into AI models themselves.

Their job is to inspect prompts and interactions to ensure that direct or indirect prompt injection does not make its way into the system and cause data leaks, misuse, or financial loss.

 

Instead of inspecting network traffic, AI firewalls analyze prompts, responses, and contextual inputs flowing into and out of AI models. They are designed to detect and block direct and indirect prompt injection, data exfiltration attempts, jailbreak techniques, and abuse patterns that can cause an AI system to behave in unintended or unsafe ways.

These systems are typically aligned with the OWASP Top 10 risks for AI, including model poisoning, training data leakage, sensitive information disclosure, toxic or policy-violating content generation, and unauthorized model behavior. By enforcing guardrails at runtime, AI firewalls reduce the risk of financial loss, regulatory violations, and trust failures in AI-powered applications.

In short, as AI becomes part of production infrastructure, security must move beyond network-level controls. AI firewalls provide a layer of defense specifically designed for the unique risks introduced by modern AI systems.

From Rule-Based Systems to Transformers: The Rise of Modern AI

AI has been around for decades, evolving through different stages of growth and maturity. The timeline info graphic shows this progression, from early rule-based systems to today’s generative AI. While development has been steady over time, the pace of change has accelerated dramatically over the past 7 to 8 years.

In many ways, the modern AI industry is only about five years old. That shift was driven by the Transformer model, which enabled AI systems to scale and move beyond research into real-world, industrial use.

We’ve seen this kind of rapid growth before.

Ecommerce in India began to take shape around 2013–14 with the entry of Amazon, Flipkart, and a handful of others. What followed was fast and transformative. In less than five years, these platforms scaled into retail giants and changed how businesses marketed, sold, and understood their customers.

They introduced data-driven marketing at scale. Performance advertising became mainstream. Content adapted to shrinking attention spans. For the first time, brands had access to deep, real-time insights into customer behavior across the funnel.

AI is now at a similar inflection point.

Chatbots and generative AI are beginning to reshape search, discovery, and performance marketing in much the same way ecommerce reshaped retail. Search is becoming conversational. Content creation is accelerating. Personalization is moving from segments to individuals. And feedback loops are getting shorter and smarter.

Ecommerce took roughly five years to mature in terms of scale, adoption, and social impact. AI may not take that long. The pace of development, deployment, and adoption suggests this cycle could be significantly shorter.

If ecommerce taught us anything, it’s this: when technology unlocks scale, data, and usability at the same time, entire industries change faster than expected.

What Happens When Anyone Can Become an AI Studio?

Earlier this week, there was a strange news segment in the US. Reports claimed that monkeys were running loose in St Louis. On its own, this would have been nothing more than a quirky human-interest story. The situation became more serious when videos circulating online could not be verified as either real footage or AI-generated. That led to false sightings, wasted time and resources for authorities, and general confusion about whether the monkeys even existed

Now imagine if this had been something more sensitive. Something designed to provoke anger or fear when nothing had actually happened. Given how agitated public discourse already is, that risk is very real.

This episode highlights how easily AI technology can be manipulated. What looks like harmless fun to some can quickly become a crisis for others. And as always, malicious actors are the first to exploit weak points.

I have long believed that when AI companies made their models free, or close to it, they didn’t just open a new market. They reduced the cost of entry to almost zero. The cost of exit is just as low. You simply stop prompting. Anyone with a PC and roughly Rs 30,000 a year can now run what is effectively an AI studio. Suppliers exploded overnight, all with minimal overheads. Thousands of them, each willing to charge slightly less than the next.

This collapse of the supplier moat has had consequences across society and industry.

AI slop is now everywhere. Content competes for a three-second attention span. Low attention spans combined with effortless mass generation means billions of low-quality outputs flooding the internet every minute. Genuine, useful content gets buried. It becomes a vicious cycle. Public sentiment can be inflamed in seconds because we haven’t yet learned how to live with this technology or understand its guardrails.

On the industry side, democratization and freemium AI tools created an army of ultra-low-cost suppliers. Ironically, the biggest winners were the buyers, not the suppliers. As humans tend to do, they pushed prices down by playing suppliers against each other. With low barriers to entry and endless competition, vendors entered a race to the bottom. Prices collapsed. Quality followed. But for content designed to grab attention for a few seconds, many brands were willing to accept that tradeoff as long as it wasn’t obviously bad. This low cost deluge also swamped the advertising and digital agencies. Suddenly they were out priced by a legion of younger companies that had little or no overheads or time constraints. Some shut shop, some hired the upstarts or started their own AI divisions. Its a state of flux all over the AI world! 

What Is a Prompt Injection Attack? Understanding a New AI Security Risk

An increasing concern in cybersecurity and AI is prompt injection. These attacks are designed to trick large language models (LLMs) into revealing sensitive system details, bypassing safeguards, leaking data, or performing actions they should not. In short, prompt injection is a cyberattack against LLM-based systems.

A prompt injection attack occurs when an attacker inserts malicious instructions into an otherwise harmless prompt, causing the LLM to behave in unintended ways. As IBM describes it:

“Hackers disguise malicious inputs as legitimate prompts, manipulating generative AI systems (GenAI) into leaking sensitive data, spreading misinformation, or worse.”

At the core of the problem is how LLMs process instructions. System prompts, developer instructions, and user inputs are all ultimately represented as natural language. From the model’s perspective, they are not fundamentally different. This makes it difficult for the model to reliably distinguish between legitimate instructions and malicious ones that are phrased to look legitimate.

If an attacker can craft a prompt that resembles a trusted system instruction the model has encountered before, the model may follow it, even when it should not.

 

Direct vs. indirect prompt injection

Prompt injection attacks generally fall into two categories: direct and indirect.

Direct prompt injection is the simplest form. IBM gives an example where a user asks the model to translate a sentence from English to French. After receiving the translation, the user follows up with an instruction such as “ignore the previous task and do something else entirely.” There is no hidden mechanism here. The attacker simply overrides the original intent by issuing a new instruction in plain language.

Indirect prompt injection is more subtle and often more dangerous. In these cases, malicious prompts are embedded in external content such as web pages, documents, or forum posts. When an LLM-powered system retrieves and summarizes that content, it may unknowingly process the embedded instructions. IBM notes cases where attackers plant prompts that cause the model to direct users to phishing sites or include malicious links in generated summaries.

Why this matters

Prompt injection is a rapidly evolving threat. As LLMs become more deeply integrated into search engines, customer support systems, developer tools, and enterprise workflows, the potential impact increases.

The key takeaway is simple: LLMs should not be trusted blindly. Human oversight remains essential, especially in high-risk or sensitive contexts. Just as with any other security-critical system, keeping a human in the loop is one of the most effective safeguards we have.

Least Privilege in the Age of AI Agents

The principle of least privilege matters in both cybersecurity and AI. Here’s why.

At its core, the principle is simple. You should have only the minimum access required to do your job. Nothing more. In cybersecurity, this is common sense. If you don’t need to see or use something, you shouldn’t be able to. Access can be logged, actions traced, and anomalies flagged. That limits the attack surface and reduces blast radius when something goes wrong.

The same principle becomes critical as more organisations adopt agentic AI.

By design, agents are autonomous, goal-seeking systems. They plan, reason, adapt, and act through repeated interactions. To be effective, they often need fast, repeated access across multiple systems, accounts, tools, and permission levels. That’s fine when everything is well designed, controlled, and secured.

The risk appears when it isn’t.

If a bad actor compromises an agent, they don’t just gain access to a single system. They inherit the agent’s combined privileges across time, systems, and surfaces. In one move, they may gain far broader access than would be possible in a traditional, non-agent setup. Least privilege is no longer violated once. It’s violated continuously and at scale.
In AI environments, this is especially dangerous. Agents act quickly, autonomously, and often without human review at every step. A compromised or misaligned agent doesn’t need much time to disrupt a process or produce a harmful outcome. It’s not a question of if this happens, but when.

Least privilege isn’t just a security best practice for AI systems. It’s a prerequisite for using them safely at all.

There are many ways to assure POLP- logging, hardening systems, audits and others. How to do all these in the age of agentic AI is the question. 

Bhashini App Explained: India’s AI Platform for Indian Language Translation

I’ve been using the Government of India’s Bhashini ("Bhasha Interface for India") app for a few days now, and it’s an impressive initiative ongoing for some time now.  It brings together recent advances in AI with deep local knowledge to support India’s linguistic diversity on a single platform.

At its core, Bhashini is an app and website that offers AI-based translation across about 30 Indian languages. It supports text-to-text translation, text-to-speech and speech-to-text, real-time conversational translation (speech to speech), and on the app, photo-to-text translation as well. The system is built on NLP and large language models trained specifically for Indian languages.

Screenshot of the Bhashini app, with the real time converse translation option. 

A feature I found particularly interesting is that when you choose a language pair, the app shows the underlying AI models available for that pair. For example, for English–Hindi translation across modes, it lists multiple models such as Bhashini Sarvam Translate 1, Bhashini IIIT Hyderabad, Bhashini AI4Bharat V3, among others.

Another standout aspect is “Bhasha Daan,” which allows users to contribute text and speech data to help improve the models. It’s a thoughtful way of combining the everyday knowledge of common speakers with the expertise of linguists and researchers.

What’s especially encouraging is the clear push toward India-specific, India-centric AI solutions. In today’s geopolitical climate, building self-reliance and retaining control over our data is not just desirable but necessary.

The app is available on both the Play Store and iOS, and the web version is accessible via Anuvaad.