Latest news with #Mayya

Mint

How reasoning AI is all set to change our lives

Varun Mayya, a Bengaluru-based entrepreneur building things at the intersection of generative AI and content creation, believes that the real world is messy and full of problems that require multi-step solutions and creative thinking. He wants AI that understands and can work with this. 'Reasoning AI models can infer cause and effect. If it's raining outside, for example, then they can infer that the ground is wet because of the rain, whereas vanilla models can get confused and assume that the events simply appear together and are not related," explains Mayya. The difference becomes apparent when the questions themselves are less like trivia and more like conundrums. Imagine, for instance, asking: 'Evaluate the feasibility of launching a vegan cloud kitchen in Gurgaon, with a break-even point within 9 months, based on current delivery and food-tech trends, along with regulatory risks." A chatbot that simply lists vegan recipes or gives a dated market estimate would be worse than useless. What's needed is a well-reasoned response that weighs demand and supply data, costs, consumer sentiment, policy shifts, and even the likely impact of next product pivot of food delivery apps. And now we have reasoning AI models that can do this, an evolution piggybacking on the rapid adoption of generative AI tools like OpenAI's ChatGPT, Microsoft's Copilot, and Google's Gemini along with a slew of AI experiences from innovative AI startups. Over the past couple of years, millions of individuals and businesses started interacting with a seemingly intelligent system using natural language, sparking a wave of experimentation and speculation about a productivity revolution. These early models were impressive content generators, capable of drafting emails, writing code snippets, and summarizing articles with astonishing speed. However, as the initial novelty subsided, businesses encountered the significant growing pains of this nascent technology. The very capabilities that made these early chatbots seem magical also revealed their profound limitations, creating a ceiling for their application in high-stakes enterprise environments. These limitations were not minor bugs but fundamental flaws that posed serious risks. These included hallucinations (fabricating information and presenting it as fact) and factual inaccuracy, outdated knowledge since the models were only as current as their last training dataset, and bias and discrimination as the AI systems inevitably inherit and often amplify the societal biases present in the data they are trained on. These limitations confined early generative AI to a sandbox of low-stakes creative work and preliminary drafting. They were powerful assistants but unreliable decision-makers. A strategic shift In response to these shortcomings, the AI ecosystem is now undergoing a fundamental and strategic pivot. The focus is no longer on simply generating more fluent or creative text. Instead, the race is on to build systems that can demonstrate genuine reasoning, engage in multi-step planning, and achieve verifiable accuracy. What users now want is not just a rapid-fire quip, but a reasoned argument; not just a list of surface facts, but a tapestry woven from context, inference, verification, and perspective. The new gold standard is smarter, not just faster, AI. This industry-wide pivot reflects a deeper ambition, one articulated by AI pioneer Yann LeCun: 'AI is not just about replicating human intelligence; it's about creating intelligent systems that can surpass human limitations". The goal is no longer just to mimic human conversation but to build cognitive engines that can solve problems at a scale and complexity beyond human capacity. An AI system must be able to analyse, interpret, and perform logical operations on that information. This is where the next evolution, often called Reasoning-Augmented Generation (RAG+) or advanced RAG, comes into play. This framework builds a layer of logical reasoning on top of the retrieved data. The 'black box" problem, where even developers could not explain why an AI produced a certain output, has been one of the single greatest barriers to enterprise adoption of AI tools, especially in regulated industries. Techniques that make the AI's thought process visible and auditable are the solution. When an AI can explain its step-by-step logic, a human expert, the 'human-in-the-loop," – be it a doctor, a lawyer, or a financial analyst – can validate that logic, challenge its assumptions, and ultimately trust its conclusion. The Rise of the AI Agent The abstract concepts of reasoning and planning find their most powerful and tangible expression in the form of AI agents. An AI agent is an autonomous system that receives a high-level goal from a user, independently creates a multi-step plan to achieve that goal, and then executes that plan by interacting with a variety of digital tools. Industry analysts predict that by 2028, 15% of all daily work decisions will be handled autonomously by AI agents, up from virtually zero today. Virtually zero, because we're not there yet. 'I recently bought a shoe using an agent and it bought the wrong one for me. It struggled through the manufacturer's website and thrice clicked on the wrong option before correcting itself. A human wouldn't have made the mistakes the agent did," says Mayya, underlining that we still have some miles to go towards critical evolution from AI as a tool to AI as a collaborator. That said, he thinks while AI agents are in their infancy, they are reasonably good for many use cases already, like solving most search-and-retrieving and information scraping tasks. The trend towards agentic AI is accelerating, with industry body NASSCOM projecting that the agentic AI market will surge from $5.1 billion in 2025 to over $47 billion by 2030. Of course, these advances bring new dilemmas. If a machine can reason, whose perspective, and whose logic, does it represent? With greater complexity comes opacity, and many models struggle to explain their thinking in human-understandable terms. Moreover, as reliance on AI for deep reasoning grows, there's the risk of outsourcing thinking entirely. Are users still equipped to spot errors, biases, or gaps in reasoning when the answer arrives neatly packaged, citations and all? The next frontier of AI lies not in flashier chat interfaces, but in robust reasoning engines that tackle the most challenging problems across domains. If there is a lesson in this shift from creative to cognitive, it is this: to extract the most value from AI, users must ask better questions—and demand better answers. And while the promise of smarter AI tempts us with visions of algorithmic sages, the truest breakthroughs may be quieter—better tools for collaboration, platforms that blend machine reasoning with human intuition, and the rise of augmented intelligence, where the best outcomes emerge from partnership, not replacement. OLD GEN AI VS NEW AGENTS Early chatbots • Content generation, summarization • Static, pre-trained knowledge, often outdated • Single-step, direct commands • Hallucinations, lack of verifiability Reasoning AI • Complex problem-solving, planning, automation • Live, real-time access to external & proprietary data • Multi-step, autonomous tasks requiring planning • Computational cost, ethical oversight complexity