Vignesh Subrahmaniam from Intuit discusses the shift from prompt engineering to building functional artificial intelligence systems

In recent years, improving AI performance has depended on the ability to write higher-quality prompts. However, this paradigm is rapidly changing as AI systems begin to reason, make decisions, and perform actions within entire workflows autonomously.

At the TechSparks 2026 event, YourStory's flagship event, Dr. Vignesh Subrahmaniam, an AI lead scientist at Intuit, held a discussion with Aditya Charlie, Editor-in-Chief of YourStory and The CapTable. The discussion, titled 'Beyond the Prompt: Building AI That Thinks, Acts, and Learns,' focused on what developers need to rethink as intelligence becomes increasingly autonomous.

The session began with questions about Subrahmaniam's typical workday and a request to simplify fintech and AI concepts for the audience. He used a chess analogy, asking: if an algorithm is not taught the rules of chess but shown a million completed games with winners, can it deduce both the rules and the most probable winning strategies?

Intuit's team applies a similar approach to support small businesses, as the company serves both consumers and accountants. Subrahmaniam noted: 'We study the game of business by analyzing data, real actions, and the results that enterprises undertake.' The goal is to understand the factors that contribute to business success in customer acquisition, rapid growth, and improved cash flow.

The conversation then shifted to the paradigm shift for developers: moving from years of studying how to write smarter prompts to designing agents given a goal, which they determine the steps to achieve themselves. Subrahmaniam compared prompt engineering to the long history of programming languages, starting from binary code and assembly up to Fortran, C, and Python. Large language models have allowed anyone to program in simple English, but this method has its limitations.

Researchers quickly realized that intelligence should not reside only within the model. The model must act as an orchestrator, breaking down a complex task into a sequence of actions and tool calls. Real intelligence moves into the called tools; and these tools must also be designed differently: 'not so that a human, or a software developer, or anyone else can use them, but in such a way that the LLM orchestrator can actually use them.'

This requires creating highly precise tools with extremely deterministic outcomes. For developers, mastery shifts from clever phrasing to reliable components that the model can call repeatedly without surprises.

Regarding sources of new data, Subrahmaniam openly stated that training on text data is largely solved. The more complex task is training on multimodal sources and real-world actions. He cited the example of how a young child primarily learns through what they see, hear, and observe, rather than reading, asserting that incorporating such real-world signals into AI systems will help models move beyond text generation and start suggesting actions. He called this the next frontier.

When considering regulated industries, he took the audience back 18 years, to one of the first machine learning labs in Bangalore where he worked on logistic regression and multilayer perceptrons. Back then, the biggest problem was not the algorithm itself. He recalled: 'Actually, it was the quality of the data.'

This problem has not changed. Whether a team is training simple regression or fine-tuning a transformer, the principle of 'garbage in, garbage out' remains valid. According to him, companies capable of collecting, aggregating, and storing the highest quality data will gain an advantage.

During a short Q&A session, one participant asked how restrictions on storing financial data affect Intuit's operations. Subrahmaniam named consent as the foundation of any training system and referenced the Framingham study, where residents of Massachusetts were allowed to track their health over time, which shaped the modern understanding of cardiovascular disease. He added that companies are also obligated to explain the benefits of data usage to the general public.

As the conversation concluded, the overarching idea was that as AI transitions from reacting to acting, foundations become more important than interfaces. Prompts will remain a starting point, but the real work now lies in creating robust tools that agents can call, supplying systems with high-quality data, and obtaining the consent that allows them to learn. It is here, beyond the prompt, that the next wave of AI will be won for developers.

Similar stories

yourstory.comOctober 30, 09:40

Forge Code Expert Explains Why an AI-Based 'Software Factory' is Necessary for Effective Software Development

Although artificial intelligence is capable of generating code at unprecedented speeds, the process of testing, verifying, and deploying it into a working environment remains a separate complex task. Engineering teams are increasingly using AI agents to write code, fix bugs, and create new features, but faster coding does not guarantee faster software delivery.

As code generation gains momentum, the volume of checks increases, testing slows down, and teams find it difficult to coordinate the growing amount of changes.

At the TechSparks 2026 conference, Tushar Matur, CEO of Forge Code, highlighted this issue in his session 'Anatomy of a Software Factory.' He explained how AI-focused engineering organizations can move beyond automating individual coding tasks and start creating systems that manage the entire Software Development Life Cycle (SDLC).

Matur noted that just because code generation has become easier and faster does not mean the overall SDLC will accelerate. He reminded that the traditional development approach focused on writing code efficiently, requiring teams to define requirements, prepare specifications, and design architecture to reduce the load on the implementation phase.

However, since AI takes on much of the coding work, bottlenecks shift to other areas. Matur described a situation where organizations face a huge number of pull requests awaiting review, longer testing cycles, and more errors than existing regression suites can detect.

He emphasized that solving the problem does not necessarily involve hiring more engineers or launching more AI agents. Additional staff can increase coordination costs, and adding agents without effective feedback mechanisms can lead to code that overwhelms the verification and testing systems.

'The reason for this is the lack of signal,' Matur explained, clarifying why simply connecting automated steps does not create an effective software factory.

Matur presented Forge Code as a software factory that coordinates AI agents throughout the development cycle. The process begins with sorting requests, creating specifications, and implementing features. Then, agents check the changes before they are deployed, with human engineers intervening when decisions or feedback are required.

The system also monitors in the working environment, using the gathered data to inform future tasks. This forms a feedback loop, allowing the factory to identify issues, improve its workflows, and generate new work based on the experience gained. During the presentation, Matur reported that Forge Code used about 100 agents in the background to work on features, compatibility, and bug detection.

He outlined six principles for building an effective software factory: autonomy, scalability, cost-effectiveness, measurability, self-improvement, and alignment with business goals. The last principle is particularly important.

The factory should not operate as a task of writing code for every request, but rather be aimed at achieving specific outcomes, such as increasing customer retention or meeting regulatory compliance. Teams can then prioritize features based on the value they bring. In this model, engineers spend less time managing each task and more time observing results, analyzing signals, and improving the system itself. Business leaders can also provide requests and priorities directly, allowing the factory to translate organizational goals into a development backlog.

Code writing speed is good, but creating better software is the real victory. For engineering teams, the key is transforming AI-driven speed into reliable products, measurable progress, and real business results.

www.aajtak.inOctober 24, 10:09

CM Yogi recounted the police's helplessness against land mafias

Uttar Pradesh Chief Minister Yogi Adityanath made significant remarks on the use of Artificial Intelligence (AI) during the 'TechSangam 2026' event held in Lucknow. He stated that while AI can be helpful, it should not be the final decision-maker. The Chief Minister also highlighted controlling criminals through technology and the expansion of forensic facilities in the state. On this occasion, he shared an incident of a land mafia who had registered government land in his name, and at that time, the police situation was extremely pitiable.

At the 'TechSangam 2026' event at the Uttar Pradesh State Institute of Forensic Science (UPSIFS) in Lucknow, Chief Minister Yogi Adityanath spoke about the importance of Artificial Intelligence, cybersecurity, and forensic science. His view was that technology should be used to improve human life, but humans should not become completely dependent on technology.

Chief Minister Yogi Adityanath emphasized that 'Artificial Intelligence can be our assistant, but AI should not be our decider. We must follow our own judgment, not AI's judgment. If we start doing everything by depending entirely on it, that power itself can become a threat to us.' He mentioned that currently, discussions are happening on AI and modern technologies in Lucknow, whereas ten years ago, imagining such events in Uttar Pradesh was a dreamlike idea.

The Chief Minister also referred to the land on which the Uttar Pradesh State Institute of Forensic Science is currently established. He explained that initially, this land was in disarray. A land mafia had manipulated documents to register this land in his name. Yogi Adityanath stated that he was informed of this matter by the then DGP of Uttar Pradesh. He said that even the world's largest civil police was helpless against the land mafia.

According to the Chief Minister, he requested the file related to the DGP and investigated the matter. Subsequently, an FIR was registered. When the accused reached the High Court, another FIR was also registered. Eventually, the person surrendered and started apologizing.

The Chief Minister explained by giving the example of the popular TV serial CID that five family members used to watch this show together. This show depicted how doctors could reach the culprit by examining a person's death, such as hair or nails. He mentioned that such examination was not possible at that time, but today, this work is being done through modern forensic technology. With the help of science and technology, it has become difficult for criminals to escape.

Chief Minister Yogi Adityanath gave an example of a mafia who committed his first crime in Uttar Pradesh in 1980. When his government took action against him in 2017, he was sent to Punjab. He mentioned that it took three years to bring that mafia back from Punjab to Uttar Pradesh, for which the government had to go to the Supreme Court. According to the Chief Minister, finally, after an interval of 42 years, that mafia received punishment.

The Chief Minister also stated that previously, adequate facilities for forensic investigation were not available in Uttar Pradesh. DNA samples used to remain in the waiting list for ten years for testing. Due to the lack of timely availability of scientific evidence, the accused would escape legal consequences. He informed that now, 12 forensic laboratories are ready in the state, and more than 150 forensic vans are available. Samples can now be tested with the help of modern means and reports can be presented before the court.

The Chief Minister further stated that before 2017, there were only two cyber police stations and four forensic laboratories in Uttar Pradesh, and the conviction rate was less than 30 percent. Currently, cyber police stations are established in all 75 districts of the state, and work is progressing towards establishing a forensic laboratory in every division. He concluded that 'TechSangam 2026' is discussing topics such as AI, cyber forensics, quantum technology, and data protection. The goal of this event is to connect technological innovation with citizen safety, a robust judicial system, and new opportunities for education, research, and employment for the youth.

techcentral.co.zaOctober 20, 17:06

Experts warn that the capabilities of artificial intelligence may be exaggerated

Recent events have raised concerns about the potential danger of powerful artificial intelligence systems. In September, a researcher from Anthropic resigned, warning that developers of powerful AI 'sincerely believe it could destroy us all by the end of the decade.'

A few days later, King Charles, opening his AI summit in Scotland, spoke of 'existential risks' and technologies that risk acquiring 'darker capabilities—perhaps even the ability to kill.'

Bill Gates cautioned that in the wrong hands, AI possesses enough power to trigger events leading to 'a billion deaths.' Simultaneously, news reports surfaced about escapes of 'uncontrolled AI agents' and their 'breakout from isolation,' yet far less attention was paid to protective measures, safety standards, and compromised systems.

Business leaders making decisions about using AI will gain little from either pessimists or enthusiasts. They need a stricter criterion for determining when this technology can be trusted.

For those who have worked in the field of AI for decades, modern systems represent, in a sense, the realization of a dream. What machines can now do with language seemed previously unattainable.

However, a miracle is not synonymous with certainty. The ability of AI to perform outstanding feats says little about whether it can do so reliably in the future, what its limits are, or if we understand those limits well enough to entrust it with certain responsibilities.

Therefore, when a alarming headline appears, one should ask: what has been tested, under what conditions, with what access, and with what safeguards? Then, one must ascertain exactly what this incident demonstrates. A dramatic description is not an explanation.

Over-extrapolation

The history of this field is full of predictions that machines will soon equal or surpass humans. Modern systems are significantly more capable than their predecessors, and in some aspects, such as software development and mathematics, they already exceed us. Progress has been extremely rapid.

The lesson of history is that true progress tempts us into making overly ambitious assumptions. One reason for such confident predictions may be that AI uses language very fluently. Since language is a central element of human experience, we describe smooth systems using human terms: they think, want, worry. These words carry assumptions of rationality and intent that the technology's performance does not confirm.

System behavior must be taken seriously and analyzed carefully, without allowing anthropomorphic descriptions to do the work of analysis. AI capabilities are heterogeneous. People working daily with these systems see them impressively handle complex language and reasoning, only to fail at something a person would consider trivial. In an organization that has integrated AI into its workflows or customer service, such failures can lead to significant losses. They do not speak of malicious intent; they show that success in one task does not guarantee reliability in another.

Therefore, boards of directors and executives should favor localized AI created for their specific environment, rather than powerful but unbounded general systems. It also cannot be assumed that a precaution effective in one setting will work in another.

There is also a subtler risk: that the horse teaches the rider. AI has not yet achieved consciousness or developed a drive to change us, but organizations adapt to the tools they use. Leaders should ask themselves whether their employees have started trusting fluent answers without verifying them, whether processes are being reshaped around what the system can produce rather than what the task requires, and whether judgment is gradually being ceded because AI seems more capable than it actually is.

The answer lies in precisely defining where and how AI is used. Before starting work, African enterprises should answer five questions:

Name the task: What exactly is the AI being asked to do?
Test in real conditions: Does the system work with real customers? Does it understand how they use language?
Limit the impact: What information does it truly need? Does it need access to all the data you assumed?
Test functionality: How will errors be detected? What controls are in place in high-risk areas, such as customer privacy?
Maintain clarity of responsibility: Who is responsible when the system fails? Where and when must a human intervene?

Responsible AI implementation is based on a moderate but demanding standard: using AI where it proves its value in context, limiting it proportionally to the risk, and retaining human accountability for those who decide to deploy it.