From factories to bazaars, what the India AI Impact Summit’s skilling panel is really arguing for

At Bharat Mandapam’s Room 9, the session titled “AI and the future of skilling” sounded, on the surface, like the usual education panel: degrees vs skills, disruption vs continuity, and a friendly dose of optimism. But listen closely and a sharper argument emerges. The panellists were not just talking about AI as a teaching aid. They were sketching a new operating system for skilling, one that tries to do three hard things at the same time: personalise learning at population scale, keep curricula from going stale in months, and reduce the cost of trust in credentials, employers, and marketplaces.

The session opened with a tight, almost deliberately eclectic lineup, and that mix is why it worked. Dr Vijay Kumar of MIT brought the long view on education at scale and the invariants worth protecting. Ashish Kulkarni, founding director of the Indian Institute of Creative Technologies (IICT), grounded the conversation in a creative economy where tools and formats mutate faster than semesters. Dr Manish Kumar, former NSDC MD and now an academic and advisor, came at it through institutional economics and competency frameworks. Shankar Maruwada, co-founder of the EkStep Foundation and one of the key figures behind India’s digital public goods thinking, argued for skilling as infrastructure. And Dr Ramesh Raskar of the MIT Media Lab pushed the most provocative frame: the difference between a centralised “AI factory” and a decentralised “AI bazaar”, where people create their own agents as digital representatives.

The first thread, “quality at scale” stops being a slogan

Dr Vijay Kumar anchored the conversation in an MIT arc that begins with MIT OpenCourseWare in 1999 and moves to a newer bet: modular, updateable AI learning designed for very large audiences. The important point was not the platform name, it was the redefinition of scale. In the OpenCourseWare era, scale could mean broadcasting one corpus to millions. In 2026, he argues, learners arrive with radically different goals, contexts, and preparation levels. Scale becomes a vector, not a scalar: magnitude plus direction. In plain terms, reaching a billion learners is no longer about distribution, it is about differentiation.

That framing matters because it changes what success looks like. “Quality” is not prestige packaging. It is active, practice-based learning (mind and hand), but delivered in ways that can adapt to adult learners and shifting needs. MIT’s new Universal AI learning experience is explicitly structured as short, self-paced modules, with foundational “AI fluency” plus vertical modules for domain application, designed so content can be refreshed quickly as tools and practices change. That “updateability” becomes a recurring motif through the whole panel. Because if the half-life of skills is shrinking, then the half-life of curricula is shrinking too.

The second thread, curricula that mutate every semester

Ashish Kulkarni brought the most concrete “this is already happening” energy, from the perspective of media, entertainment, and immersive tech. His claim was blunt: AI has entered every process in the modern content pipeline, from ideation and scripting to final output, at a speed that breaks the classic semester model. The result is a design problem for institutions: how do you teach fundamentals that endure, while tooling and workflows churn underneath?

His answer is a kind of two-layer curriculum. Layer one is the grammar of storytelling, creative fundamentals taught early. Layer two is relentless retooling, because production methods and distribution economics are moving targets.

This is where the panel’s argument intersects with policy and industry scaffolding. The Indian Institute of Creative Technologies (IICT) is being positioned as a national centre of excellence for AVGC (animation, VFX, gaming, comics, extended reality), and Budget 2026-27 has included funding for talent development and “content creator labs” in 15,000 schools and 500 colleges.

Whether those labs deliver outcomes or just optics will depend on how tightly they connect to real production pathways and assessment that employers trust. But the direction is clear: India wants creative and immersive skills to be treated as a mainstream pipeline, not an extracurricular indulgence.

Kulkarni’s most interesting twist, though, was that the industry is not only about artificial intelligence, it is about emotional intelligence and behavioural intelligence in a “screen era”. That is less a soft sentiment and more an admission that the differentiator for humans may increasingly sit in taste, judgement, and narrative empathy. He even name-checked the “human in the loop” role of the editor as something AI cannot reliably replace yet, because creative sense is contextual and experiential.

In other words, in at least one high-growth sector, the skilling argument is not “learn prompts”. It is “learn fundamentals, then learn how to constantly relearn tools without losing your voice”.

The third thread, education’s lag is structural, not motivational

While the industry has sprinted from 1.0 to 4.0 while education is still behaving like 2.0, was how Dr Manish Kumar diagnosed the current state of affairs. The obstacle is not lack of awareness, it is the system’s desire for predictability and stability. Institutions are built to resist rapid change.

His prescription was competency-based learning, translating syllabi into competencies that map to industry needs, and updating those mappings continuously. The reasoning is practical: even industry struggles to predict what it will need next year because tooling is improving so fast, with coding assistants and platforms changing materially within months. He cited Replit as an example of rapid improvement. (The broader point holds regardless of which tool you pick.)

But his second idea is the one that quietly complements everything else on the panel: intergenerational learning as a bridge between “knowing what to do” and “knowing how to do it”. Younger learners often pick up tools quickly but lack problem framing and purpose, older workers often have judgement and context but struggle with tooling. Put them together in mentorship-heavy environments and change can accelerate without waiting for institutional reform to crawl through committees.

It is a practical social design for adaptation, and it sidesteps a common trap in education debates: assuming the classroom alone is the unit of change.

The fourth thread, DPI as the missing plumbing for skilling markets

Then Shankar Marwada reframed the question with a distinctly Indian instinct: treat skilling like infrastructure. His point starts with an observation India has already lived through. When problems grow nonlinearly, solutions have to scale nonlinearly too. That is the story of Aadhaar, UPI, DigiLocker and DigiYatra as DPI (Digital Public Infrastructure), not as apps. The “road and cars” analogy matters here, because it argues for neutral rails that many innovators can build on top of.

Applied to skilling, the bottlenecks are familiar: language barriers, paper-based trust, and weak discovery of jobs, training, and scholarships. His version of the future is not just “AI tutors for everyone”, it is marketplaces that find you (the scholarship finds the student), and credential systems that reduce the cost of trust.

This is not purely theoretical. EkStep has been involved in building education-facing digital public goods, including learning infrastructure and concepts around registries and verifiable records. The “trust layer” is the difference between skilling as self-improvement and skilling as economic mobility. If employers cannot trust credentials, the market collapses into reputational gatekeeping. If they can, opportunity can move faster.

The capstone, escaping “AI factory mode” with decentralised agents

Finally, Ramesh Raskar pushed the panel into a more provocative frame: the risk of “factory mode” AI, where a handful of companies build the dominant models and everyone else becomes downstream labour, even if jobs do not disappear outright. He described the danger as “intelligence slavery”: work that is technically employed but creatively hollow.

His counter-model is decentralisation, a “bazaar era” where people and communities can create their own micro AIs (agents) as representatives, and where discovery and collaboration happens across an open network rather than inside a few closed platforms. That vision maps directly to his work on Project NANDA, described as infrastructure for networked AI agents in decentralised architectures. 

The PowerPoint analogy he used is telling: Microsoft sells the tool, but the pride sits with the person who crafts the deck. In this model, the agent becomes the user’s “digital envoy”, a personalised capability you own, train, and deploy. The link back to DPI is clear: decentralisation needs identity, discovery, trust, and interoperable rails, otherwise it collapses into chaos or recentralises by default.

So what’s the takeaway?

Taken together, the session was not merely about AI in classrooms. It was proposing a new national loop involving a modular learning layer that can be updated at speed (quality at scale, but personalised). This would be aided by a competency layer that translates learning into employer-readable signals (what you can do, not what paper you hold). Followed with a DPI layer that reduces friction in discovery, language access, and trust (verifiable records, marketplaces that match people to opportunities). And lastly there ought to be an agentic layer that stops AI from recentralising power by making personal and local AIs viable (the bazaar instead of the factory). 

If that loop works, it changes the most stubborn constraint in skilling: not the supply of courses, but the conversion of learning into livelihood at scale. Of course, none of this is guaranteed. Personalised learning can become surveillance if governance is weak. Credential systems can become exclusionary if standards harden too early. Agent networks can become messy if safety, authentication, and accountability are treated as afterthoughts. And “every semester must change” is a thrilling line right up until it hits faculty workload, assessment integrity, and procurement cycles.

But the panel’s core message is pragmatic: India does not need a single heroic reform. It needs rails that let many reforms happen, locally, continuously, and credibly.