AI Impact Forum: Democratising AI Resources

Detailed Summary

Question: How can governments and industry jointly ensure AI growth while embedding security and trust from design onward?
Key Insight (Anne Neuberger):
- Trust must be built from the ground up (root of trust). Every AI layer should inherit that trust.
- Users must know what data a model trained on and how decisions are made.
- India already possesses foundational trust‑building infrastructure: Aadhaar (digital identity), DPI (API for financial transactions), and curated national datasets (e.g., healthcare).
- Real‑world impact examples: rural‑village health assistants, real‑time language translation for STEM lectures.
Implication: A national “trust stack” can accelerate responsible AI diffusion.

Speaker: Gokul Subramaniam (Intel)
Core Points:
- Start with workload & user experience – design AI systems around the problem, not the hardware.
- Heterogeneous compute (XPU): Blend CPU, GPU, NPU depending on deployment tier (data‑center training, edge inference, end‑user devices).
- Scalable, affordable architecture: Choose the right compute unit for the right performance‑power envelope.
- Open ecosystem: Promote open standards to avoid vendor lock‑in.

Speaker: Kalyan Kumar (HCL Software)
Argument: Enterprises linger in pilot/POC mode because they build applications first, then force data into them.
Data‑First Imperative:
- AI needs access to clean, well‑catalogued data (metadata, lineage, observability).
- Orchestration & Process Intelligence bridge deterministic workflows with probabilistic AI.
Tipping Point: When data stewardship matures—i.e., data is discoverable, governed, and exposed—AI can scale beyond isolated pilots.

Speaker: Anshu Sharma (Skyflow)
Practical Approaches:
- AI‑as‑exoskeleton: Every employee should have a role‑specific AI assistant that augments productivity (e.g., coding, design, analysis).
- Skill‑shifts in semiconductor space: Engineers must learn to use AI in chip design; early adopters get a competitive edge.
- Broadening access: Pair AI tools with role‑based training so non‑technical staff become power users rather than mere ChatGPT consumers.
Follow‑up Question (Moderator): How to address public distrust in black‑box AI?
Answer (Panel):
- Transparency – reveal training data sources.
- Explainability – show model reasoning (e.g., how a water‑purifier recommendation deviates from norms).
- Continuous learning & monitoring – keep models up‑to‑date and auditable.
- Regulation – act as a safety net; not just a compliance checkbox but a dynamic benchmark system.

Speaker: Anshu Sharma (again) (with nods from other panelists)
Historical Analogy: 1990s India met software talent demand via training “mini‑universities” run by IT firms.
Three‑fold Recommendation:
1. Curriculum Evolution – move from basic coding to systems thinking and AI‑augmented engineering.
2. Multidisciplinary Skills – embed critical thinking, persistence, problem‑solving into degree programs.
3. Self‑directed Learning – encourage individuals to learn, unlearn, relearn via MOOCs; content is abundant, adoption is the bottleneck.

Speaker: Kalyan Kumar (data theme)
Problem Statement: AI requires massive data aggregation, but nations demand sovereign control over sensitive datasets (e.g., health).
Proposed Solution – Federated Learning:
- Train models where data resides, share only model updates.
- Pilot projects with US HHS demonstrated feasibility for privacy‑preserving health‑AI.
- This balances global collaboration with national data protection.

Speaker: Kalyan Kumar (second part)
Three Pillars:
1. Data as Product – Catalog, metadata, knowledge‑graph discovery; treat data as a first‑class asset.
2. Governance & Observability – Track lineage, enforce policies, monitor data quality continuously.
3. Real‑World Constraints – Respect latency, economics, and data‑sovereignty laws; avoid naive centralisation.

Speaker: Sunil Gupta (Yotta) (brief interjection)
Key Observation: AI compute is energy‑intensive; building small, efficient language models is crucial.
FinOps Discipline: Apply cloud‑cost‑management principles to data‑compute budgeting; optimise query‑per‑watt.

Speaker: Gokul Subramaniam (hardware focus)
Security‑First Mindset:
- Application isolation, VM isolation, secure inter‑processor links (Intel SGX, TDX).
- Goal: Confidential AI – protect both data and model during execution.
Democratising AI Benefits:
- Indic‑language translation removes educational barriers for 245 M Indian students.
- Low‑cost compute devices essential for equitable access; AI can amplify impact if hardware is affordable.

Moderator thanked panelists, noted time pressure, and introduced Dr Vishal Sikka for a bonus keynote on AI democratisation.

Root‑of‑trust architecture (identity, data provenance, hardware security) is essential for responsible AI diffusion.
Heterogeneous compute (XPU) enables cost‑effective AI deployment across cloud, edge, and device levels.
Enterprise scaling hinges on data‑first practices: cataloguing, governance, and process‑AI orchestration.
AI exoskeletons: Providing every employee a role‑specific AI assistant accelerates talent democratisation.
Transparency, explainability, and continuous regulation are non‑negotiable for public trust.
Industry‑academia partnership must shift curricula toward systems thinking, multidisciplinary skills, and lifelong learning.
Federated learning offers a path to global AI collaboration while respecting data sovereignty.
Data‑as‑product mindset, combined with robust metadata/knowledge‑graph tools, underpins AI‑ready enterprises.
Energy‑aware infrastructure (small LLMs, FinOps) is crucial to keep AI scaling sustainable.
Hardware‑level security (SGX/TDX, confidential AI) should be baked in from chip design onward.
Language‑translation breakthroughs demonstrate tangible democratisation benefits for education and inclusion.

See Also: