IT, Artificial Intelligence & Cyber Security

GS Paper: GS Paper III | Subject: Science & Technology | Last updated: 2026-07-20

---

Prelims

(Key facts, data, schemes, laws, organizations — MCQ-ready points)

AI Solves 80-Year-Old Maths Problem (Indian Express, 04-06-2026)

• An OpenAI internal model (maker of ChatGPT) reportedly solved the planar unit distance problem — open since 1946
• Problem first posed by Hungarian mathematician Paul Erdős (such challenges are nicknamed "Erdős problems")
• Significance: claimed as a landmark in AI reasoning — mathematician Tim Gowers (Univ. of Cambridge) noted a human producing it would merit publication in a top maths journal
• Caveat: AI models still "hallucinate" (produce false output); claims of AI maths proofs have drawn scepticism before

AI Geopolitics — "Trusted Tech" & Export Controls (The Hindu, 15-06-2026)

• At the India–France summit a Joint India–France AI Working Group was created; Macron attacked the US ban on non-US citizens accessing Anthropic's frontier AI models as "closing up AI models" into a "power tool"
• India & France pitched "cooperative, human-centric, trusted AI" against AI fragmentation by the US–China duopoly; Macron claimed France's LLMs can rival US/China
• Signals AI sovereignty / techno-nationalism: model access, compute, and standards are becoming instruments of geopolitics

Sarvam — Sovereign-AI Funding (The Hindu, 16-06-2026)

• Sarvam, an Indian "sovereign AI" firm, raised $234 mn (first close of a $300 mn Series B) at a $1.5 bn valuation; HCLTech led with $150 mn (lead strategic investor), alongside Bessemer, Khosla Ventures and Peak XV — a step toward a "globally competitive" Indian AI ecosystem (cf. IndiaAI Mission)

Microsoft's Pay-as-You-Go Model for AI Agents (The Hindu, 21-06-2026)

• Microsoft introduced a consumption-based ("pay-as-you-go") pricing model for its AI agents — pay by usage rather than only fixed per-seat licences — moving AI billing toward the economics of cloud computing (rent compute, pay for what you use)
• Why subscriptions don't fit AI: flat per-employee fees assume every user gets equal value, but AI usage varies enormously (one worker summarises a meeting weekly; another runs document analysis all day); AI agents that autonomously perform tasks widen this gap further
• The supplier-side reason: unlike selling Word once, every prompt/response/agent action consumes compute → ongoing token-generation cost for the provider; usage-based pricing aligns price with cost (Microsoft leverages its Azure metered-billing experience)
• Caveat: usage-based pricing makes bills unpredictable — as AI agents embed in workflows, organisations may struggle to forecast/budget spend (the "surprise cloud bill" problem)

Amazon's Additional $13 bn for AI & Cloud in India (The Hindu, 26-06-2026)

• Amazon CEO Andy Jassy (after meeting PM Modi) announced an additional $13 bn to expand AI & cloud (AWS) in India by 2030 — within 6 months of a $35 bn commitment; total AI/cloud spend >$21 bn (2026–2030), $48 bn over 5 years, and cumulative India investment >$88 bn (2010–2030)
• Will expand AWS data-centre capacity in Mumbai & Hyderabad (custom AI chips, managed AI services); aligns with India's priorities of "democratising access to AI", digitising small businesses and enabling exports — part of a wave of hyperscaler data-centre FDI (India's data-centre market is projected to cross $200 bn)

Keeping Humanity at the Centre of AI — Pope Leo XIV's Encyclical (The Hindu, 26-06-2026, op-ed)

• A senior-advocate op-ed frames the ethical-guardrails debate around Pope Leo XIV's encyclical "Safeguarding the Human Person in the Time of Artificial Intelligence", which warns against AI-driven "new forms of dehumanization" and the "idolatry of profit", urging humanity to remain "profoundly human" and to anchor AI in the dignity of the individual
• Promise vs peril: AI's gains (cancer screening, education access, disaster/weather forecasting, targeted aid) are weighed against risks — a "global epidemic of stress", a predicted "useless class" from job disruption, data-privacy loss, misinformation, electoral manipulation, rogue weapons, surveillance/censorship
• Digital sovereignty: control over data is tied to national security and strategic autonomy → the op-ed (and PM Modi at VivaTech 2026, Paris and the India-AI Impact Summit 2026, New Delhi) argues for a robust, enforceable global regulatory framework, not voluntary/non-binding commitments — to "democratise access to frontier AI" and build a trustworthy ecosystem

Embodied AI — The Next Frontier in Robotics (The Hindu "Keyword" explainer, 20-07-2026)

• The trigger: on 14 April 2026, Boston Dynamics and Google DeepMind announced that Spot — the yellow quadruped robot dog that had spent most of its life running narrowly scripted routines — is getting an AI brain. Gemini Robotics-ER 1.6 is being folded into Spot and into Boston Dynamics' Orbit inspection platform, giving stronger spatial reasoning, autonomous decision-making, and continual learning inside complex industrial sites
• What "embodied" actually means (the key conceptual point): the common error is to read embodied AI as merely "AI in a robot body." The deeper claim is that the body is not a delivery mechanism for intelligence — it is part of the computation itself. Intelligence is distributed across brain, body and environment, not confined to one part
• The foundational texts and figures:
• Rolf Pfeifer (longtime director, AI Laboratory, University of Zurich) and Josh Bongard (University of Vermont), How the Body Shapes the Way We Think — argue thought is tightly constrained, and simultaneously enabled, by the body
• Rodney Brooks (MIT, late 1980s–90s) — argued against the then-dominant symbolic-AI paradigm that robots do not need internal world-models or elaborate planning to act intelligently. His "subsumption architecture" — layered simple reflexes tightly coupled to sensors and motors — produced robust real-time behaviour without programming any representation of the world
• "Morphological computation" — offloading cognition onto the body. Two examples to memorise:
• A passive-dynamic walker strolls down a slight incline with no motors and no control system at all, purely because leg geometry and joints are shaped to do the computing
• A soft, compliant robot hand grasps an oddly-shaped object without a controller that has modelled that shape, because the material itself deforms and adapts
• Why it is hard — two named obstacles: unlike chatbots trained on text/images/video, embodied AI must master gravity and balance across countless physical scenarios; and it faces the "simulation-to-real gap" — success in simulation rarely transfers to the real world
• The gap in numbers: the embodied-AI market is projected to reach $23 billion by 2030, yet most humanoid robots last only ~90 minutes on a charge, and policies that succeed 95% of the time in the lab drop to roughly 60% in the real world. "The gap between demo and deployment is the field's central, unglamorous problem."
• Embodied AI vs neuromorphic AI — a likely Prelims distinction:

	Embodied AI	Neuromorphic AI
Question it answers	Where intelligence lives	How the processor is built
Core claim	Cognition is distributed across brain, body and environment	Hardware/algorithms mimic biological neurons
Hardware stance	Agnostic — can run on an ordinary GPU cluster	Central — spiking neural networks (SNNs)
Mechanism	Morphological computation	Neurons integrate signals and fire only when a threshold is exceeded
Key advantage	Offloads computation onto physical form	Power efficiency — only actively spiking neurons consume energy; suits time-sensitive tasks like motion sensing

• The two increasingly overlap: research on "embodied neuromorphic intelligence" puts spiking, event-driven chips inside physical robots precisely because low power draw and fast response suit real-world, always-on tasks
• Evolutionary computation — designing the body, not just the brain: Yaochu Jin (Alexander von Humboldt Professor, Bielefeld University) argues neural control and physical form must be developed together, the way organisms grow nervous systems and bodies in tandem — rather than bolting an AI model onto whatever frame engineers happened to build. Evolutionary methods let simulated populations of robotic forms compete and replicate on task performance before anything is built, addressing the bottleneck that robot bodies are frequently mismatched to their tasks and re-engineering hardware each time is slow and expensive
• The unresolved problems (a clean five-point list): (1) the sim-to-real gap — policies trained cheaply in simulation degrade sharply on real hardware; (2) speed mismatch — sophisticated reasoning models are too slow for a limb that must react in milliseconds, forcing a split between heavy "thinking" off-device and lighter reflexive control on-device; (3) short runtimes from battery drain and vulnerable components; (4) data scarcity — there is no internet-scale corpus for embodied action (efforts: the Open X-Embodiment dataset, and Generalist AI's GEN-0, pretrained on hundreds of thousands of hours of manipulation data — but tens of millions of hours are needed); (5) governance — safe deployment depends on sensors, hardware robustness, operational design limits, human interaction, cybersecurity and organisational processes, "not algorithms alone", so regulators must start asking what counts as proof for deploying learning-enabled robots in the real world
• Where it already is: Boston Dynamics' humanoid Atlas performed on the football field during FIFA World Cup 2026 (demonstrating adaptation to uneven surfaces); China's Unitree G1 robots performed alongside a dancer on America's Got Talent Season 21; Spot now assists e-commerce delivery agents

---

Mains

(Analysis, dimensions, significance, critique, policy angles — for 10/15 mark answers)

Embodied AI — Why the Next AI Wave Is a Systems Problem, Not a Software One (The Hindu, 20-07-2026)

• The closing insight, worth quoting: "Intelligence was never just a matter of better software sitting inside better computers. It was also, at least partly, a matter of testing out new form factors." This reframes the AI debate: the LLM boom scaled cognition; embodied AI must scale action — and action is bounded by physics, energy and materials, which do not obey scaling laws
• Why this matters for India specifically:
• Manufacturing & the China comparison: embodied AI is the technology layer atop robotics and precision manufacturing — where India is weakest and China (Unitree and peers) is scaling fastest. The $23 bn by 2030 market is one India could miss the way it missed semiconductors, unless National Robotics Strategy-type efforts and PLI for electronics/machine tools are connected to AI policy
• The labour question is different from the LLM one: LLMs threaten white-collar cognitive work; embodied AI threatens manual and service work — the sector absorbing most of India's informal and low-skill labour. For a country counting on a demographic dividend and labour-intensive manufacturing (see poverty-employment), the automation of physical labour is a more direct threat than chatbots — and arrives just as Gen Z graduate unemployment is at 29–37%
• The advantage India does hold: the field's binding constraint is data scarcity, not compute. India's scale, diversity of physical environments and manufacturing/logistics volumes are a potential source of embodied training data — a strategic asset if data-governance frameworks permit its use
• The governance argument the piece makes explicitly: embodied AI must be regulated as a systems challenge — safety depends on hardware robustness, sensors, operational design limits, human-robot interaction, cybersecurity and organisational process, not model behaviour alone. This is closer to aviation or automotive safety certification (type approval, operational design domains, incident reporting) than to the content-moderation model that dominates current AI regulation. India's DPDP Act 2021 and the IT Rules address data and content; neither governs a machine that moves through physical space and can injure someone. The sim-to-real gap is a safety problem, not just an engineering one — a 95%-to-60% reliability drop is a liability question
• Liability parallel worth drawing: exactly like the missing Space Activities Act for private launch (see space-technology), India has no statutory liability regime for autonomous physical systems — leaving injury, product liability and insurance to contract and general tort law. Both cases show the same pattern: capability arriving ahead of statute
• The philosophical dimension (usable in GS4/essay): if intelligence is distributed across brain, body and environment, then the sharp human/machine cognition boundary blurs further, and the "is it really thinking?" question becomes less tractable. It also supplies a counter to pure digital-utopianism: intelligence is embodied, situated and materially constrained — a claim with roots in phenomenology (Merleau-Ponty) and, in Indian thought, in traditions that treat body and mind as continuous rather than dual
• UPSC angle: embodied AI vs neuromorphic computing & SNNs, robotics & advanced manufacturing, automation and the future of manual labour, AI safety & the sim-to-real gap, liability for autonomous systems, sectoral vs horizontal AI regulation, India's robotics competitiveness, data as strategic resource

A Moral Compass for the AI Revolution (The Hindu, 26-06-2026)

• Beyond efficiency — the "destiny of intelligence": the deepest question is whether functional efficiency and material abundance should prevail over human dignity, emotion and self-worth — AI can now replicate cognitive skills and even read emotions, so society must decide whether it is "ready for a new narrative of humanity". Thinkers cited (Ortega y Gasset, Eagleton) and the papal encyclical supply a humanist-centric benchmark: the individual at the centre of every AI decision
• From ethics to governance: moral vision must translate into enforceable, sovereignty-respecting global regulation (India's pitch at VivaTech & the India-AI Impact Summit) — voluntary codes are inadequate against data-privacy, misinformation, electoral-manipulation and autonomous-weapons risks; ties to AI sovereignty/digital sovereignty and India's bid to democratise frontier AI
• The labour & inequality dimension: the "global epidemic of stress" and a feared "useless class" make social protection, reskilling and inclusive growth integral to AI policy, lest AI "exacerbate inequalities" — links to ethics (GS4) and the future of work
• UPSC angle: ethical AI governance, human-centric/"trusted AI", digital sovereignty & data, global AI regulation (enforceable vs voluntary), AI & employment/inequality, technology vs human values (GS4)

AI Moving from Tool to Researcher (Indian Express, 04-06-2026)

• Reasoning leap: Solving a long-open research problem (not just retrieval/pattern-matching) signals AI moving into genuine scientific discovery — could accelerate research across maths, materials, drug design
• Verification challenge: AI proofs need human expert checking; "hallucination" risk means outputs can't be trusted blindly → raises the reproducibility & trust problem in AI-assisted science
• Strategic/sovereignty angle for India: Frontier reasoning models are concentrated in a few US/foreign labs → reinforces the case for indigenous compute, AI talent, and the IndiaAI Mission to avoid dependence in a foundational technology
• UPSC angle: AI and the future of R&D, frontier models, AI safety/hallucination, India's AI strategy (IndiaAI Mission), tech sovereignty

AI Sovereignty & Fragmentation Risk (The Hindu, 15-06-2026)

• Model-access controls (e.g., the Anthropic non-US ban) turn frontier AI into a geopolitical chokepoint like chips — pushing middle powers (India, France) toward indigenous models, sovereign compute, and "trusted-AI" coalitions; strengthens the IndiaAI Mission case and India's stake in open, interoperable AI governance (GPAI)
• UPSC angle: AI sovereignty, export controls on frontier tech, IndiaAI Mission, GPAI/global AI governance, strategic autonomy in technology

The Limits of AI — Extrapolation & Judgment (The Hindu, 16-06-2026)

• AI fails at the rare, high-stakes event: a Science Advances study found leading AI weather models (GraphCast, Pangu-Weather, Fuxi) match or beat physics models on normal weather but systematically under-predict record-breaking extremes (heat/cold/wind). It's an extrapolation problem — AI interpolates within its 1979–2017 training data but cannot anticipate the unprecedented, unlike physics-based models (ECMWF HRES). Risk: under-stated extremes → inadequate disaster early-warning/response
• Information ≠ judgment (AI in medicine): LLMs (ChatGPT/Claude) democratise cancer knowledge and informed consent, but a JAMA study shows accuracy drops on complex, context-dependent cases; presenting risks without clinical judgment can erode doctor–patient trust and cause patients to delay care — a caution for AI in any high-stakes domain
• UPSC angle: AI strengths vs limits (interpolation vs extrapolation; data-driven vs physics-based), AI in healthcare & early-warning, human-in-the-loop, AI safety & governance