Artificial Intelligence: Capabilities and Limits

Narrow AI is what actually exists

Every artificial intelligence system genuinely deployed today, including the most capable ones, is narrow AI: trained to perform one specific task, or a bounded family of related tasks, extremely well, with no ability to transfer that competence to an unrelated task. A system that plays chess at a superhuman level has no capability whatsoever to diagnose a disease, drive a car, or hold a general conversation about an unrelated subject, unless it was separately built and trained to do those things too.

Artificial General Intelligence (AGI), a hypothetical system with human-like flexibility across arbitrary, previously unseen tasks, does not exist and has not been achieved by any current system. This distinction is precisely what a "which of the following are realistic capabilities of current-generation AI" question is testing: options that describe a narrow, specific, already-demonstrated capability are realistic, while options that imply general, human-like understanding, judgement, or creativity across unrelated domains describe AGI, which remains unrealised. Optimising a specific industrial process is a narrow-AI task already in real use; something implying genuine, unprompted original creative judgement in the way a human artist has it edges toward the AGI framing that is not yet real.

Machine learning: how a narrow AI system actually gets built

Machine learning (ML) is the dominant technique behind current AI systems: rather than being explicitly programmed with rules for every situation, a system is trained on large quantities of data, from which it statistically identifies patterns, and then applies those learned patterns to new, unseen inputs. Deep learning, using layered neural networks loosely modelled on the structure of biological neurons, is the specific machine learning approach behind most of the recent, highly visible advances, including image recognition and large language models.

This training-on-data foundation is also the source of AI's most-discussed limitation: a system trained on biased, unrepresentative, or incomplete data will reproduce and often amplify those same biases and gaps in its output, since it has learned only the patterns present in what it was shown.

India's institutional response: the IndiaAI Mission

The IndiaAI Mission, approved by the Cabinet with a budgetary outlay of over Rs 10,300 crore in March 2024, is India's dedicated institutional push to build both AI capability and AI safety infrastructure domestically, structured around a small number of specific pillars.

The compute pillar addresses a genuine bottleneck: training and running large AI systems requires substantial specialised computing hardware, principally GPUs (Graphics Processing Units), which are expensive and were previously accessible mainly to a handful of large global technology companies. India's common compute facility has onboarded more than 34,000 GPUs, made available to Indian startups and academic researchers at subsidised, affordable rates, precisely to prevent AI capability building in India from being gated by who can independently afford this hardware.

The Safe and Trusted pillar established the IndiaAI Safety Institute, tasked with working across academia, startups, industry and government to address AI risk and safety challenges as capability scales up, recognising that expanding compute access and building in safety oversight need to happen together rather than one after the other.

Quick revision points

• Narrow AI (task-specific, no transfer to unrelated tasks) is what exists today. AGI (human-like general flexibility) does not exist yet. A "realistic capability" question is testing exactly this line: narrow, specific, already-demonstrated tasks are realistic; general human-like judgement or creativity across unrelated domains is not.
• Machine learning: systems learn statistical patterns from training data rather than being explicitly programmed with rules. Deep learning (layered neural networks) underlies most recent visible advances. A system trained on biased or incomplete data reproduces and can amplify that bias, since it only learned the patterns it was shown.
• IndiaAI Mission: approved March 2024, outlay over Rs 10,300 crore. Compute pillar: over 34,000 GPUs onboarded, offered to startups and academia at affordable rates, addressing the hardware-access bottleneck. Safe and Trusted pillar: established the IndiaAI Safety Institute for AI risk and safety work across academia, industry and government.