MAINS Current Affairs includes Energy Imbalance & Changing Dynamics of El Niño & Training of Large Language Models (LLMs) by Indian Firms

GEOGRAPHY

1. Energy Imbalance & Changing Dynamics of El Niño

Context

A new study shows that the sharp rise in Earth’s Energy Imbalance (EEI) during 2022 was driven by the shift from a rare triple-dip La Niña (2020–2023) to a strong El Niño, combined with long-term global warming. This explains why temperatures spiked globally in 2023–24. 

Earth’s Energy Imbalance (EEI)

Energy Imbalance = Incoming solar radiation − Outgoing longwave radiation

• When more heat is trapped than released, the planet warms.
• EEI has been rising due to greenhouse gases, but 2022 saw an abrupt jump.

Why was 2022 exceptional?

• A rare three-year La Niña kept warm water trapped at deeper ocean layers
• Surface remained cooler → emitted less heat → energy accumulated
• When El Niño began in 2023, the “lid” was removed
• Heat surged upward → record-breaking warming

ENSO (El Niño–Southern Oscillation)

A natural climate cycle involving Sea Surface Temperature (SST) and atmospheric pressure in the equatorial Pacific.

ENSO consists of:

1. El Niño (warm phase)
2. La Niña (cool phase)
3. Neutral phase

It influences global temperatures, monsoon, cyclones, droughts, and floods.

El Niño (Warm Phase)

• SST in Central–Eastern Pacific warms by ≥0.5°C
• Trade winds weaken

Global Impacts

• Droughts in India, Indonesia, Australia
• Floods in Gulf Coast & SE USA
• Warmer winters in Northern U.S. & Canada
• Global temperatures rise

India Impact

• Monsoon weakens → deficit rainfall
• Drought risk rises
• 2023–24: El Niño caused 4% fall in foodgrain output
• Heatwaves more frequent

La Niña (Cool Phase)

• SST in equatorial Pacific cooler
• Trade winds strengthen → push warm water westwards

Global Impacts

• Floods in Australia
• More rainfall in Canada
• Dry Southern U.S.

India Impact

• Above-normal monsoon rainfall
• Cooler summer in parts of India

SCIENCE & TECHNOLOGY

2. Training of Large Language Models (LLMs) by Indian Firms

Context

Bengaluru-based Sarvam AI unveiled two indigenous large language models (LLMs), highlighting India’s effort to build sovereign, multilingual, compute-efficient AI systems as global competition intensifies.

What are Large Language Models (LLMs)?

LLMs are AI models trained on very large datasets using deep learning to:

• Understand and generate text
• Summarise and translate
• Recognise patterns and predict next tokens

They operate through probabilistic modelling, learning relationships between characters → words → sentences → concepts.

India’s Growing Indigenous LLM Ecosystem

1. Sarvam AI Models

• Two new models: 35B and 105B parameters
• Prioritise efficiency, Indian languages, and local deployment
• Claimed as open-source, pending external verification

2. BharatGen (IIT Bombay-led)

• 17-billion parameter multilingual LLM
• Tailored for education, healthcare & governance

3. Gnani.ai

• Lightweight speech + TTS (text-to-speech) models
• Optimised for Indic languages

How LLMs Are Trained?

(Exam-important process)

1. GPU/Compute Infrastructure

• Training requires massive GPU clusters (thousands of GPUs)
• Continuous training for weeks or months
• High-performance compute remains India’s main bottleneck

2. Data Collection & Curation

LLMs require huge, diverse datasets:

Sources used in India:

• Government publications
• Literature & regional media
• Judicial/administrative documents
• Academic datasets
• Synthetic data (AI-generated)

Why this is critical?

Indian languages are low-resource → quality data = better native performance.

3. Pre-Training (General Language Learning)

LLMs learn general patterns by:

• Predicting next tokens on massive unlabelled text
• Learning grammar, semantics, world knowledge
• Building reasoning foundations

4. Fine-Tuning (Domain Specialisation)

Uses curated task-specific datasets for:

• Government services
• Education
• Law
• Healthcare
• Conversational agents
• Indic linguistic nuances

Tools: Hugging Face, LangChain, LoRA, etc.

5. Alignment (RLHF – Reinforcement Learning from Human Feedback)

Human evaluators rank the model’s responses.

Helps the model:

• Avoid harmful or biased outputs
• Increase accuracy
• Follow instructions better
• Improve safety & trustworthiness

This is the stage separating raw LLMs from usable AI assistants.

Challenges in Training LLMs in India

1. Lack of High-Quality Indian Language Data

• Limited corpora for Hindi + 21+ Indian languages
• Many models rely on translation → English, causing:
  • Higher computational costs (more tokens)
  • Latency
  • Loss of meaning
  • Reduced adoption among regional users

2. Huge Capital Requirement

Frontier models need:

• ₹100–600 crore training cost
• Repeated training cycles
• No immediate commercial returns for startups

3. Compute & Infrastructure Deficit

• Limited access to NVIDIA A100/H100-grade GPU clusters
• Heavy dependence on cloud providers
• Need for national compute platforms (AI compute cloud)

Download Pdf | Study Material | Downloads | Daily Quiz  | FREE Youtube Videos