Research

Mixture-of-Experts (MoE) is the architectural innovation quietly powering some of the world's most capable AI models — including GPT-4, Gemini 1.5, and Mixtral. Understanding how MoE works explains why the next generation of AI can be simultaneously more powerful and cheaper to run than its predecessors.

Google DeepMind has launched Gemini 2.5, a next-generation multimodal AI model that demonstrates a significant leap in multi-step reasoning, coding, and scientific problem-solving — setting new state-of-the-art scores across leading benchmarks.

Meta AI has released Llama 4, its most powerful open-source large language model featuring 400 billion parameters. Available for free download under a permissive research licence, the release challenges the dominance of closed proprietary models from OpenAI and Google.

A new frontier AI lab led by a prominent academic researcher is reportedly raising capital at a multi-billion dollar valuation, underscoring investor appetite for yet another contender in the race to build next-generation foundation models.

Researchers at Stanford Medicine have unveiled an AI diagnostic model that identifies rare genetic disorders from patient symptom profiles and genomic data with 94% accuracy — significantly outperforming specialist physicians who averaged 72% on the same test cases.

Chinese internet giant Baidu has unveiled ERNIE 5.0, its flagship large language model, boasting performance metrics that it claims match or exceed OpenAI's GPT-5 on Chinese-language tasks and come close on English benchmarks — a milestone that underscores China's rapid AI advancement despite chip export restrictions.

Researchers at IIT Bombay's Centre for AI Research have released BharatLM 2.0, an open-source large language model trained natively on all 22 scheduled Indian languages — the most comprehensive Indic AI model ever built, designed to serve the 900 million Indians who prefer to communicate in languages other than English.

New reports paint conflicting pictures of how many jobs AI will displace versus transform, fueling political debates over reskilling, social safety nets, and productivity-driven wage growth.

A six-month AIIMS Delhi pilot shows AI triage cut emergency wait times by 40% and correctly escalated 94% of high-acuity patients — with a national rollout under evaluation.

IIT Madras's HydraAttn mechanism cuts transformer training memory by 60% while matching full-attention accuracy — a potential breakthrough for long-context AI on constrained hardware.