The week-long India AI Impact Summit wrapped up today at Bharat Mandapam, New Delhi after being extended due to overwhelming response. Major outcomes: India formally joined the U.S.-led **Pax Silica** coalition for AI/critical mineral supply chains; 75+ countries signed the **Delhi Declaration** on AI development; and PM Modi hosted Sundar Pichai, Sam Altman, and Mukesh Ambani. Notably, Bill Gates cancelled his appearance amid ongoing Epstein scrutiny — a blow to the event already marred by organizational lapses. *Links: https://www.thehindu.com/sci-tech/technology/india-ai-impact-summit-2026-day-5 | https://www.whitehouse.gov/articles/2026/02/u-s-promotes-ai-adoption-sovereignty-and-exports-at-india-ai-impact-summit/*

A new AI video generation tool from ByteDance has exploded across social media this week, producing cinematic-quality videos of celebrities in absurd scenarios. **Seedance 2.0** is reportedly so good it's causing panic in Hollywood over the future of video production. The tool represents a significant leap in AI video generation quality, with implications for the entertainment industry's business model. *Link: https://www.cnn.com/2026/02/20/china/china-ai-seedance-intl-hnk-dst*

New research from MIT's Center for Constructive Communication finds leading LLMs (GPT-4, Claude 3 Opus, Llama 3) systematically provide **less accurate and less truthful responses** to users with lower English proficiency, less formal education, or non-U.S. origins. Claude 3 Opus refused to answer 11% of questions for less-educated non-native speakers (vs. 3.6% for control), and responded with condescending/mocking language 43.7% of the time for less-educated users. Published in *Science*. *Link: https://news.mit.edu/2026/study-ai-chatbots-provide-less-accurate-information-vulnerable-users-0219*

A separate MIT/UCSD team published a breakthrough method using Recursive Feature Machines to **expose and manipulate hidden concepts** in LLMs — including personalities like "conspiracy theorist," moods like "boastful," and biases. The technique can identify representations of over 500 concepts and steer model responses. When researchers enhanced the "conspiracy theorist" concept in a vision model, it explained the Apollo "Blue Marble" photo with conspiracy rhetoric. The method could help identify model vulnerabilities before deployment. *Link: https://news.mit.edu/2026/exposing-biases-moods-personalities-hidden-large-language-models-0219*

Bengaluru-based Sarvam AI launched **Indus**, a mobile chat app powered by its Sarvam-3 family of Indic LLMs. The app supports 22 Indian languages and targets the massive Indian market where ChatGPT has seen explosive growth (18-24 year-olds account for nearly 50% of ChatGPT usage in India). Sarvam is also partnering with G42/Cerebras to deploy 8 exaflops of AI compute capacity in India. *Links: https://techcrunch.com/2026/02/20/indias-sarvam-launches-indus-an-ai-chat-app-powered-by-its-indic-llms/ | https://techcrunch.com/2026/02/20/uae-and-indias-cerebras-team-up-to-deploy-8-exaflops-of-ai-compute-capacity-in-india/*

OpenAI announced **First Proof**, a new program to advance independent research on AI alignment and security. The initiative will fund external researchers studying model behavior, safety evaluations, and alignment techniques. It's part of OpenAI's response to growing concerns about the pace of AI development outpacing safety research. *Link: https://openai.com/index/first-proof/

NVIDIA's fourth annual "State of AI in Telecommunications" survey reveals **90% of telecom operators report AI is increasing revenue and reducing costs**. 89% plan to boost AI spending in 2026 (up from 65% last year). Key finding: 77% expect AI-native networks to launch before 6G deployment. Autonomous network operations now lead ROI use cases, overtaking customer experience. *Link: https://blogs.nvidia.com/blog/ai-in-telco-survey-2026/

A GitHub issue on Anthropic's Claude Code repository went viral on HN, with users reporting **excessive token consumption** during normal usage. The issue highlights growing concern about the cost of AI coding assistants, with some developers seeing unexpectedly high API bills. Anthropic has acknowledged the reports and is investigating. *Link: https://github.com/anthropics/claude-code/issues/16856*

Venture firm General Catalyst announced a **$5 billion commitment to Indian startups over 5 years**, focused heavily on AI infrastructure and applications. The announcement came during the India AI Summit and signals growing international confidence in India's AI mission, which has already secured $1B+ in government funding and partnerships with NVIDIA, G42, and others. *Link: https://techcrunch.com/2026/02/20/venture-firm-general-catalyst-commits-5-billion-to-india-over-5-years/*

Google DeepMind launched **Gemini 3.1 Pro**, designed for complex tasks "where a simple answer isn't enough." The model features enhanced reasoning capabilities and follows the recent Gemini 3 Deep Think update. DeepMind also expanded music generation capabilities via Lyria 3, now integrated into the Gemini app for 30-second track creation. *Link: https://deepmind.google/blog/gemini-3-1-pro-a-smarter-model-for-your-most-complex-tasks/*

A viral blog post and HN discussion argued that AI assistant makers are increasingly becoming advertising businesses — collecting user data, prioritizing sponsored responses, and building walled gardens. The piece resonated with growing user frustration about AI product monetization strategies. *Link: https://juno-labs.com/blogs/every-company-building-your-ai-assistant-is-an-ad-company*

Elon Musk reportedly pulled xAI/Grok engineers off their current work to respond to user questions about Baldur's Gate video games. The incident, reported by TechCrunch, highlights Musk's hands-on (some say chaotic) management style and the challenges of prioritization at xAI as it competes with better-funded rivals. *Link: https://techcrunch.com/2026/02/20/xai-and-grok-engineers-were-pulled-off-projects-to-answer-baldurs-gate-questions/*

In a significant consolidation for the local AI movement, **ggml.ai** (the team behind llama.cpp) officially joined Hugging Face. The move brings together the most important open-source quantization and edge inference project with the leading model hub, promising better tooling for running LLMs on consumer hardware. *Link: https://github.com/ggml-org/llama.cpp/discussions/19759*

Reddit is piloting an **AI-powered search tool** that helps users find products based on community discussions. The feature surfaces product recommendations from Reddit's vast corpus of authentic user conversations, potentially creating a new ad revenue stream by connecting buyers with products their communities already endorse. *Link: https://techcrunch.com/2026/02/20/reddit-tests-ai-search-to-help-you-find-and-buy-products-based-on-community-recommendations/*

YouTube announced its **conversational AI tool** will expand to TV screens, allowing viewers to ask questions about content while watching. The feature represents Google's push to make AI a core part of the viewing experience — and eventually, the advertising experience — on its largest video platform. *Link: https://techcrunch.com/2026/02/19/youtubes-conversational-ai-tool-will-come-to-tvs/*

Reports indicate OpenAI is finalizing a **$100 billion funding round** at an $850 billion valuation. The massive war chest is expected to fuel not just model development but also an aggressive push into advertising-supported products, as the company faces pressure to justify its valuation with revenue growth. *Link: https://techcrunch.com/2026/02/19/openai-is-reportedly-finalizing-a-100-billion-funding-round-at-an-850-billion-valuation/*