Deepseek is Jambavanth moment for India
summary
Company builds small fast SLMs since Nov 2022–Jan 2023, aimed at privacy‑first, VPC‑hosted deployments for GDPR customers in Europe because Indian budgets lag. Claim: specialized, cheaper models for defense, healthcare, education outperform generic LLMs for local use. Technical claim about Deepseek: it skips SFT, uses GRPO — relying on reinforcement learning to cut compute and reach strong performance. Strategy: reuse open source pretraining, start domain tuning at stage six/seven, then RL with high‑quality use‑case data. Own your data is the operating principle.
AI-written summary · about
chapters (7)
- 0:00AI sovereignty and DeepSeek
- 1:05Small models and Rabbit AI
- 3:12Europe and India strategy
- 5:41DeepSeek disruption and India's AI
- 7:25Model-building stages and approach
- 9:10DeepSeek as India's Jambavanth
- 11:15Conclusion
related talks
44mApropos DeepSeek
43mHarnessing DeepSeek-R1 and OpenSearch for RAG-Powered Intelligent Info...
01h 00mDeepSeek-R1 Thoughtology: ＜Thinking＞ about LLM Reasoning
34mMapping the Open Model Landscape in 2025
54mThe Future of AI: Building the Most Impactful Technology Together
17mDEEPSERVE: Serverless Large Language Model Serving at Scale
16mNous Research
10mvLLM & Deepspeed Updates