InCommon

Excited to welcome Naveen Bachwani to InCommon’s advisory group. We’ve been lucky to have a strong set of operators and leaders around us, and Naveen’s perspective will add a lot to that bench! Looking forward to building together. #GCC #India

I had the chance to speak with Entrepreneur India on how the next wave of Global Capability Centers in India is being led by small and mid-sized companies. We’re seeing this shift firsthand at InCommon. PE/VC-backed companies are increasingly looking for a faster, lighter way to build in India - one that mirrors enterprise-grade playbooks but fits their scale. And that’s exactly what we’re enabling through our GCC 2.0 model. The article does a great job highlighting and analyzing this shift and what’s driving the next phase of India’s GCC growth. Thanks for driving this conversation, Ayushman Baruah. Link to the article in the comments.

Why do LLMs hallucinate? A new paper from OpenAI argues it comes down to incentives. They compare today’s training and evaluation setup to students taking MCQ exams. When unsure, students guess because there’s no penalty for being wrong, only upside if they get lucky. LLMs learn the same habit. I think it’s a strong analogy. LLMs are “under pressure” to give some plausible answer because we reward them for guessing. If scoring were different: ✔️ +1 for right ✔️ 0 for “I don’t know,” ✔️ –1 for wrong …the incentives would flip immediately (Anyone who’s taken JEE or NEET knows this system well) I put together the full breakdown and why it matters if you’re building with LLMs in today’s newsletter. (link in comments)

This chart sums up the last 7 years of AI: open vs closed models • 2018: GPT-1 makes paper open, weights closed. • 2020–2023: GPT-3.5 and ChatGPT push access behind APIs. Great for usage; weak for learning and reproducibility. • 2023: Llama weights leak. Anyone with 10 to 20 GB VRAM runs real experiments. Derivatives like Vicuna and Mistral sprint. • Sep 2024: O1 adds reasoning signals and jumps about 20 points on AIME-style evals. Closed lead restored. • Jan 2025: DeepSeek R1 reaches frontier-level reasoning with far less capital. Market belief resets; the gap narrows again (Nvidia’s ~$600B drawdown that week was a belief reset). Why does this matter? When the lead is closed, you learn via abstractions; when it swings open, you can read runs, poke failure modes, and ship weird ideas locally. Careers compound when you switch modes quickly. API craftsmanship in closed phases; evals/finetuning/local orchestration in open phases. I break it down in today’s newsletter, coming to your inbox at 5 (link in comments)

I went through Amplify Partners 2025 State of AI Engineering report and highlighted the 6 stats that stood out to me the most. ✅ Half of engineers with 10+ years of software experience have three years or less in AI, and one in ten started only this year ✅ 94% of LLM users run two or more use cases, 82% run three or more ✅ 70% of teams use RAG, 41% fine-tune, with LoRA/QLoRA and SFT the most common approaches ✅ More than half update models monthly, 17% weekly; 70% update prompts monthly, one in ten daily; 31% don’t manage prompts at all ✅ 80% say LLMs work well at work, fewer than 20% say the same for agents, yet fewer than 10% will never adopt them ✅ 65% already use a dedicated vector database, one third self-host, one third managed What these numbers show is a field that has quickly settled on a common stack while still figuring out how to run it well. The pieces are in place: multi-use LLMs, retrieval, fine-tuning, context storage. Now the challenge is making these systems reliable under constant change. That’s where the real work of AI engineering is happening. Credits: Barr Yaron

India’s AI & Tech Market in 2025 1. Market Overview → 27,332 open tech jobs (Global # 2). But demand is shifting toward AI/ML and infra roles, away from generalist software (main reason for massive unemployment in tech) → Tech sector valued at $767B (Global # 4) with 68 unicorns. Fewer new unicorns, more focus on scaling existing ones. 2. Regional Shifts → Bengaluru: Still dominant in AI talent density. → Mumbai & Delhi-NCR: They are catching up through fintech-led AI adoption and R&D investment → Pune & Chennai: Rising as cost-effective hubs with fresh engineering talent pools. 3. Hiring Trends → AI Engineer, ML Engineer, MLOps/Infra roles in high priority. → Senior AI Research salaries touching ₹50 LPA; mid-tier salaries tighter as compared to 2021–23. AI hiring is becoming selective, skill-driven, and geographically distributed. Engineers aligning with niche, high-impact skills will secure the best positions. #ai #tech #jobs