vikram srinivasan

@paraschopra Good luck. I think you are doing an amazing job. My path is exact opposite of you, started as a Prof and now in the start up world. Both are very different muscles. All power to you.

Honestly, I feel like a beginner again. Learning what good scientific research is about and how to build a lab that creates impact. Just like my first time building a startup in my 20s, I'm making mistakes, reflecting, iterating - all of it while shipping. It's so refreshing and fun. The key difference from my startup days v/s now is that I feel (for good or bad) there are more eyes on me, so I have to keep reminding myself that it's okay to fail publicly occasionally as long as each iteration is better than the previous one. Reputation, social image, legacy - all these are mental prisons and devices of fear. What matters ultimately is if you're having fun playing at the edge of your capabilities, which prevents things from either becoming boring or impossible. And that's what I keep recalibrating to.

Everyone’s excited about Claude Code / Claude Cowork in legal & finance. Under the hood, they rely on folder structure + filenames + grep to find context. Fine for clean setups. Fragile in messy enterprise data. At scale, only true semantic context engines work.

Looks good, hopefully quality and impact have also gone up commensurately.

I thought I was an irrational optimist, must have for entrepreneurs, glad this has a name!

Super proud of the work by @OncoStemDx and @ManjiriBakre

In high-stakes domains, safe AI isn’t about bigger models - it’s about better retrieval, governance, and guardrails. Enterprises need source-grounded answers, strict access controls, workflow-bound copilots, and human-in-the-loop validation. If the system can’t show sources, enforce policies, or explain itself - it isn’t safe.

70% of Uber cabs I take in India, the seat belts aren't functional!

AI works when the data is messy, the workflow is painful, and the cost of error is high. That’s the real frontier.

Generic copilots are stalling. Domain copilots with strong retrieval are winning.

AI isn’t winning through “agents” or “AGI.” It’s winning in boring but high-value places: retrieval, synthesis, copilots built around real workflows, and automated reporting. This is where enterprises see ROI today.

@vaibhavbetter My concern, most folks in India, especially VC ecosystem, lack the depth to understand the difference between true deep tech and snake oil.

the scariest part of india deeptech is the sudden (over) marketing of it.

@fooobar We are lucky to have someone as brilliant and kind and motivated as you in India pushing bleeding edge research. Folks who know you enough know the intent is always positive. Unfortunate to see some of the comments.

Not gonna lie - thoda sa dukh to hua tha kuch comments padhkar, but as we say in India, dukh bhi apne hi to dete hain 🙏

Feedback is always welcome, whatever shape or form it comes in - and there is never any shame to take feedback and make changes 🙏 Twitter has been like a family to me. No offence taken, none will be taken in future. Let's do our bit to innovate, in whichever way we can!

This resonates. When we built our market-intelligence stack at Needl.ai, the biggest learning was that the challenge isn’t collection, it’s context - understanding what matters for a specific role, sector, or portfolio. The more personalized the signal, the more valuable the insight.

Holy shit... Someone just built an AI agent that monitors the entire internet for you 24/7 and only messages you when something actually matters. It's called Scouts, and here's how it works:

At the AWS AI Policy dinner #AWSreInvent with @tarunkatial @vikramsrinivasa @FirstHive @awscloud

Why is calling everyone a Member of Technical Staff a trend? At Bell Labs everyone was Member of Technical Staff before all the new fangled titles kicked in.

There’s a lot of buzz right now about “RL + agents” as the answer to everything in AI. I’m genuinely excited about RL – I worked on it during my PhD and in academia a couple of decades ago – but it’s worth remembering the foundations. Classical reinforcement learning assumes there is an underlying Markov Decision Process (MDP): A (mostly) well-defined state space > Actions > Rewards And, crucially, stationary transition dynamics between states Under those assumptions, RL is beautiful and powerful. It shines in closed, well-specified systems: Games like chess, Go, or Atari Domains governed by stable physics and control laws Many operations research problems, where the environment is engineered and predictable But a lot of the places where RL is now being hyped – financial markets, human organizations, multi-agent strategic settings – do not look like stationary MDPs: The state space drifts as new products, regulations, and narratives appear Human behavior (greed, fear, incentives) changes with regimes and feedback Other agents are adapting their strategies in response to you (true game theory, not a fixed environment) Sometimes even what counts as a “state” or “reward” changes over time In those worlds, vanilla RL tends to be brittle. You either: Simplify the problem until it’s no longer that useful, or Chase a moving target and overfit to yesterday’s regime. That doesn’t mean RL is useless there – just that it’s one component in a larger toolkit, alongside: > Robust statistics and risk management > Game theory and mechanism design > Causal and counterfactual reasoning > Human judgment and domain expertise RL is incredibly valuable today for: > Simulation and control > Recommendation, ranking, and local policy optimization > Aligning models with feedback (RLHF/RLAIF) > Structured decision-making in bounded environments We should absolutely keep pushing on RL and agents. But we should also be honest about where the assumptions match reality – and where the world is messier, non-stationary, and fundamentally strategic. Excitement is good. Treating RL as a universal solution to complex human systems is not.

@akshay_pachaar How is this different from a standard reranker? This sounds like what most folks doing RAG do, what we've been doing for 2 years now.

Meta just solved the biggest problem in RAG! Most RAG systems waste your money. They retrieve 100 chunks when you only need 10. They force the LLM to process thousands of irrelevant tokens. You pay for compute you don't need. Meta AI just solved this. They built REFRAG, a new RAG approach that compresses and filters context before it hits the LLM. The results are insane: - 30.85x faster time-to-first-token - 16x larger context windows - 2-4x fewer tokens processed - Outperforms LLaMA on 16 RAG benchmarks Here's what makes REFRAG different: Traditional RAG dumps everything into the LLM. Every chunk. Every token. Even the irrelevant stuff. REFRAG works at the embedding level instead: ↳ It compresses each chunk into a single embedding ↳ An RL-trained policy scores each chunk for relevance ↳ Only the best chunks get expanded and sent to the LLM ↳ The rest stay compressed or get filtered out entirely The LLM only processes what matters. The workflow is straightforward: 1. Encode your docs and store them in a vector database 2. When a query arrives, retrieve relevant chunks as usual 3. The RL policy evaluates compressed embeddings and picks the best ones 4. Selected chunks are expanded into full token embeddings 5. Rejected chunks stay as single compressed vectors 6. Everything goes to the LLM together This means you can process 16x more context at 30x the speed with zero accuracy loss. I have shared link to the paper in the next tweet!