Rahul Patil

OpenAI's CFO Sarah Friar published a piece laying out how the company thinks about delivering more intelligence at lower cost, and the framing is worth sitting with for a minute. The argument is that gains compound across the entire stack: chip efficiency, compute infrastructure, model improvements, and product delivery all reinforce each other. It's not one lever pulling harder, it's multiple layers improving in parallel. That compounding logic is actually what makes the economics of AI hard to predict from the outside. If you only track model benchmarks or headline API pricing, you might miss that a quieter efficiency gain two layers down is what actually changes the unit economics for builders and enterprises. Having a CFO make this case publicly is also a bit of a signal in itself. The conversation is shifting from "look what the model can do" toward "here's why the cost curve keeps moving" which is the kind of narrative that matters when you're trying to close large enterprise deals or justify continued infrastructure investment. Whether the compounding holds at the scale OpenAI is targeting is still an open question. But the framework they're describing, full-stack efficiency rather than model-only improvements, is a reasonable way to think about where durable cost reductions actually come from. https://lnkd.in/gUQBTAan

1

My highlights from PyAI Conf - Mar 10 in SF 𝐀𝐈-𝐠𝐞𝐧𝐞𝐫𝐚𝐭𝐞𝐝 𝐨𝐩𝐞𝐧 𝐬𝐨𝐮𝐫𝐜𝐞 “𝐬𝐥𝐨𝐩” 𝐢𝐬 𝐧𝐨𝐰 𝐚 𝐫𝐞𝐚𝐥 𝐦𝐚𝐢𝐧𝐭𝐚𝐢𝐧𝐞𝐫 𝐭𝐚𝐱 This came up repeatedly: low-signal AI PRs/issues/reviews are increasing triage load for maintainers. Some ideas discussed were creating a "credit score-like system" for contributor reputation management. This has been on everyone's mind, and even outside of PyAI, Pete Steinberger (OpenClaw) has been very vocal about this and Daniel Stenberg (curl) has documented similar pressure at maintainer level. • Pete Steinberger post: https://lnkd.in/dWYgy9dU • Daniel Stenberg, Death by a thousand slops: https://lnkd.in/d9zhu4Su • Daniel Stenberg, The end of the curl bug bounty: https://lnkd.in/d2NtryCQ 𝐌𝐨𝐧𝐭𝐲 (Samuel Colvin / Pydantic) looks like a useful middle-ground runtime The Monty demo and discussion was strong because it���s intentionally minimal + fast, not trying to be a full VM/container replacement. It’s a Python interpreter in Rust designed for safe execution of model-generated code with explicit boundaries to host functions. That gives more flexibility than strict tool-calling, while staying more controlled than open-ended sandbox compute. • Monty repo: https://lnkd.in/dm4tzTJH • Pydantic deep dive: https://lnkd.in/dcqdSQ3P 𝐥𝐚𝐭.𝐦𝐝 - 𝐀 𝐤𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐠𝐫𝐚𝐩𝐡 𝐟𝐨𝐫 𝐲𝐨𝐮𝐫 𝐜𝐨𝐝𝐞𝐛𝐚𝐬𝐞, 𝐰𝐫𝐢𝐭𝐭𝐞𝐧 𝐢𝐧 𝐦𝐚𝐫𝐤𝐝𝐨𝐰𝐧 Yuri’s https ://lat.md stood out as actionable: it creates a markdown-based knowledge graph for codebases with cross-linking, reference checks, and agent-friendly prompt expansion/search. This is interesting for onboarding + architecture navigation + reducing context loss in larger repos. • Repo: https://lnkd.in/du2chFrf 𝐁𝐚𝐮𝐩𝐥𝐚𝐧 - "Git for Data" Ciro’s Bauplan section was compelling because it maps software-dev primitives (branches/commits/transaction-like execution) to data workflows. Their transactional pipeline model is especially important: run in isolated branch, merge on success, keep main untouched on failure. That aligns well with agentic workflows where reproducibility, rollback, and auditability matter. • Bauplan: https://lnkd.in/dhjrwxMQ • Git for Data: https://lnkd.in/dPtDNM4n • LLM quick start: https://lnkd.in/d-wV6d7x What's New In 𝐅𝐚𝐬𝐭𝐀𝐏𝐈 Also worth adding: Sebastián's “What’s new in FastAPI for AI” highlighted that FastAPI is becoming increasingly AI-native in practice, improving dev UX via pyproject.toml-based setup and VS Code/Cursor extensions. He demoed new production-friendly streaming patterns (JSON Lines, bytes, and especially SSE) for real-time LLM responses; also the Pydantic return-type performance, making FastAPI feel optimized for modern AI backends: https://lnkd.in/dcXJSr4Z

65

9 Comments

Chat GPT / Claude / Gemini is not a reliable source of benchmark data. In the last 6 months I've heard: 🩺 The average spend on healthcare is only $X at Series A... 🌴 Did you know we don't have to offer PTO? 💰 Most ML engineers in SF make $Y. 📈 Everyone has plans with mega backdoor Roth functionality now! Reality? 🙈 Health Spend: Data source was *three* founders on a Reddit thread. We talked about how health plans are rated, how we use PEOs to offer stronger benefits at early stage, and how we need to make decisions about offering reasonable, cost-effective coverage to *our* group as part of a comprehensive total rewards strategy. Bonus: Brokers have high-fidelity data sets on your peer group! 🫠 PTO: Legally, yes. Practically this is noise that tells you to achieve goals by controlling how your employees work. If there are issues with velocity or goal attainment, we look at the dozens of other levers to inspire and reach outputs that are more effective than throttling or taking away PTO. 🤑 Salary: It's a smattering from Blind, Levels, Glassdoor... Self-reported salary data is not high-fidelity. $100K-500K salary ranges on job postings are not helpful. What's is? Professionally managed, scrubbed and validated (including levels, not titles) salary benchmark data. Pave is still free if you connect for T1 benchmarks, or you can pay $5-10K for a data set to do this right. It's an investment in fair comp, better offer accept rates, and retention. 🤦‍♀️ 401K: They sure don't outside very big tech, and we aren't Google. We pulled from someone's Hampton conversation cut-paste. Your 10-person startup doesn't need a retirement plan this complex; let's talk about your distinctives and holistic Total Rewards philosophy. GPT can be a thought partner to a busy leader or professional, but it's also highly likely to bring you incomplete data that reinforces the position you ask it to validate, and then help you workshop how to sell it. ❓ Ask non-leading questions, and ask for different points of view, risks, and pitfalls when sparring with LLMs. 💡 Seek out better, high-fidelity data. 🧠 Tap someone with technical expertise that can help you avoid unnecessary mistakes and get the outcomes you're actually aiming for. Fractional Chief Financial and People Officers offer 1-hour consults and can help you quickly cut through the noise and find a high-impact solution for *your* unique business. Thankful for the many fractionals, VC partners, and high-fidelity benchmark reports out there doing the work. Let's help founders build better foundations, scale intelligently, and unlock teams to do their best work! 🚀

36

12 Comments

We've been scaling in one direction ⬆️, with bigger models, more parameters, and more compute. That vertical progress will continue. But there's a second axis we've we've been ignoring, ➡️ outward. Today, one agent figures something out, that knowledge lives and dies with that agent. Every agent optimizes for its own goals. We need systems that optimize for emergent multi-agent goals. Connecting agents or APIs doesn't fix this - that's passing opaque data, not meaning. We need both vectors: scaling up AND scaling out. Sat down with Axios to unpack what that means and why it changes the path to distrbiuted superintelligence. Link in the comments.

54

1 Comment

Arvind Jain sees context graphs emerging from enterprise search and observability. I agree on the destination and disagree on the starting point. Arvind - a fellow IITD alum who's devoted his career to enterprise search - has built a tremendous business at Glean, and his "capture the how, learn the why over time" framing is sharp. But the foundation needs to be different: 1 - Search sits in the read path, not the write path. By the time activity data reaches the index, the decision context - why an exception was granted, what inputs were weighed, who approved - has already been flattened. 2 - Inferring intent from observed patterns is fundamentally different from capturing decision context in the operational flow. One reconstructs; the other records. 3 - The ~80% accuracy Arvind cites on task understanding is impressive, but when agents affect customers, contracts, and compliance, you need an authoritative record, not a probabilistic inference. (The math of compounding errors is brutal - I’ve argued for 99%+ accuracy in past editions of my Substack.) -- We believe orchestration-layer startups - where the context graph is built by actually executing work - are better positioned: 1 - They sit at the point of decision. When an agent triages an escalation or approves a discount, it pulls from multiple systems and applies policies in real time. That's the moment to capture the decision trace, not reconstruct it later from activity signals. 2 - They create decision traces as first-class artifacts. The orchestration layer captures the full picture: what inputs were gathered, what policies applied, what exceptions were granted, and what state existed at the moment of decision. That's not inferred; it's captured - which means enterprises can answer "why did we do that?" definitively. 3- They can learn across customers. Arvind notes enterprise data can't be aggregated for privacy reasons. But orchestration-layer startups focus on bounded workflows - which means they can refine ontologies within a customer and across deployments without ever sharing raw data. -- PlayerZero, Maximor, and Oliv are great examples of this. PlayerZero builds the context graph by automating L2/L3 support - sitting at the intersection where code, config, infrastructure, and customer behavior collide. Maximor AI captures decision lineage by orchestrating finance workflows where reconciliation logic and exceptions actually live. Oliv AI builds it for sales - starting as a co-pilot that performs specific tasks (update CRM, send follow-up) and capturing the decision traces along the way. There will be multiple context graphs within each org, and enterprise search can certainly be a starting point for one of them. Glean's approach creates value for knowledge retrieval and understanding how work flows. The question is whether the authoritative record of decisions will emerge from observability or from orchestration. We're betting on orchestration.

221

48 Comments

Today we’re launching Mercury 2 (https://lnkd.in/gP8pZuDA), the fastest reasoning LLM and the first reasoning dLLM: ~1,000 tokens/sec, ~5× faster than leading speed‑optimized LLMs. What I’m most excited about isn’t just the speed benchmark line, but what this unlocks for building real AI systems: multi‑step agents that don’t stall, voice assistants that can reason inside tight latency budgets, and long-form coding loops that stay in flow. Diffusion changes the generation loop: Mercury works more like an editor iterating in parallel than a typewriter committing one token at a time. We’re hiring engineers across research, systems, infrastructure, and product—If you’re excited about working on frontier LLMs and care deeply about engineering quality, DM me—I’d love to connect (https://lnkd.in/g28qHbNV).

155

8 Comments

Everyone's AI agents (or openclaws...) still need great context. We recently published an article sharing various lessons in Context Engineering while building Dash at Dropbox. We still believe building an index across all your 3rd party apps gets you much higher quality context (and far more reliable). You also don't have to just use the Dash product to access that index. Connecting the Dash MCP to your favorite AI app is one of the fastest ways to bring in a ton of your work context. You get connectors, content understanding, cross-app graphs, and search all in one. Other lessons I touch on: - Using knowledge graphs for cross-app intelligence - Challenges we faced with MCP tool calling - Promising wins using DSPy at scale - Heavy use of contextual LLMs as a Judge Let me know if folks have additional questions in the comments. https://lnkd.in/eqqnvN9F

122

8 Comments