I don't blog often, but when I do, I buy the domain names. 😜 TheNewMoats.com UnitOfValue.com VCrisks.com TheOnDemandEnterprise.com EvolutionOfCloud.com CastlesInThe.Cloud OpenSourceVsCloudCastles.com TheNewNewMoats.com

Excited to be speaking at Modal Runtime alongside an insanely stacked group of speakers! Will be talking about how we're making agents on niteshift.dev faster, cheaper, and more reliable using @modal!

i raised my first 3 rounds of capital ($20M+) before he graduated high school

Excited for @GreylockVC to partner with @CarlSchoeller and the @parallax team to build the leading AI energy company @CarlSchoeller is on a relentless mission to build purpose built turbines that will power the next wave of intelligence

😄

Excited for @SnorkelAI to announce that they have raised $350M at a $3.5B valuation and have crossed $375M in annual revenue run rate. While AI has changed dramatically since Snorkel was founded in 2019 and we at Greylock led their Seed and Series A, Alex Ratner @ajratner and team have been consistent in their belief that data is the very core of AI. Spending years at Stanford, and then at Snorkel, pioneering a different approach to data: treating it as a product and technology problem, combining expert knowledge with deep technology to drive much higher quality data, significant technology leverage, and very attractive unit economics. Snorkel's approach is powering the agentic and RSI wave across the frontier labs, applied AI companies and large enterprises. On a personal note, I couldn't be more proud of Alex. He has navigated an industry that evolves at an extraordinary pace while maintaining unwavering focus on the company's north star and building tremendous business momentum. He is a very special leader. And as a fun relic of history: a photo of the original Snorkel team when they just got started building out of @GreylockVC's Menlo Park office :)

Switching Spark platforms typically means rewriting jobs and re-validating pipelines before savings show up. The drop-in replacement design for AWS EMR removes that step.

Congratulations to @howardting and the @opal_sec team on today’s launch of Opal Zero, an end-to-end access governance platform for AI agents AI agents are forcing companies to rethink the basic unit of enterprise access and the architecture built around it. Howard and I discussed why inventory alone won’t be enough, why human approval processes cannot keep up, and how Opal is approaching the problem. - Agents are becoming some of the most powerful and least governed identities in the enterprise. Few teams can say what a given agent has accessed, who owns it, or why it has that access. - Howard estimates that once permissions are scoped to each task, agents could generate millions of access decisions for every employee. - Today, teams face a bad tradeoff: give agents broad access so they can work, or lock them down until they’re useless. Neither approach works at agent scale. - Opal Zero is built to close the gap between inventory and governance. It finds agents and their credentials, ties each one to an owner and purpose, and gives them scoped access for each task. Full conversation in the comments.

Your lakehouse has a new job: fueling open models. Today we're launching AI training data pipelines on Onehouse for @baseten, @FireworksAI_HQ, and @togethercompute. The problem: Most teams "version" training data by dumping JSONL files into cloud storage. When labels change or eval surfaces a gap, there's no way to trace examples back to source records or rebuild the dataset. What we built: ✅ Quanton curates training examples from lakehouse tables using Spark SQL—join conversations to outcomes, filter by quality, validate schema, export versioned datasets. ✅ Onehouse submits training jobs to Baseten, Fireworks, or Together AI and tracks status. The provider returns fine-tuned weights or adapters. ✅ Lakegres serves point-in-time context during training or eval—queries tables for exact documents and state needed to replay traces. How it works: OneFlow ingests source data → Quanton curates examples → Airflow coordinates → provider trains → Lakegres serves context. The dataset stays reproducible. Engineers can inspect selection queries, trace examples to sources, and rebuild when labels change. Go to quanton.dev and give it a spin.

🔥 📣 Your lakehouse already has the data. Your training pipeline doesn't know how to use it. We're announcing AI training data pipelines on Onehouse for @baseten , @FireworksAI_HQ , and @togethercompute Training data curation today feels like carrying lake water in a leaky bucket. Most teams "version" training inputs by dumping JSONL files into a cloud folder. When a label changes or an evaluation surfaces a gap, there's no way to trace which examples came from where, or rebuild the dataset from updated source tables. Consider a support agent learning from past conversations. Useful examples require joining each conversation to its outcome: Was the issue resolved? Was the answer correct? Did someone intervene? The conversation alone isn't enough. The evidence lives in operational tables. When the model handles a class of requests poorly, the next training dataset needs to reflect updated labels, corrected data, and new selection logic. Teams need the selection query, the source-table versions it read, and the history of those corrections. 🔄 These are familiar data engineering requirements: joins, versioning, lineage, backfills. The lakehouse already handles them. Here's what we built: 1️⃣ Quanton curates examples from lakehouse tables using SQL selection logic—join conversations to resolution status, filter by outcome quality, batch vector search for good/bad examples at scale. It writes versioned exports as training datasets. 2️⃣ Onehouse submits the training job to Baseten, Fireworks, or Together AI and tracks status in the UI. The provider trains the model and returns the fine-tuned weights or adapter. 3️⃣ Lakegres serves point-in-time context during training or eval—queries lakehouse tables for the exact documents, records, or state needed to replay a past trace. The dataset remains reproducible. Engineers can inspect the selection query, trace examples to source records, and rebuild the dataset when labels change or evaluation surfaces gaps. Go to quanton.dev for a free account and give it a spin, let us know!

America is banning AI in schools. China is using AI to create geniuses. Introducing Aristotle: The AI tutor that solves America’s broken education system. heyaristotle.com

Companies spend billions on Spark compute every year. Most assume open-source accelerators can close the gap to commercial engines. We ran TPC-DS 10 TB across 5 engines on the same 11-node cluster to find out. OSS Spark: 12,200s (baseline) Comet (tuned): 9,122s (25% faster) Gluten (tuned): 8,563s (30% faster) Databricks Photon 18.2: 2,550s (5× faster) Quanton: 2,384s (5× faster) Open-source accelerators can get you partway there—under ideal conditions.

just read Derek Parfit's paper on personal identity. TLDR "human identity is less stable than we imagine + memory is partly reconstructed each time we revisit it" - feels v relevant to designing good products around agents / subagent forking

Was great to be at the @GoldmanSachs Comm+Tech conference today with @saranormous, @guruchahal & @jerrychen. We covered a lot of ground on where we are in AI, from open vs. closed models and the pace of software creation to security and the infrastructure required to support it all. The common thread for me is the curve. I’ve never seen a demand curve as steep as the one we’re seeing for intelligence. And every time we think we understand how much people and businesses will consume, we underestimate it. You can see it in software. Developers pushed nearly 1 billion commits to @GitHub in all of 2025, and as of this August, GitHub sees 2.9 billion per month. You can see the impact as agents become a meaningful part of how software gets built. We’re about to produce software at a scale we’ve never seen before. And there’s little evidence today that this curve is slowing down. That has implications across the entire stack. More software means more software to secure. More intelligence requires more compute. And more capable models create opportunities for entirely new products and companies. We’re spending a lot of time debating who will capture which slice of the AI market. I think the bigger question is just how large the pie is going to get.

Thank you @GoldmanSachs for having us on the VC Panel! @glennsolomon @guruchahal @saranormous

Greylock is excited to lead @antiochrobotics's $32M Series A, with participation from @A_StarVC, @Category_VC, @BoxGroup, @IcehouseVenture. AI has shrunk software development cycles from weeks to days, but that speed has not yet reached the physical world. Physical validation processes have been a barrier to physical AI development and adoption speeds, as slow, costly testing bottlenecks progress. Introducing @antiochrobotics, the simulation platform for physical AI. Antioch enables robotics and autonomy teams to develop, test and validate AI systems in high-fidelity simulation, scaling from a single experiment to thousands of parallel evaluations before real-world deployment. By calibrating high-fidelity simulations to each customer's hardware and running them at cloud scale, the platform enables companies to accelerate development, increase test coverage and bring new capabilities to production faster. With experienced, second-time founders and partnerships with leading physical AI teams already in place, we believe Antioch is building the core platform for the future of physical AI. We’re excited to partner with @HarryMellsop, @AlexLangshur @michaeljcalvey and @CollinSchlager for what comes next. Read more on our investment blog (link in comments).

We’re excited to announce that @GreylockVC is leading the Series A in @Antioch to bring the speed of software development to physical AI. We believe the ability to test, learn, and iterate at software speed will be critical infrastructure to enable the next generation of physical AI companies. Dozens of teams including Amazon’s Ring are already using Antioch to accelerate development and bring new physical products to market much faster. We’re thrilled to partner with @HarryMellsop, @AlexLangshur, @michaeljcalvey, @CollinSchlager, and the entire Antioch team on this mission.

hosting din (flour & water 🍕) + onboardings tmrw. ping us if you're in SF & want to onboard IRL

Sometimes I wonder if we spent too much time obsessing over the fundamental primitives in niteshift.dev But then “new model weekend” arrives, and it 100% pays off. Makes vibe coding so much more feasible and fun

ahh ando in grokbot 😍

One of the most exciting (and challenging) applications of physical AI is thinking about how it can accelerate scientific discovery. Imagine a robot that can pipette, handle stem cells, operate existing lab machinery, 24/7 - a physical AI scientist. This is what the new episode of @GreylockVC Change Agents is all about, featuring @michellearning, founder and CEO of @medra_ai. We talk about building an AI-powered lab, how accelerating physical AI lab experimentation is the key to unlocking drug development, and how Medra is automating experimentation today inside pharma companies. 00:00 Teaser 00:56 Medra’s founding mission 04:17 Hypotheses are not enough 13:27 Medra’s differentiation 20:22 The experimentation bottleneck 27:06 Refining the agent harness 29:05 The role of the physical lab 33:23 Managing context across different customers 34:59 Working with models 40:26 Towards personalized medicine and drug discovery 43:15 AWS for life sciences Thank you for joining @michellearning