Jason Liu

In the latest Greylock Partners Change Agents I spoke with Arvind Jain from Glean and Ping Wu from Cresta, two companies that are leading some of the largest agentic AI enterprise deployments. A few themes we dug into: - Why agent performance depends less on model improvements and more on enterprise context (systems, permissions, tribal knowledge - see Arvind’s follow-up post about this topic) - How agent sprawl is emerging inside large companies (Glean has customers with up to 2,000 unique agents built) and the subsequent orchestration needs - How they are building an ‘enterprise brain’ by observing human workflows over time, connecting various systems of record, and embedding that context into every agent Both Glean and Cresta started years before “agentic AI” was a term, which means enterprise understanding is the strong base upon which their agents are built. If you’re thinking seriously about agents in enterprise environments, check out the full episode below.

177

5 Comments

I joined Periodic Labs a few months ago after a very rewarding few years building commercial-scale TRISO technology. Watching the Periodic team build something existentially new at the intersection of machine learning and physics has been truly eye-opening. Today, we're sharing a little more on what we've been building. We built automated, high-throughput materials labs (in my hometown of Menlo Park!) to create a closed loop between experiments and machine learning. Our labs generate data, our models help us characterize and learn from the data, and then we iterate. We’re learning first on hard problems in materials science, including superconductors, magnets, and semiconductors. Using only a fraction of the computational resources of the frontier labs, plus months of creating our own experimental data, we mid-trained and performed reinforcement learning on an open-source model to surpass GPT-6 Astra on our analysis benchmark: predicting x-ray diffraction (XRD) from powder synthesis. We intend to extend our prediction capabilities well beyond XRD soon. Grateful to be a small part of the team that Liam Fedus & Ekin Dogus Cubuk have built. https://lnkd.in/gasdfh2s

80

11 Comments

One of our latest companies DN Capital - directionally this is where agentic AI can help augment humans in the workplace rather than replace humans. Congratulations to our partner Guy Ward Thomas and the rest of the DN Capital AI investment team on this newest portfolio company. Nenad Marovac Thomas Rubens Ian Marsh Gülsah Wilke Raoul Oscar F. Jack Pierrepont Maddie Coughlin Celine Xiao

19

Worth paying attention to... Anthropic vs. Cursor. While Cursor was paying Anthropic hundreds of millions in API fees to power its AI code editor, Anthropic seems to have studied exactly what developers wanted, then launched Claude Code to "eat" their biggest distribution partner. The result. Claude Code hit $1B ARR in just six months and has already surged past $2.5B, proving that when you're a "wrapper" startup, you're often just paying your provider to learn how to replace you. Keep an eye on Cursor layoffs...

89

13 Comments

Easy to blow through token budgets Early data shows Architect ’s Liquid Inference (LI) prices below OpenRouter 99% of the time LI routes each prompt through a live auction across inference providers Markets benefit from competition, making them efficient, which ultimately benefits end-users. A per-prompt auction forces providers to bid their true marginal cost every request, which is structurally more efficient than OR’s posted-price model If you’re a builder / developer (or really anyone these days) you’ll benefit and should check out LI — quick & easy base-URL swap liquidinference.ai

2

LMArena just raised $150M at a $1.7B valuation to build AI performance evaluation infrastructure, largely powered by crowdsourced human preference. That’s a huge vote of confidence in humans-in-the-loop benchmarking. But I’m not convinced this is an unambiguously good direction. Crowdsourced evaluation optimizes for what people like, not necessarily what’s correct. At scale, that risks pushing models toward consensus and groupthink rather than accuracy.

5

Jev. Some of the largest companies in the world have been paying tens of millions of dollars to have contractors and experts advise them on how to use LLMs and build complex LLM pipelines at their companies... to setup the software development lifecycle for the next decade.... And then Jev comes out and fucks up everyone's plans and makes those people look like idiots. The game just changed... the LLMs were not ready for this... and the next few months are going to be SUPER interesting. I know the LLMs weren't ready for this because Astra and Fable suck at using Jev and can't wrap their minds around it, so they clearly weren't trained on this kind of AI.

60

19 Comments

Yesterday, two DCVC-backed companies raised about a half billion dollars. In today's funding environment, that's not that exceptional... Together, the two companies just happen to enable gigawatt-scale AI compute, except in the unprecedentedly small space of only about 40 shipping containers, no exotic cooling required. You have to wonder what the US military might do with a gigawatt-equivalent of compute they could fit in a few C5 flights... or a 100MW-compute platform that fits in a single rocket launch... Even that, in DCVC's world, isn't that exceptional. We like to say we help make the impossible and essential, profitable and available. But what is exceptional, and this is a theme with us folks, is that they are each only two of many of our companies delivering robustly on outcomes considered impossible on VC dollars or in reasonable time, at the point when we backed each company from very early on: - a safe, cost-effective, meltdown-proof, truck-portable 1 megawatt nuclear reactor (Radiant, see https://lnkd.in/gF3JAHwS) for American energy dominance - American-made AI semiconductor chips that match or outperform everything else at 100x the energy efficiency, on off-the-shelf models (Mythic, see https://lnkd.in/gWFdvZFT) We were able to make the calls on these, and how they could work together, along with similar calls on all of their brother and sister impossible companies, because of a very deep bench of unified scientific, engineering, operational, and business building experience considered passé by a lot of other folks. It's how we make the exceptional routine.

103

6 Comments