Operator Collective 🔆

One of the clearest signals we get at OpCo is when one of our own Operator LPs decides to go build something instead of just advise on it. Emily Heath, the CISO who sat on Wiz's board through its $32 billion acquisition by Google and ran security for United Airlines and DocuSign, just did exactly that when she joined Glow as COO. Glow is going after the endpoint, the actual laptops and devices people work on every day, with an AI-native platform that inventories everything running on them and fixes what's risky automatically. Founded by Roi Tiger (previously Onavo, then 9 years at Meta), alongside Omer Singer (ex-Snowflake) and Ophir Arie (ex-Claroty), the team is betting endpoint security needs to shift from reacting after an attack to preventing it before it starts. More on why we're invested in their $1.2B Series B led by Sequoia Capital, Cyberstarts, Greenoaks, and Redpoint from Mallun Yen:

We've been data driven since the beginning, thanks to Anna Jacobson who was Mallun Yen's third hire at OpCo. She's now our Chief AI Enablement Officer and designs all the systems that fuels our Collective Venture Model and OpCo Intelligence: our systematic knowledge-seeking initiatives drawing on 250+ senior operators driving enterprise AI transformation, and knowledge-sharing back to our LPs, portfolio companies, and the broader market. Congratulations to Anna on the well-deserved recognition as a Data Driven VC Top 100 Thought Leader!

Fireworks AI solves a problem most companies run into once they start building with AI: getting a model to run well, cheaply, and at scale on their own data. Enterprises increasingly want AI that reflects their business, not just generic intelligence. Fireworks has built the infrastructure to make that practical. The company now serves more than 43 trillion tokens a day, up from 15 trillion at its last round, and its annualized revenue run rate crossed $1 billion. We were introduced to CEO and co-founder Lin Qiao by founding OpCo LP Erica Ruliffson Schultz at one of our AI Luminaries dinners. That's the Collective Venture Model at work: our 250+ operator LPs open doors to founders building the infrastructure the rest of the industry will run on. Congratulations to Lin, Dmytro Dzhulgakov, George Hu and the whole Fireworks team! The next chapter of AI will be defined by who can turn proprietary knowledge into a durable competitive advantage. More on why we’re invested in Fireworks' $1.5 billion Series D from Mallun Yen ⬇️

Benchmarks are dead (for us). An RSI system makes benchmarks too easy. Give our system a benchmark and it builds its own solution—then beats the state of the art. We just did it on six at once: competition math, scientific coding, long-horizon planning, long-context retrieval, tool use, and web apps. No human tuning. Here's the idea behind it. Most of AI research is betting on one strategy: buy intelligence through massive, incremental weight updates at a cost of billions in compute. At Poetiq, we took a different bet. We don't believe intelligence has to live inside model weights. We treat the LLM as a single component of a larger, self-improving reasoning system. That system—our Metasystem—runs a Recursive Self-Improvement (RSI) loop. Hand it a benchmark, and it autonomously builds the entire harness needed to solve it: code, prompts, tools, and search strategies. No hand-tuning, no privileged access, no fine-tuning of weights. We just pick the benchmark. Even better, we frequently reached SOTA without using the leading model. Because we decouple reasoning from weights, our approach is completely model-agnostic—we use frontier and open-source models as interchangeable tools. As models improve, our Metasystem simply picks them up and keeps getting better. And every harness it builds becomes reusable material for the next problem—search strategies from factory logistics resurface in tool use, and prompt structures from competition math accelerate scientific coding. The benefits compound. Once our system builds and beats harnesses on its own, a fixed test stops being a measure of capability and becomes a to-do item. So maybe it's fairer to say: the static benchmark is dead. Long live the living benchmark—dynamically generated tests (that we’ve asked our own RSI to generate) that no single model can train against. We're now turning the RSI loop on and leaving it on, improving it by solving real problems for our  partners, with complete data privacy, sovereignty and real results. Curious what intelligence beyond Fable, at a fraction of the cost, looks like? Read the full post: https://lnkd.in/gmq5vQJM Want to help us push the frontier of AI & Recursive Self Improvement? We’re hiring. https://poetiq.ai/careers/

"The half-life of any security defense is shorter than we've ever seen, and every prompt, every agent action, every model call is a potential attack vector." Ana Pinczuk joined SentinelOne, the autonomous cybersecurity firm that secures nearly one-fifth of the Fortune 500, in 2022 as a board member and is now President, Chief Product and Technology Officer. In our latest Operator Spotlight, she gets into: 🔸 The three transformations happening at once inside her product and engineering org as SentinelOne crosses $1B in revenue 🔸 How she evaluates fast-moving AI security partnerships with OpenAI, Google, and Anthropic, and why the ones worth keeping are worth more in year two than year one 🔸 Why the dual challenge of AI as threat vector and defense tool is actually a three-prong problem, and what SentinelOne is investing in for all three

💡 AI enablement has a dirty secret right now: the maintenance burden is enormous. Everyone's racing to build custom agents and configure internal tools. But here's what I keep running into: setting something up properly is hard enough. Keeping it running well is a whole other job. The promise of AI is to "automate the work." The reality is a new category of work - prompt engineering, quality monitoring, workflow updates, and constant care and feeding of systems that break in quiet, hard-to-detect ways. This is why a lot of people I'm talking to in the AI enablement space are defaulting to third-party platforms over custom builds, even when custom might be technically superior or more tightly coordinated with existing workflows. Because marginal upfront performance gains don't make up for the ongoing cost of ownership. A few things I'm watching closely: ❶ The agent manager role is emerging as a real job - not an engineer, not a traditional manager, but something new. The people doing it well tend to be those who were good people managers prior to AI, but who also have good technical instincts. ❷ Consumption-based pricing is forcing an honest reckoning. A few months ago it was "token max everything." Now it's "wait, this costs how much?" I think that's actually healthy. ❸ Evaluation fatigue is real. I pulled a list of 90 tools in a particular problem space recently - but there's no way I'm evaluating 90 tools. Nor should anyone. So the first problem is figuring which few to even start with. Here's what I think matters most right now: choosing well, taking the time for well-designed set up, maintaining ruthlessly, and staying honest about the total lifecycle cost of what we're doing.