
🚨 Big news: We’ve raised $160M for @Work_Bench Fund IV to back Seed-stage founders with massive enterprise ambitions. Enterprise is in our DNA and with Fund IV, we’re quadrupling down.👇

Enterprise VC @Work_Bench | We Lead Seed Rounds with $2-4M Investments | Organizer @NYETM, @KauffmanFellows 19 | My better half is @MichaelaLehr | 305 Native

🚨 Big news: We’ve raised $160M for @Work_Bench Fund IV to back Seed-stage founders with massive enterprise ambitions. Enterprise is in our DNA and with Fund IV, we’re quadrupling down.👇

Check out @ProbyShandilya's latest pod drop!
Episode 2 of Thinking About Thinking is live! In 2018, I published a book titled Surrounding Factors. The essential premise was that how a basketball player performs is not just a function of their individual skillsets and talents - how they fit with the "surrounding factors" around them (the coaching style, the other players on their team and complementary skillsets, the strategy the team employs, the overall environment, etc) is a *very critical* determinant. Whenever I talk about this with folks in the startup ecosystem, I joke that this is essentially "founder market fit for basketball players" - because it's true! The surrounding factors mental model applies universally, whether it's for players on a sports team, people within an organization, or even products within certain time periods (the "why now") and surrounding environments (early customers/ICPs, business models, GTM motions, etc). Hope this is a fun listen!

Pumped to share that we’ve led a $5.7M round into @EarthlyTech AI coding agents have 100x’d developer productivity, but as more code gets written by agents, it’s getting harder to ensure that the same standards are followed across the organization. Earthly Lunar 🌖 helps teams enforce engineering standards across thousands of repos. Teams can define a rule once and check it as AI agents write code, when developers open PRs, and before anyone deploys. Lunar’s enterprise footprint already spans more than 100,000+ repositories across some of the largest companies in financial services, communications, defense, and more. Most importantly, Earthly is led by @VladAIonescu, a deeply technical founder who has turned himself into a true enterprise seller. Working with Vlad has been an absolute pleasure — he's exactly the kind of founder we love backing at @Work_Bench! Plenty of engineers can build great products, but fewer can walk into a Fortune 500 company and close a massive enterprise deal (time and time again). Vlad can do it all! Here are a few ways Earthly’s already helping customers: → A large bank turned 65 pages of software delivery policy into 41 automated guardrails that run in its existing pipelines. → A global communications company immediately rolled out Lunar across 11,000+ repos and 2,000 engineers. → A large employment tech company brought engineering standards to 90,000+ repos that its golden-path templates couldn’t reach. → A defense tech company automated the compliance checks and software bills of materials that used to take weeks for every DoD delivery. This is just the start for Earthly Lunar. Next stop, the moon! 🚀 🌖 Read more about their coverage with the links in the comments.

this is the stat I was hoping for when we got Klay have Bam and Giannis dominate inside and let Klay just catch and shoot loving this 🔥
He only dribbled on one field goal

🔥🔥🔥
incredibly excited to be partnering with @OpenAI on embedded assessments! we'll be embedding interpretability researchers who will use the @GoodfireAI platform to monitor and debug training runs to improve safety and alignment our primary audience for our assessments are internal OpenAI researchers, but we'll publish our findings as the collaboration progresses

Last night, @fendien and I hosted an AMA Dinner with @jaykreps (founder/CEO of @confluentinc and creator of @apachekafka) and it was a blast. Confluent has always been one of my favorite enterprise software companies to study. It all started when Jay, alongside Neha Narkhede and Jun Rao, created Kafka, an open source data streaming platform, within LinkedIn. They had deep conviction that continuous, real-time data was something that had to exist in the data infrastructure stack, and before long the market agreed with them…they experienced pull from outside customers and decided to start a company, Confluent, around it. They started the company in September 2014, launched the platform in Jan 2015, released their cloud product in 2017, and went public in 2021 en route to being bought by IBM earlier this year for $11B. So many incredible learnings from Jay last night. Sharing a couple here: - He shared a very elegant framework to think about GTM: think about how a customer buys your product and figure out how to engineer that as much as possible. Lessons from other companies and sales leaders are great, but what matters most is anchoring to your unique buying decision and trying to accelerate that as much as possible. - “It’s not about what can we do, it’s about what must we do” - this was one of the core ideas that drove the cloud product in the early days of Confluent. There a lot of things that a company can do, but there’s one or two things that a company must do to get to their North Star / where they want to be in 3-5 years. Having that sense of focus and clarity is critical. - A company should feel like a startup for as long as possible and you as the founder should be in the trenches for as much of the company’s life as possible, but at the same time, the CEO job starts from day one. Super important for founders to deeply identify with the CEO role as early as possible and learn the operational and management skills they need to in order to run the company effectively. Left the dinner so energized - few things beat spending time with one of the greats in enterprise software, and many of the timeless principles around company creation / leadership / hiring / enterprise sales will stay timeless. Thank you Jay for spending your Thurs night with us and thank you to everyone for coming and bringing great energy and an insatiable curiosity!

the long awaited technical follow-up to Neural Networks are Already Clean is here! it dives deep into detailed neuron experiments, and we release thousands of neuron pages with neurons representing everything from the birth and death of a hero in a story to diplomacy and international treaties. it also introduces the concept of our Neuron Scientist, which autonomously develops and tests hypotheses about a neuron's computational role, running steering experiments, circuit analyses, and iterating on feedback until it generates a strong description. it's a hefty paper that reveals so much about the nature of neural networks. it's quite beautiful and i think you should read it.
In Neural Networks are Already Clean, @nickcammarata argued that neurons were far more legible than previously thought – and that this was great news for interpretability and safety. Now, we’re releasing the first technical companion to Nick’s essay. It shares our findings from studying thousands of neurons using our novel automated interpretability pipeline, and outlines why we now think of neurons by their computational function, instead of whether or not they are polysemantic. Here’s what we found: 🧵

Vmax is building Campaign, a platform for RL task generation using self-play. We worked with them to teach agents to use the Modal SDK, translating Python code into working Modal workloads by generating tasks that target their failure modes. Modal's elastic compute lets them run and check solutions, from GPU calculations to inference endpoints that reuse state across requests.
1/ As a teaser for Campaign, Vmax’s infrastructure for automated task generation, we are sharing work we did earlier this year for @modal using self-play to generate tasks that target weaknesses of models when using the Modal-SDK.

btw @harris check out this old post of mine. we share a lot of similar thinking! medium.com/work-bench/how-to…

"The most reliable way to measure real interest is to ask for something valuable that requires work after the meeting." Excellent advice from @harris for founders planning to raise
how to have a super awkward conversation with a vc where you might drink coffee and will hopefully get information and maybe even a term sheet or least build some fomo


A must read companion to @nickcammarata's research publication the other week Very cool to see that interpretability is far more tractable than previously thought...
In Neural Networks are Already Clean, @nickcammarata argued that neurons were far more legible than previously thought – and that this was great news for interpretability and safety. Now, we’re releasing the first technical companion to Nick’s essay. It shares our findings from studying thousands of neurons using our novel automated interpretability pipeline, and outlines why we now think of neurons by their computational function, instead of whether or not they are polysemantic. Here’s what we found: 🧵

This is a must listen @ProbyShandilya is a true student of the venture craft, and while the mental models he talks about vary across industries he always does an excellent job tying things together to drive relevant learnings
TLDR: I’m starting a new podcast - Thinking About Thinking! The origin story for this is a fun one. About a month ago, we at @Work_Bench hosted our NYC50 Operator Retreat, which was a blast. When we were brainstorming potential breakout sessions during the day, I gave a very “Proby” idea - a breakout session titled "Thinking about Thinking: The Mental Models that are Key in the Startup Journey.” Much to my surprise, this was one of the most attended sessions at the retreat! I had an awesome time co-leading the session with @fendien - at a moment when so much is changing in the surrounding startup landscape, it was fun, refreshing, and energizing to talk about various frameworks and mental models that have stood the test of time. This podcast is an attempt to continue and open source the conversation, so to speak. I’ll be discussing a new mental model every week, talking through things such as the origin of the idea, its nuances, where it’s applied across software/tech/business (and beyond as well), how people can use the model as a critical tool in building companies, how this framework can help us parse today’s reality, and so much more. The first episode is on incentives, inspired by one of my very favorite Charlie Munger lines: “show me the incentive and I’ll show you the outcome.” Enjoy!


We're running this back next Tuesday, 11/3! cc @Somil_Agg If you want a great poker game and even better conversation, shoot one of us a DM
Another great night of poker at the @Work_Bench office! Always a fun cohosting with @Somil_Agg. Fun bringing together great engineers from @DecagonAI, @OpenAI, @percepta, @SierraPlatform, and @DistSpectrum. Awesome convos on all things personal agents, SW factories, and more

I’m finding compute efficiency to be the most interesting metric from @AnthropicAI's S-1. Compute efficiency is how much revenue a company brings in for every $1 it spends on compute. Yes, Anthropic has obviously scaled phenomenally well, but as with all businesses, there’s a cost to build and deliver the product, and after reading the S-1, it’s clear frontier labs are truly built different… In its S-1, Anthropic’s single largest OpEx line item is ‘Compute and infrastructure’, which helps power everything from training new models to serving customers and more. In the first half of 2026 alone, Anthropic spent ~$9B on compute and infrastructure. That’s almost 4x YoY and already 22% more than all of 2025. Despite all that spending, compute efficiency has dramatically improved 5.2x in just five quarters from $0.40 of revenue for every $1 of compute in Q1 2025 to $2.08 in Q2 2026. I find compute efficiency so interesting because it’s a simple proxy to see how much wiggle room is opening up between what Anthropic earns and what it costs to produce AI. It's also a big reason for why Q2 2026 was Anthropic’s first quarter of operating profit of $974M. So, the real question becomes how long does this hold? We’re obviously in a compute-constrained environment, and a lot of Anthropic’s compute spend is locked into long-term contracts. On top of that, frontier labs are constantly racing to train and release the next model, which pushes them to procure more and more compute. At least in the short-term, neither long-term compute contracts nor the need to create SOTA models flexes with revenue, so if growth slows, how much will compute efficiency start falling? More importantly, will that newly positive operating income go with it? There are a lot of threads to pull on here and I’m still learning in real time. If you have any thoughts on this topic, I’d love to chat. Here are some bonus questions I’m thinking through: 1) How much further can revenue scale while compute is constrained? 2) Given models are some of the fastest depreciating assets, how will a lab’s compute strategy change over time? 3) If revenue slows because usage slows, how will variable compute change? How quickly can labs actually redirect capacity to training and other workloads? 4) What specific optimizations is Anthropic doing so well? Multichip architecture? Fungible pool of compute? 5) Most importantly, how does OpenAI compare on compute efficiency? (Though not apples to apples, since Anthropic books cloud revenue gross.)

"Indeed, arguments and disagreements are very present — but even this morning, people who will vote differently from one another at the ballot box will then get in the same tank and share a shabbat dinner. They are willing to give up their body and soul, one for the other, for something greater than themselves." 🙏🙏🙏
I was quite emotional this morning. I davened (prayed) at 6am. Then, I voted. In democratic elections. I then read about the $2+ Billion acquisition of 🇮🇱 quantum computing company Qlassic by intel. I was thrilled for our friends and colleagues who backed and built Qlassic. And, on a small personal level, I am excited by the new book I was privileged to write that is arriving in stores today. On the voting line at 7 a.m. I met a female soldier with an M16 slung on her shoulder. A yeshiva student. Elderly people and professors. Some with yarmulkes and others without. We live in neighborhood in this city where you see everything and everyone. I also thought about my children who are voting at special ballot boxes because they’re in the army or on the front lines now. I want you to stop and think for a minute. Almost any other country that would have gone through what we in Israel went through would not have made it to this moment. Massive protests, covid, a 7 front war, fractious politics and cacophonous social media. Indeed, arguments and disagreements are very present — but even this morning, people who will vote differently from one another at the ballot box will then get in the same tank and share a shabbat dinner. They are willing to give up their body and soul, one for the other, for something greater than themselves. Before I cast my vote this morning, I paused and thought about those who unfortunately and tragically didn’t return from the front lines and the hostages that did not return from captivity. Our biological family and our Israeli family. It is our first election since October 7th. My eyes teared up. I then thought about the heroism and self-sacrifice of young Israelis: the hundreds of days of service by our soldiers—our children—and their families stationed on the front lines and holding down the home front so that we may merit to live as free people in our land after two thousand years. This is no trifling matter. Our ancestors would have given arms and legs to have this gift. Many people around the world yearn for our freedom. We are privileged to be part of this incredible people and this amazing country at this extraordinary time. We are fortunate and blessed. I feel personally fortunate and blessed. Am Yisrael Chai. 🇮🇱