Yunkai Zhou

The uncomfortable truth is that most GPU waste comes from treating models like services instead of ephemeral compute. Inference traffic is spiky, stateful, and multi model. Static allocation guarantees idle silicon. GPUs don’t need better schedulers. They need a different runtime model. And We’ve been building that at InferX Great article by Anyscale

10

Martin Eberhard played a key role early at Tesla and even brought in Elon Musk. Then he was pushed out of the company for pennies. Founders get most of the attention in Silicon Valley, but the ecosystem is built on the people who take a chance and join early stage ventures. Unfortunately, those same people are often the easiest to cut out. Founders can dilute early employees, change vesting terms, and restructure equity with little effort and zero consequences (you can literally do it in Carta in 5 minutes). Daniel Kottke was a close friend of Steve Jobs and helped build the Apple I and II computers. He got no equity (Steve Wozniak eventually gave him a few of his own shares). Brian Acton co-founded WhatsApp and stayed through the Facebook acquisition. When he left Facebook over privacy concerns, he had to walk away from $850 million in unvested stock. There’s more awareness now around how vesting works, what happens during a liquidity event, and how to think about re-ups and dilution. That’s good. But the truth is, if someone really wants to screw you, they can. It still comes down to one thing: trust the people you work with. Or don’t work with them at all.

85

8 Comments

Sequoia just published their thesis on the next trillion-dollar company. It won't sell software. It will sell the work. Julien Bek, Partner at Sequoia, mapped out 10+ verticals where AI-native services firms will replace both software vendors AND incumbents. His core insight: for every $1 spent on software, $6 is spent on services. AI is turning that $6 into software. If you sell the tool, you're in a race against the next model. If you sell the work, every improvement in the model makes your service faster, cheaper, and harder to compete with. A year ago, this was a niche take. Today, it's consensus: → Sequoia: "Services: The New Software" — full framework for AI-native services firms → YC: Spring 2026 RFS calls for full-stack AI companies and AI-native agencies → a16z: backing AI-native companies that own the outcome, not just the tool We've been living this thesis before it was cool. At ClaimMind, we don't just sell AI tools for claim processing. We sell the work — case mix optimization as a service. Hospitals and insurers get outcomes (faster claims, fewer rejections, better groupings), not dashboards they have to figure out. At KeyCenter, we don't sell consulting hours. We sell business advisory as an outcome — AI-native analysis that used to take weeks, now delivered in hours. The model gets better, our service gets faster. The pattern is clear: Lawhive in legal, Rillet in accounting, WithCoverage in insurance. Now AI-native services in healthcare claim processing and business advisory. When Sequoia, YC, and a16z all converge on the same thesis — pay attention. This isn't a trend. It's a structural shift. If you're in healthcare and want to see what AI-native claim processing looks like — let's talk.

6

1 Comment

Demis was talking about RL environments and dropped this gem: "Humans don't have single objective functions. We're multi-objective and continuously weighting those objectives differently." This hits different when you're building voice AI. It’s clear, you can't grade voice AI on just "did it answer correctly?" Real evaluation needs multiple dimensions: → Follow instructions? → High-quality response? → Didn't piss off the customer? → Hit the business goal? Single metrics = single-minded AI = broken in production. Real-world environments are messy. Every customer is different. Emotions change mid-call. Context shifts constantly. At Leaping AI (YC W25), we grade every call across multi-dimensional evals simultaneously. Because you can't optimize for clean when reality is chaos.

17

5 Comments

In partnership with Anthropic and Perplexity, I brought together Stanford engineers to go from idea → working product in a single day. It was our first ever Stanford Accel Leadership Program hackathon. 🔍 Perplexity’s Kesku unveiled their newest Search API, giving builders the flexibility to choose any model for their LLM stack. 💻 Anthropic’s Thariq Shihipar and Kashyap M. ran a live 5-minute quick start on Claude Code that allowed even non-coders to ship fast. 7 teams. 12 hours. Zero excuses. Winning projects: 🥇 Voicera (AI fraud detection) 🥈 Daily Brief (custom daily podcasts) 🥉 Trail (local personal memory) The gap between idea and product has never been smaller. The builders who move fastest and choose wisely, win. Blog in comments Loubaba El Ayoubi Fahad Nabi Alex Georgiou Naima Patel Roya Meykadeh Angela Tsao Abbie Maemoto Jisha Kambo Advit Deepak Anshika Agarwal Katie Li Jenny S. Jesse Dwyer Sara Ittelson Mo Fong

268

22 Comments

If 2025 was the year of the AI agent, 2026 will be the year of AI managers. Actual AI is pioneering a new set of AI agents and tools that help engineering managers to manage the new reality of engineering teams being composed of humans and AI agents that work 24/7 to build the next generation of companies. I'm excited to partner with John Kennedy, Ethan Byrd who's building Actual AI with stellar team from AWS and GCP with support from Jean Sini, at Irregular Expressions, Vik Ghai at G2C Venture Partners.

36

1 Comment

Reposting for visibility. Don't know details but if Jon Maroney is helping out, worth a look if you have these skills

5

I had the honor of learning about fellow Knight-Hennessy Scholar Yuncong Toby Mao’s research work that leverages AI and machine learning to advance medical technology. Please follow Toby and DoubleCheck AI to watch our interview!

34

4 Comments

We are announcing Potpie AI's $2.2M pre-seed fundraise to advance Spec-Driven Development for large enterprise codebases. The round is led by Emergent Ventures, with participation from All In Capital, DeVC, and PointOne Capital, along with the support of some amazing angel investors from companies including Atlassian, OpenAI, Meta, Razorpay and Flobiz. As AI accelerates code generation, the constraint inside large enterprises has shifted from coding to maintenance and assurance. The limiting factor is no longer writing code, but understanding complex systems, aligning teams around intent, and safely evolving large, interdependent codebases. In most organizations, specifications exist as static documents, while production systems evolve independently. Context is fragmented across repositories, tickets, logs, reviews, and floating documents making reliable AI adoption difficult. We are building the foundational layer that makes Spec-Driven Development executable at scale. By unifying engineering context and operationalizing the spec as a structured source of truth, we enable AI systems to reason with architectural awareness rather than surface-level code completion. We are already working with large enterprise customers, including Fortune 500 organizations. This milestone allows us to deepen those partnerships and support more teams transitioning from experimental AI usage to structured, production-grade AI-first engineering workflows. If you are leading engineering at scale and evaluating how AI should integrate into mission-critical systems, we would love to chat with you!

808

132 Comments

When AI models generate a pipeline as a single code block with provider agnostic vocabulary, chainable collections, typed entities, and namespaces, it feels like a natural fit. Instead of juggling raw JSON and multiple orchestration steps, the workflow becomes cleaner, more reliable, and easier to reason about. Full article: https://lnkd.in/guZCwsg4

7

Just wrapped up a few days at #RaySummit2025 organized by Anyscale. Last year, it was all about RAG. Now, the conversation around AI has evolved beyond text-based foundation models. Here are my takeaways: 1. Post-training is where the real breakthroughs are happening. Over the last year, attention has shifted to reinforcement learning after deployment. Frameworks like Verl and GRPO are pushing the limits of model refinement through long-horizon, user-grounded feedback. 2. Synthetic data is the new scaling lever. As natural data sources plateau, teams are turning to simulation and synthetic generation to stress-test models and expand coverage, effectively creating compute for cognition. This is an ode to all the work we did in autonomous vehicle simulation back in the day… fun to see this theme become more larger relevant. 3. The “cognitive supply chain” is the emerging bottleneck. As Jimmy Ba (xAI) put it, “We’re hitting a blocker on the cognitive supply chain.” Enterprises trying to reuse legacy infrastructure for AI workloads are discovering that this transition demands architectural, not incremental, change. 4. Multimodal and physical intelligence are converging. Jim Fan (NVIDIA) described the 'Physical Turing Test' .. can you tell if a robot interacted with the physical world? Progress here depends on bridging reasoning and embodiment, not just text and pixels. 5. Foundation models are fragmenting into specialized layers. From User Foundation Model built by Grab to OpenPI from Physical Intelligence for robotics, we’re seeing verticalized models emerge — tuned for specific data modalities, user behaviors, and physical environments. I’m super curious to see what this next year will bring. What’s clear: opportunities abound in marrying human expertise with post-trained foundation models tackling problems in verticalized AI domains. Building or hiring AI talent? Would love to hear your thoughts!

29

4 Comments

I make my employees suffer together. It's the best management decision I've ever made. NVIDIA 's worst cultural habit made them a $3T company, and I accidentally copied it. Jensen Huang says the most important thing at NVIDIA is that everyone “suffers” together. He uses that exact word. After building two companies, I think he's underselling it. Building a company takes an emotional and physical toll. There's no way around it. You're going to work long hours. There's going to be a lot of uncertainty. You're going to work really hard on something, feel accomplished and great about it, and then a day or two later it becomes irrelevant because the market shifted. You're going to have competitors come in and challenge you in ways you weren't expecting. You're going to get negative feedback from customers, from investors, from people you respect. All of the above is part of suffering through it. The question isn't how to avoid suffering, but how to make sure your team suffers together instead of alone. At Windmill, we share challenges openly. When something breaks, everyone knows. When we're behind, everyone knows. I don't pretend to have all the answers because pretending would be the fastest way to lose people who trust you. We read startup books together as a team and talk about where we're struggling with the same things the authors describe. Struggling together through problems is what builds the toughness to survive. But not all suffering is worth it. Employees suffering through performance reviews is just a waste of time. I have watched managers spend days writing reviews from memory. I've watched employees dread a process that was supposed to help them grow. That kind of suffering doesn't make your team stronger. It just makes them want to leave. We built Windmill because we believe in suffering together through the hard stuff, like market shifts, competition, and uncertainty. But we also believe in eliminating the suffering that doesn't need to exist, like traditional performance reviews. Shared suffering builds companies. Pointless suffering kills them.

33

3 Comments

This is the real filter: if nobody is willing to put their name on it, it’s not automation — it’s noise. The AI layer has to be approval-native: policy checks, audit trail, and explicit handoffs to accountable owners. Otherwise “smart” outputs never become decisions.

1

Recently, a friend told me he was "long on compute." My first reaction — obviously, isn't everyone long on compute? Then I started actually using Claude more seriously. I don't think the market has priced in what personal agents do to compute demand. People are already buying dedicated Mac Minis just to keep their agents running 24/7. That's not a niche hobbyist thing. That's a signal. We're used to thinking about the compute build-out in terms of hyperscalers and GPU clusters. But the real story might be at the individual level — billions of people running personal agents in the background, continuously. Every individual consuming data center-scale compute. Not shared, not enterprise. Personal. We may be a lot earlier than we think.

24

10 Comments

YellowDog.ai just set a 10x benchmark uplift in scale computing, delivering 40,000 tasks per second (TPS) and managing 100,000 compute nodes in the cloud. What's even more interesting, that's 2x IBM Symphony and opens an intriguing pathway for these until-know closed/captive systems.

11

It was a pleasure to chat with Rakesh Agrawal over at PCMag on his recent article that dug into how Waymo is changing late-night rides, especially for women. A few key takeaways: 🛡️ Many women prefer Waymo at night because there’s no driver involved, which reduces the risk of uncomfortable or unsafe interactions. 💸 That added sense of safety often comes at a premium. In one example, Uber and Lyft rides were under $20, while Waymo was over $50 for the same trip. 🚗 Waymo has a fixed fleet. Unlike Uber and Lyft, it can’t add more cars when demand spikes, so prices rise or rides disappear. 🎀 The result is an indirect “pink tax.” Men are more likely to choose cheaper human-driven rides, while women are more likely to pay extra for safety. ⚖️ Ride-hailing platforms are trying to close the gap with features like women-preferred driver matching, but smaller pools can also mean higher prices. ❓ The big question: should safety come at a premium, or should companies like Waymo rethink pricing so safer options aren’t effectively a luxury?

10

5 Comments

Happy Monday :) SF AI-Native Operator Takeaway #1: Forward-deployed engineers and how AI companies are ~actually~ getting built. One of the biggest differences between AI-native companies and traditional SaaS right now isn’t just the technology or the model. It’s how and where the product actually gets built. The strongest AI teams I met aren’t optimizing pitch decks or even demo environments, they’re building inside customer environments. That’s why forward-deployed engineers and implementation strategists keep coming up. In practice this means: engineers sitting directly with customers; shipping integrations, workflows, and edge cases in real time; product managers working with those FDEs to understand what needs to be done; and learning what actually matters to the customer before anything gets productized. This flips the old SaaS playbook. Traditional SaaS assumed engineering was scarce. You build once, sell many times, and customize through Sales and CS. When bespoke engineering was required, it was often rational to walk away. AI-native teams are operating under a different assumption: engineering and GTM are treated as equally flexible. With that, services are often the wedge to becoming a platform, not something to run from. The vision is still to be a platform, but the momentum starts with high tough delivery. That said, there’s an important caveat that came up repeatedly. The forward-deployed model only really works when contract values can support it, and yet right now, it feels like almost everyone is trying to use it. Forward-deployed work drives speed, but it also blurs: ➖ Product versus services ➖ Pricing models, such as software plus implementation or usage plus services ➖ Gross margins once delivery, support, and compute costs normalize The best teams aren’t blind to this, but they still use the approach to rapidly build product. They focus less on feature validation and more on customer willingness to pay for something (before building it!), as well as what multiple customers repeatedly ask for versus what is truly bespoke. With that, they decide what should stay services versus what makes it into the core product. In this rapidly changing time, speed beats elegance, but only if teams are honest about what they’re actually selling and what the economics can support. Forward-deployed teams aren’t a scaling strategy on their own, but they’re an incredibly compelling learning strategy. And right now, learning (and implementing) faster than everyone else is the advantage. Next up: how PLG and GTM are changing in the AI landscape, and why many teams are fixing the wrong problem. #FDE #AI

14

2 Comments

🚀 At Instacart, we just launched PIXEL — our in-house generative AI platform for product imagery. From DreamBooth fine-tuning to automated quality review with VLMs, PIXEL helps teams generate high-quality food images 10x faster — boosting conversions and reducing manual effort. Proud of the cross-functional AI work here! #GenerativeAI #ComputerVision #Instacart #ML

28

https://lnkd.in/g6JBTDux “Steam, Steel, and Infinite Minds.” Ivan Zhao’s latest piece stopped me in my tracks. It made one thing painfully clear: we are entering the first era where human progress is no longer constrained by the scale of human cognition. Just as steam and steel propelled us from agrarian labor into industrial civilization, AI is introducing a new kind of material into our world: Infinite Minds. This isn’t a faster tool. It’s not smarter software. It’s a fundamentally new production factor, one that can multiply human thinking exponentially. A few realities we can no longer ignore: 1. AI is not an efficiency tool. It’s a cognitive amplifier. When thinking itself can be replicated, parallelized, and run continuously, team size no longer defines intellectual capacity. The real advantage isn’t having AI, it’s knowing how to collaborate with infinite cognition. 2. The office was never the final form of work. Human labor moved from factories to desks. AI may now dismantle the idea that humans must be the sole cognitive center of work. Knowledge workers won’t disappear, but roles built purely on individual brainpower will be fundamentally reshaped. 3. Tools are becoming the factories of the future. As Ivan points out through Notion, we are no longer building “software.” We are constructing cognitive infrastructure for the next civilization. Ignoring AI isn’t cautious, it’s choosing to stand against the direction of history. At Genomii, this question confronts us every day: When AI stops being “advice-giving software” and becomes something that can continuously understand, reason, execute, and stay with humans through long-term goals, where does true human value belong? I’m increasingly convinced of this: The most important skill of the future isn’t knowing how to use AI, but knowing how to form long-term collaboration with it. If steel built empires, what kind of civilization will infinite minds create? This isn’t science fiction. It’s already unfolding. And I choose to see it as an opportunity, an opportunity to redefine the human role itself. What do you think? #AI #InfiniteMinds #FutureOfWork #Genomii #Leadership #AGI #notion

3