About
Technology entrepreneur working with leading hotel brands of the world to create…
Activity
7K followers
-
Jasen Lew posted thisMy Customer Success Manager built a brilliant AI automation workflow. I told her to shut it down. When my CSM sent me the auth request, I paused: "Manage all business transactions." In Xero, I can create a user with edit-only access to draft invoices, no credit card permissions, and limited expense visibility. Dozens of granular controls. But when my team tried to automate invoice drafting through the API or an app script? The only option: "Manage all business transactions." Contacts. Payroll. Expenses. Everything. This wasn't necessarily a problem when only developers used APIs. They needed broad access and could manage it themselves. But now AI agents need to interact with our systems. And we're handing them permissions we'd never give a junior employee. And this binary access model is a security nightmare. You either give the agent keys to the kingdom or you block it entirely. Here's what most companies will do: Wait for software vendors to update their APIs with granular controls. Or, just grant full access despite the risk. Here's what we're doing instead: Computer use agents. They interact through the UI like humans do. Own login. Own permissions. No skeleton keys. Software companies spent years building robust permission systems. We just need AI that can use them.
-
Jasen Lew posted thisEveryone says TikTok is ruining attention spans. I thought so too... until I started running agent teams. This morning: - I kicked off a Claude Code agent team to research wellness protocols for a health app we're building. - While it ran, I spun up another agent team to process a community marketing meeting transcript, research a couple concepts and draft updates to a strategy doc. - While both agent teams worked, I responded to two Slack messages, responded to a designer email, and started drafting this post. All this within 10 minutes. Each agent run takes 2-15 minutes. The skill isn't sustained focus on one thing. It's rapid context switching between parallel threads - checking output quality, redirecting when something's off, holding the full picture while five things run at once. And I realized: I'm not doing deep work during my day. I'm doing something closer to air traffic control. Gen Z's TikTok-trained brains might actually be better wired for this than mine. Now here's the nuance... Deep focus matters more than ever. Maybe more than any time in history. You need it to set direction, define strategy, think through problems that agents can't solve. But execution with agents? That's a different cognitive muscle. Rapid switching. Pattern recognition across parallel threads. Knowing when to trust output and when to intervene. We've fallen into a 𝘀𝗶𝗻𝗴𝗹𝗲-𝗺𝗼𝗱𝗲 𝗮𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 𝘁𝗿𝗮𝗽... Optimizing entirely for deep, sustained focus while the agent era demands both depth and multiplexing. Deep thinking sets the vector. Rapid switching multiplies the force. We trained a generation to switch contexts in seconds and then told them it was a weakness. What if it's their edge? How are you managing attention and context (pun intended) when you're running multiple agents at once?
-
Jasen Lew posted thisWhen did we decide that easy to share matters more than hard to forget? I was staying at the Mandarin Oriental Hong Kong, running late for a meeting across town. Rushed through the lobby. Needed a cab. Reached for my wallet. No cash. I turned around to head back to my room. One of the Mandarin Oriental staff noticed. Asked what was wrong. I told them. They nodded, slipped to the front desk, and came back in 10 seconds with more than enough for a taxi there and back. No name. No room number. No questions. Just: "Here you go." That was years ago. I still remember it like it happened this morning. Not because it was flashy. Not because I posted about it (until today). Because of how it made me feel. That experience stayed with me. It's part of why I've spent the past decade building technology for luxury brands like Oberoi, Rocco Forte, and Mandarin Oriental. The moments that last aren't designed to be shared. They're designed to be felt. The same principle applies to building AI products today. Most founders optimize for metrics that spike and disappear - shares that don't convert, clicks that don't retain, attention that doesn't compound. But the products that last solve problems people can't stop thinking about. I've built multiple seven-figure SaaS businesses over 20 years. None of them went viral. All of them kept customers who never wanted to leave. The question isn't: Will this spread? The question is: Will this matter tomorrow? Optimize for memory, not momentum. Build things that are hard to forget.
-
Jasen Lew posted thisI told my engineers to stop getting better at coding. Because "good at coding" stopped being the point. For two decades, I saw great software engineering as three skills: Writing code, designing systems, and debugging. Then, AI agents entered our workflow. Our fastest engineer wasn't writing better code. He was talking (literally!) to his computer, providing better instructions (prompts), and describing precisely what needed to be done and directed context so that agents could work in parallel without stepping on each other. What used to take one engineer several days of sequential work shipped in an afternoon with zero merge conflicts. Our best debugger wasn't reading stack traces. He was catching where agent context had drifted, where an agent had "forgotten" a constraint from four turns ago and silently built the wrong thing. The pattern became obvious. After a couple of decades in software, I should have seen it sooner. Coding was always just the implementation layer. AI didn't change what software engineering is. It just burned away the part that was never the real job. What's left is what always mattered: Orchestration and context management. Orchestration is deciding what builds what, in what order, with what constraints. The value is speed. Context management is deciding what each agent knows, when, and how much. The value is quality and cost. I see this as the 𝗰𝗼𝗱𝗶𝗻𝗴 𝗶𝗱𝗲𝗻𝘁𝗶𝘁𝘆 𝘁𝗿𝗮𝗽. We spent decades defining engineers by the one skill that was most automatable. Now leaders are scrambling to figure out what their engineering teams should actually do... because they never clearly named the other two pillars. The engineers who thrive won't be the best coders. They'll be the best orchestrators. So yea... I told my engineers to stop getting better at coding. And they've never shipped faster.
-
Jasen Lew posted thisAfter burning through my team's monthly Cursor credits in 3 hours, the real lesson wasn't "spend less." It was to start tracking token mechanics. Most teams treat their AI coding tool as one thing. It's actually three layers, and the middle one is quietly destroying your budget. The 3-layer stack every team uses: 𝗠𝗼𝗱𝗲𝗹: Opus / Sonnet / Codex / Gemini 𝗥𝘂𝗻𝘁𝗶𝗺𝗲: orchestration loop (planning, tool calls, retrieval) 𝗖𝗹𝗶𝗲𝗻𝘁: IDE / CLI / chat UI Costs explode when runtime and client add hidden work you never see. The killer is feedback loops: Truncated context → re-explain Re-explain → more tokens More tokens → tighter caps Caps → more retries Your bill becomes a systems problem, not a people problem. A 30-minute audit any team can run: Pick one task (bugfix, refactor, feature). Run it two ways with the same engineer, same model: A) Direct runtime (Claude Code, Codex CLI, or API) B) IDE-layer runtime (Cursor agent mode) Track three metrics: 1. Tokens per task 2. Round trips per task 3. Re-explain events If B uses far more tokens than A without better output, you've found the tax. The top fixes: 1. Move shared context out of the prompt loop. Every major tool has a memory file for this: CLAUDE.md (Claude Code), AGENTS.md (Cursor), or equivalent config. 2. Reduce round trips. Plan → execute → verify in fewer loops. Batch related changes instead of one file at a time 3. Fix context truncation. Tools that cap below model limits cost you in re-explains. Larger effective windows often cost less total tokens by preventing retries The point isn't spend less on AI. It's to stop paying for redundant orchestration that isn't buying speed or quality. Where's your token waste hiding: model, runtime, or client?
-
Jasen Lew posted thisEveryone obsessing over YC interview tactics is optimizing for the wrong thing. I went through YC W18. I've also evaluated hundreds of startups as an investor. Last year a founder pitched me with a flawless deck. Beautiful slides. Rehearsed answers. I asked one question: "Why you?" He paused for 10 seconds. That pause told me everything. YC doesn't care about your pitch deck. They're looking for three things: Vision. Conviction. Team. Vision means you see a path to $1B+. Not theoretically. You've talked to 100+ customers, you live in your numbers and the problem, and you can articulate why this becomes massive, not just successful. Conviction means you've already started building and sacrificing. You quit the FAANG job. You dropped out of school. You have traction because you solved a problem people actually have. You're not testing ideas. You're proving a thesis. Team means the right people who share the vision and conviction. A resume is a signal and a starting point only. People with relevant lived experience win more often. They've worked in the industry, felt the problem firsthand. The ones who won't quit when it gets hard. YC partners spot this in minutes. That's why the interviews are short. The interview doesn't make or break you. Your work before the interview does. Stop optimizing your pitch. Start validating your vision, building conviction through action, and assembling the right team (which includes asking yourself honestly, "Am I the right person to work on this?"). Vision. Conviction. Team. Simple to understand. Hard to find. Unmistakable when spotted. What would you change about your startup if you focused only on these three things?
-
Jasen Lew posted thisI burned through my team's entire monthly Cursor credits in 3 hours. On a Sunday. On Monday, I posted about $600+ in Cursor overages in under two weeks, primarily from one engineer. 58 comments. Some said cap the spending. Some said switch the tool. Some said the metric was wrong. The most useful comments revolved around the same concept: stop looking at cost. Look at what's spending it. So I tested it myself. I joined our Cursor team plan. Used the same model (Opus 4.5) I'd been using all week in Claude Code. By 3pm, I used the entire month's allotment. Me. Same model. Wildly different consumption. That told me what I needed to know. Cursor layers its own prompting on top of whatever model you select. Larger system prompts, compressed context windows, more round trips. Its default context caps well below what Claude Code uses directly for the same model, and even higher tiers doesn't reliably close the gap. Every extra round trip burns tokens. Every truncated context forces a re-explanation. Seems like an "AI tool layer tax", i.e. the hidden cost of using AI through intermediary tools that silently multiply your token usage without improving your output. So instead of capping spending, we removed the inefficiency. We created a Claude Team plan for our engineering team. Every engineer now gets Claude Code access. Cursor stays available as an optional IDE for engineers who still want it, but we're setting hard caps on overages within Cursor. If someone hits Claude's throttle limits repeatedly on the the tier their on (twice a day, couple days in any given week), we upgrade them for the next couple of sprints. No approval workflows. Just one rule: if you're getting blocked, surface it. Engineers should be building. Building smartly with the right tools. The fix for overspending on AI tools wasn't spending less. It was removing the layer that was spending for us. How is your team deciding which AI tools to pay for directly vs. through an intermediary?
-
Jasen Lew posted thisAI didn't make your job easier. It made you responsible for three jobs at once. One of our engineers told me last week he's worried about his future. He's coding much less these days and isn't sure what his role even is anymore. The confusion makes sense. His job didn't disappear - it multiplied: Most people treat AI like a faster typist. They're still doing the work, just with a better tool. The real shift is different. Stop being the hands. Start being the brain and the coach. Your leverage isn't in typing anymore. It's in three things: • Designing the game (before execution) • Managing the players (during execution) • Improving the system (after execution) Most of the value comes from preparation and review, not the work itself. You're now wearing three hats at once: 1. Individual Contributor (unblocking when needed) 2. Product Manager (deciding what's worth building) 3. Orchestrator (setting standards and giving feedback) That's the role expansion that rarely is discussed. And here's the compounding part: A one-off fix teaches the AI how to solve one problem. A well-structured plan teaches it how to approach a whole class of problems. The difference? Encoding what you learn after each run. Every bug, edge case, and insight becomes a durable system upgrade for next time. Example: We kept hitting compaction issues in our code generation. Instead of fixing each one manually, we encoded two rules into our process: extract key details upfront, and break tasks into smaller chunks. Now our AI produces cleaner code and catches issues in reviews. One problem solved, entire class of problems prevented. Treat AI as a new team member who needs context, clear briefs, and feedback to grow. Not a faster typist.
-
Jasen Lew posted thisThe best AI analysis I've ever seen came from forcing two models to admit their mistakes to each other. I needed a comprehensive wireframe review for an agentic AI autoimmune wellness app. When you're building for people managing brain fog and fatigue, a confusing interface isn't just bad UX. It's a barrier to care. I ran the same analysis through Claude Opus 4.5 and ChatGPT 5.2 Pro independently. Then I did something different: I had each model review the other's analysis. THEN, I had each model review that comparison. The outcome was quite unexpected: Claude flagged "Sleep Hygiene tracking" as missing. It wasn't. The feature existed in the wireframes. ChatGPT caught it in the comparison. Claude acknowledged the error and corrected its analysis. This happened with multiple features. Attribution badges. Question status tracking. All flagged as missing. All actually present. The comparison step forced explicit accountability. Neither model's solo analysis was trustworthy. But when they had to defend their work against each other, quality emerged. The final document caught 8+ errors from the initial analyses, corrected 3 priority assessments, and prevented issues that would have sent designers rebuilding features that already existed. Here's what I learned: For any high-stakes analysis, run parallel reviews with different models. The divergences aren't bugs. They're features that surface blindspots. The comparison isn't extra work. It's where the actual work happens. This works for virtually any complex task: design reviews, proposal evaluation, document analysis. Run independent reviews with two models, have each compare the other's work, then have each review that comparison. The meta-analysis step is where errors surface. Then synthesize. The process is replicable.
Jasen Lew commented on a post
4mo
This one's well earned, Ishaan. You've consistently raised the bar on how we operate, and there's no one I'd rather have stepping up as we take Glowing into its next chapter. The best is ahead. 🚀
-
Jasen Lew reacted on thisJasen Lew reacted on thisFive Y Combinator startups are demonstrating how AI drug discovery is moving well beyond virtual screening and molecule generation. Ditto Bio (YC W26) is exploring parasite-derived proteins for autoimmune therapies, CellType (YC W26) is building biological world models that simulate human biology, 10x Science (YC W26) is automating next-generation protein characterization, Strand AI is developing multimodal foundation models that predict missing patient biology, and Atlas Discovery (YC S26) is modeling human drug responses to improve clinical-development decisions. As highlighted by Akiko Yoshida, MSc, these companies represent a broader shift toward applying AI across the drug-development workflow—from biological discovery and therapeutic characterization to patient selection and clinical-response prediction. https://lnkd.in/gaBdp4qW Want to stay ahead in biotech? Subscribe to our newsletter for more 🧬 https://lnkd.in/g3m5628R
-
Jasen Lew reacted on thisJasen Lew reacted on thisJuly 2026 was Clueso's biggest month yet. We added more ARR than in any previous month in our history. On top of this, we launched the most important product update we’ve ever shipped: 𝗖𝗹𝘂𝗲𝘀𝗼 𝗔𝗴𝗲𝗻𝘁𝘀. Engineering teams have spent the past few years becoming dramatically more productive with AI. But the teams responsible for explaining what they build (across marketing, sales, support, and training) have struggled to keep up. Clueso Agents changes all of that. Teams can now automate video production and keep their content up to date as their products evolve. Content teams can finally move at the speed of engineering. The response has far exceeded anything we expected, and our team has been working late nights and weekends to support the demand. Customers are truly seeing the value, and some are expanding their contracts by 3x or more to adopt Clueso Agents. But the most rewarding part has been hearing what customers are saying themselves: • “I was updating some old videos when this email arrived, and I thought, ‘Oh my God, finally!’” • “We were able to build a customer education academy with around 70 videos in just two weeks.” • “Everything you’ve shown has my mind going in a thousand happy directions because of the efficiency gains.” • “The platform has really supercharged us. It’s allowed us to move so fast.” I’m incredibly proud of what this team accomplished, and even more excited about what comes next. Stay tuned for August.
-
Jasen Lew liked thisJasen Lew liked thisHTML plans vs Markdown plans. Realised HTML plans are great for iterating on complex workflow planning. --- I was working on a complex workflow that combined human and agent-executed steps. There's loads to figure out. Skills, automations, queue objects, review points, and how everything should connect. I started by planning everything in markdown. It was useful for capturing the details, but once the plan got longer, it became hard to inspect the workflow as a whole and see if everything connects well. So I tried a different approach: I asked Codex to turn the workflow into an HTML visualization and open it in the in-app browser. It worked surprisingly well. I could see the full workflow, annotate the exact sections that were unclear, and ask Codex to update them. After a few rounds, I discovered a workflow that works well for complex systems: 1. Research and capture the details 2. Visualize the overall workflow in HTML 3. Open it in the in-app browser 4. Annotate the specific parts that need changes 5. Iterate until the system is clear 6. Only then create the spec and implementation plan The useful part isn't that HTML looks nicer than markdown. It gives me and Codex a shared visual artifact to inspect and improve together. And yes... I am using Codex too.
-
Jasen Lew reacted on thisJasen Lew reacted on this😇 Took a week off social media and AI to go to Iceland on a camper van trip and celebrate a year of starting my company. Probably the best week I’ve had in a bit. Building LevelUp Labs from the ground up has been one of the most meaningful journeys of my life. But I also think it sometimes takes away one very underrated joy of being an employee, which is the ability to switch off, forget about the outside world for a bit, and just exist outside your work. For an entire week, my husband and I drove through Iceland living in a camper van. Every day was basically waking up, driving, stopping at some ridiculously beautiful place, figuring out where to park for the night, cooking, sleeping, and doing it all over again. And somehow just that grind (and not having the time to think about anything else) gave me more energy to come back and continue work than any productivity system ever has. I don’t want to turn this into a “7 lessons from living in a camper van” post 🫠 but, the biggest thing I came back with was this: don’t take yourself so seriously. The world is beautiful and strange and wonderful, even when you’re not paying attention to it. Nothing massive happened that I absolutely had to witness. Nothing I did was so important that the world couldn’t continue without me. And just that thought it incredibly freeing! It made me feel like I can keep building, keep creating, keep sharing the things I care about, and keep playing the game without making myself the center of the universe. I came back with a lot more energy, but also with a little more distance from my own seriousness!
-
Jasen Lew liked thisJasen Lew liked thisDario Amodei surprised me today in going against the grain of AI drug discovery hype. At today's Anthropic event, while others have claimed AI will "cure all disease," Dario was notably measured: he said that while in 10 years we might see AI compressing discovery 10-fold, the technology is still just beginning to take off. My key observations: - Pharma leadership is serious about compressing timelines across every stage of the lifecycle, across discovery, manufacturing, clinical trial recruitment, and internal ops. I counted CEOs and other executives from at least 10 major pharma companies in the room. - Early discovery is already seeing real gains. Aviv Regev (Genentech) cited 130% more programs added to their early pipeline from internal AI transformation. - Most large orgs are targeting 5-10% productivity gains overall. That sounds modest, but actually would lead to billions in efficiency for large-cap pharma if these efforts pay off. - Getting large organizations to adopt AI can be challenging because of both the speed of technological development, and the need to fundamentally reimagine legacy processes. This last point is exactly why I think startups like ours at Ditto Bio (YC W26) have a competitive advantage. We're building novel therapies for autoimmune disease by harnessing the biology of parasites – organisms that have spent millions of years evolving precise ways to modulate the immune system. We use AI/ML to mine that biology at a scale and speed that wouldn't have been possible even a few years ago. Building from scratch also means we don't have to refactor old processes; we can plan for the future from day 0. Today's event included the launch of Anthropic's newest product, Claude Science (essentially Claude Code/Cowork built for life sciences). We've been trialing it in private beta ahead of today's launch and have been impressed by its capabilities. Time will tell if Dario's predictions come true – perhaps we'll even be surprised by how much faster we can build with these tools. Emily Weiss, PhD Adair Borges, PhD
-
Jasen Lew liked thisJasen Lew liked this😲It's an honour to have been recognised as a HotTopics Global CIO:100 for 2026. It's not often I'm lost for words... I don't quite know what to say... I'll be right with you...
-
Jasen Lew liked thisJasen Lew liked thisI'm looking forward to seeing hospitality industry friends at the upcoming HTNG Tech Summit in Redwood City, California. Lots for us to discuss!
-
Jasen Lew liked thisJasen Lew liked thisHeading to sunny San Diego ☀️ June 22-25 for BIO International! I'll be representing Ditto Bio (YC W26), where we're developing building a new class of autoimmune therapies inspired by parasite biology. Our AI-native platform mines parasite genomes for proteins that retrain the immune system. We engineer these molecules into therapies that restore immune tolerance without broad immunosuppression. If you're around, let me know! I'd love to catch up. #BIO2026
Experience & Education
Licenses & Certifications
Languages
-
English
Native or bilingual proficiency
-
Chinese (Mandarin)
Professional working proficiency
-
Chinese (Cantonese)
Elementary proficiency
Recommendations received
8 people have recommended Jasen
Join now to viewOther similar profiles
Explore more posts
-
Anjli Jain
**Uber's Strategic Moves in Autonomous Vehicles: The Waabi Partnership** Uber's partnership with self-driving startup Waabi underscores the importance of strategic collaborations in advancing the autonomous vehicle (AV) space. This billion-dollar investment signals not just a focus on trucking but a broader commitment to revolutionizing transportation. At ElevenX Capital, we understand that strategic positioning in this rapidly evolving sector can yield substantial returns. As Limited Partners, how does your investment strategy align with these transformative trends? #investing #innovation #venturecapital #entrepreneurship
1 Comment -
Dave Lambert
We invest at Pre-VC and Pre-Seed, and then help our founders find a Seed round when they're ready. Here are 27 Seed investors that we *know* are active right now. Blue Moon Bonfire Ventures Chingona Ventures Flying Fish Partners Foothill Ventures Great North Ventures ground game HealthX Ventures InsurTech Venture Partners KCRise Fund Las Olas Venture Capital Leva Capital Maven Ventures Monte Carlo Capital Moonshots Capital OneSixOne Ventures Oval Park Capital Pixel Perfect Ventures Range Ventures Revolution SNAK Venture Partners Startup Capital Ventures Supernode Global Susa Ventures Trilogy Equity Partners True Ventures True Wealth Ventures
15 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content