Kenneth Yesh

These 10 recent YC startups have some of the cleanest, most thoughtful product experiences I’ve seen 👇🏼 1. Parrot (YC F25) 🦜 Feels like TikTok but every swipe is structured for real language acquisition, which is why power users happily sit in it ~49 minutes a day. 2. Hyperspell (YC F25) ✨ UX is basically "connect Slack, Gmail, Notion, Drive and you are done", with a clean console that shows exactly what your agents remember. 3. Bear (YC F25) 🐻 Looks and feels like Google Search Console for AI search, with simple dashboards that show how ChatGPT, Perplexity and Google AI actually talk about your brand. 4. VibeFlow (YC S25) 🎛️ You describe your app in natural language then edit the backend as a visual workflow, so semi technical users can literally see and change app logic instead of guessing. 5. Bitrig (YC S25) 📱 The whole experience is chatting on your iPhone while a custom Swift interpreter shows your native app live, which users say feels magical compared to wrestling with Xcode. 6. Design Arena (YC S25) 🕹️ Turns model evaluation into a simple head to head game where real users pick winners, and the UI turns that voting into clear leaderboards devs actually understand. 7. Golpo (YC S25) 🎬 You drop in docs or workflows and get a polished whiteboard style explainer video without touching a timeline, which is why schools and enterprises keep using it. 8. Sim.ai (YC X25) 🧩 Figma style canvas where you drag blocks, connect tools and instantly deploy agents, with devs calling it the first agent builder that feels visual, fast and production grade. 9. Better Auth (YC X25) 🔐 Gives TypeScript devs a single, well documented auth framework so most flows are copy paste snippets instead of bespoke logic, which is why HN and OSS reviews rave about its DX. 10. Nomi (YC X25) 🎧 Joins your Zoom or Meet, pops real time phrase suggestions on screen and then auto writes notes and CRM updates, with reviews citing double digit lifts in close rate. I’m biased as a YC founder, but honestly, great UX is one of the strongest moats left. These teams are shipping products that feel less like "AI demos" and more like tools you want to live in every day.  - - - - - - - - - 👉 Who else should be on the "beautiful product" list? Drop your favorites in the comments so I can check them out.

147

14 Comments

Everyone assumes the most ambitious biology AI company in the world sits in Boston but its actually in London :)... Basecamp Research just closed an oversubscribed $140M Series C led by S32, the fund of Bill Maris, former CEO of Google Ventures. Anthropic's Anthology Fund, NVIDIA's NVentures, the NATO Innovation Fund, Singular and True Ventures joined the round. redalpine first invested in 2024 and we doubled down here. Most biological AI models are trained on the same narrow slice of organisms science has already catalogued, which is a rounding error of what is actually out there. Basecamp went and collected the rest - years of fieldwork across more than 200 locations in over 30 countries, from Colombian jungles to deep sea sediment to polar ice, turned into one of the largest genomic datasets in the world and into EDEN, a foundation model that learns how DNA works across the full diversity of life. EDEN is already in production. It runs inside Claude, where researchers use it to design antibiotics and prioritise vaccine targets in minutes. Next comes cell therapy, reprogramming a patient's own cells inside the body, with cancer and autoimmune disease as the first targets. The future of medicine will be designed, not discovered - good part of that design work is happening in Europe. Congratulations Glen Gowers, Oliver Vince and the whole Basecamp team. Proud to be backing you.

64

🔥 The Heat Death of Product Management (and we're doing it to ourselves) 🔥 There are two key constituencies for "AI PMs" using LLMs for everything: A) The people who know what to do and use LLMs as a force-multiplier to do everything better, while still keeping the real thinking, judgement and collaboration human. These people know what to do, and use "AI" to do it better. B) The people who DON'T know what to do and use LLMs as a quick emergency fix, don't check their half-assed work and wave it through with a smile before moving onto the next thing. They don't know what to do, and they're never going to learn because everything looks good when you don't know what good looks like. AI tool vendors will, probably in good faith, claim to be targeting group (A), but group (B) is undoubtedly bigger, and both groups look the same in ARR and MAU terms. The second order effect is that people in group (B) will either prove one of two things: (1) Product management is easy to automate, humans don't add any value, and therefore anyone can do it. Net result = product management is dead (2) Product management is not easy to automate, humans do add value, but the majority of people are doing it badly so leaders don't trust it anymore. Net result = product management is dead Product managers would be doing themselves, their companies and their customers a favour by relentlessly focusing on being in group (A). Who's with me?

80

19 Comments

Despite all the big funding rounds and flashy demos in US robotics, K-Scale’s inability to raise more money should worry us. We're at risk of replaying the LLM story all over again in robotics: - Chinese companies are going open-source and collaborating across the value chain (from EV suppliers to downstream integrators) - most western teams are going full-stack proprietary, closed-source, all in-house Guess which robots the next wave of research labs and startups will actually be able to build on when they want to invent new algorithms or tackle unseen real-world use cases?

273

33 Comments

Plot twist: LLMs can understand rocket science. They just can't code it without breaking everything. New paper tests LLMs on GTOC 12, where the world's best aerospace engineers compete on "nearly impossible" trajectory optimization problems. The findings are more nuanced than "AI good" or "AI bad." Here's what they found: The Good News: LLMs get the physics. They understand motherships, gravity assists, objective functions. One model scored 18/26 points, double the score from 18 months ago. They've conquered the conceptual barrier. The Bad News: They fail on execution. Not because they don't understand orbital mechanics, but because they: • Mix up km and meters • Make syntax errors   • Format outputs that crash validators • Fall into debugging death spirals (fix 1 bug, create 2 new ones) They act like a good architect who can design a perfect building but breaks the bricks when trying to build it. This isn't about "can LLMs reason?" They clearly can. The question is: can they execute with the precision engineering demands? Not yet. We're at a weird inflection point: AI that understands rocket science but can't reliably convert kilometers to meters. That's either hilarious or terrifying depending on your use case. The gap between conceptual understanding and flawless implementation might be the last major barrier to fully autonomous engineering. Full paper: https://lnkd.in/dhYKQ9yu #LLM #AI #Reasoning #Aerospace #Benchmarks

9

2 Comments

The AI startups that will actually endure will be: - Embedded into workflows - Trustworthy - Cross platform - Better than people - Better than other models - Propietary So many AI companies I know aren’t these things and it shows. The VCs who throw money at stuff without doing DD are gonna lose a lot and the founders who don’t account for these things are going to lose fast. It won’t be so much of a bubble pop as much as tire slowly deflating.

20

7 Comments

I’ve been looking at “platform-first” startup stories across software vs. hardware, and one pattern keeps jumping out: Software platforms can go horizontal from day one. Hardware platforms rarely do. In software, the decacorn playbook is almost familiar now: ship a reusable primitive (payments API, identity, observability, data platform), make it self-serve, iterate fast, let developers pull it into any industry. Think Stripe / Twilio / Cloudflare / Datadog / Snowflake / Databricks. In hardware, the incentives feel different. What I keep seeing in the last ~10–20 years Most “successful” hardware startups that look horizontal on paper still win via one of these paths: Vertical wedge first, platform later DJI (camera drones), Skydio (enterprise + public sector), many sensor companies… they had to pick a buyer + ROI early. Sell a complete system, not a component Carbon, Formlabs: the “platform” is real, but adoption happens when you package it into workflows customers can trust. Deep infrastructure with concentrated buyers Cerebras-style compute: massive technical moat, but distribution is “sell to a small set of high-need customers.” And the closest thing to truly horizontal, developer-driven hardware platforms (Arduino / Raspberry Pi vibes) tends to be community/ecosystem-led—often open, or at least ecosystem-native—because distribution + trust are the hard parts. Horizontal hardware as a startup: the real challenges Long iteration cycles + supply chain risk High upfront capex and “physics tax” Support burden (integrations, reliability, lifecycle) The “first customer” problem: who funds the platform before it’s a product? My takeaway (still forming) If you’re building enabling hardware tech, a vertical beachhead might not be a compromise—it might be the financing mechanism for the platform. Have you seen a hardware startup scale to $1B+ as a truly horizontal enabler without open-community pull or a vertical wedge? Drop examples—I’d love to learn.

20

Welcome Natan Linder, Co-Founder and CEO of Tulip Interfaces on this episode of 🔥The Lantern 🔥 Tulip Interfaces helps manufacturers build production systems that put people first with no code, low code, and other AI tools to turn operational knowledge into apps, workflows, and continuous improvement. 🔍 What happened in 2025 as the team heads into 2026? 🔎 Tulip wrapped their largest-ever “Operations Calling” conference with 700+ attendees to end 2025. Their platform is processing north of 620 million automations, 70 million application-based processes, and 35 million station hours. They're about to cross 30 billion contextualized data points! They also acquired the talented Akooda (acquired by Tulip) team, led by Yuval Gonczarowski, bringing enterprise AI search and deep learning expertise into their platform to help customers sift through massive operational datasets. In January, they announced a $120M fundraise led by Mitsubishi, that values Tulip at $1.3B. 🤔 Why does manufacturing need specialized AI? 🤔 Unlike software engineers who can freely experiment with their AI agents, real world operations require more rigor. Downtime and/or destroyed raw materials costs millions. Tulip is building AI copilots, composer products, and full agentic frameworks that are safe for the factory floor in regulated industries like pharma, medical devices, and aerospace. 🧠 What has Natan learned from building Tulip? 🧠 Work very close to customers. Natan believes a startup's focus should be split into 80% on customers, 10% on markets, and 10% for hallucinations. That customer obsession drove their Akooda (acquired by Tulip) acquisition. Operations leaders are knowledge workers too and need AI verticalized for their specific needs to democratize access to manufacturing expertise and elevate operations teams. 🤝 How can we help Tulip scale? 🤝 Tulip is in full sgrowth mode heading into 2026. If you're an engineer passionate about manufacturing, AI, or hard tech, or know world-class operators who want to transform how physical products get made, reach out! Boston is building the future of intelligent manufacturing and Tulip is leading the charge. The Lantern is brought to you by MGMT Boston, Rho, & Invest Northern Ireland – Americas this month

39

2 Comments

Anthropic's Project Vend fails, while Waymo is thriving. Why? And how is this related to #moltbook and #openclaw -- Prologue Waymo vehicles make "life-or-death" decisions every day, in open environments, among humans, cars, cyclists, and chaos, and they do it at scale. Anthropic's Project Vend, by contrast, put an AI in charge of a vending machine business: pricing snacks, managing inventory, handling suppliers. And it couldn't reliably manage the economics. That should feel surprising. But it isn't. Because the gap here isn't intelligence. It's systems engineering. Read again: IT IS SYSTEMS ENGINEERING. Albeit an AI system, it is still, a system. We are not close to an era where we can let AI build its own engineering, unless the environment is highly constrained: a simple CLI tool, or a social network for agents - maybe. -- Similarity At a high level, Project Vend and Waymo look similar: • Both operate in the real world • Both take autonomous actions • Both deal with uncertainty and edge cases • Both can cause real damage if they fail And importantly: • Neither is "just an LLM" So why does one work, and the other collapse? -- Differences Waymo is a complete, end-to-end AI system. Not a model with tools. A system. It has: • Clearly bounded domains and objectives • Redundant sensors and inputs • Continuous verification loops • Explicit safety cases • Tight feedback between action and reality • Rich observability for humans, with live controllability • Years of accumulated edge-case learning • Every decision is constrained, validated, and monitored. Project Vend, by design, exposed the opposite. • Even with a strong model: • Memory was shallow and brittle • Economic feedback loops were weak • Verification was mostly implicit • Errors compounded instead of being corrected • No real understanding of consequences over time At least based on what I could infer from the blog [1]. The AI wasn't irrational, it was underspecified, or more accurately, under-engineered. -- Conclusion: Waymo proves something important: Trustworthy autonomy isn't achieved by smarter models alone. It's achieved by engineering the environment around the model. Project Vend failed because autonomy without systems collapses under reality. This matters far beyond vending machines. If we want AI agents to manage money, operations, workflows, or organizations, we need to stop asking "is the model good enough?" And start asking: • Where is memory grounded? • Where is verification explicit? • Where do errors get caught? • Who can observe, intervene, and correct? • What happens when incentives drift? Waymo answers these questions systemically. Most agentic AI today does not. So please accept this warning: Be careful with #OpenClaw (ex #clawdbot / #moltbot). #moltbook working (so cool!) isn't evidence that you should let these agents operate for you, it works because it's a closed world where agents can only break their own toys, not your life.

227

23 Comments