The future of the internet will be dominated by tool calls, not clicks. We're building Smithery to orchestrate this new era of AI-native services for AI agents. Read our mission 👇

4 years of building. 3 years of no traction. 2 years of moving to Malaysia to reduce burn. 1 year of insane growth. Now at $200K MRR!

Pitched her in 2018 Finally closed the round in 2025 💍

Smithery just launched Toolbox 🧰, a single MCP that dynamically routes to all 4000+ MCPs on our registry based on your agent's needs! Your agent can now use multiple @SmitheryDotAI MCPs without knowing which ones are needed in advance: 👇

Smithery instantly adds thousands of hosted MCP servers via a single link. Use @ExaAILabs and @browserbase in @OpenAI's new responses API playground.

After 6 years, it's a bittersweet moment for me to share that I've exited from Jenni! ​ Building my startup from $0 to $6M ARR has been the craziest adventure of my life. I've learned so much and made many friends on the way. I'm proud of what our team has accomplished. ​ My co-founder @Davidjpark96 will take the baton and I'm confident he'll push Jenni to the next level. 🫡 Super grateful to our investors ( @jason, @antler, @AIGrant) who believed in us and the family & friends who supported me in this journey! ​ I'm spending the coming months reflecting, tinkering, and talking to people to figure out the most important thing to build as we usher in a new era of artificial intelligence this decade. I'm excited to be joining the @southpkcommons community during my exploration phase. ​ Open to meeting & catching up with AI builders and startup founders!

Taken to the limit, AI coding is perhaps more akin to an advanced compiler. Compilers take high level code and turn it into low level instructions. AI compilers do the same but take natural language or human goals as input.

Excited to share that Smithery ( @SmitheryDotAI ) is backed by @southpkcommons! I started Smithery in Dec ‘24 to grow the community of developers building MCP servers. Today, we’re the #1 MCP platform with 4K+ servers listed & hosted. We're just getting started. Context is the biggest bottleneck in agentic AI, and there's a ton of infrastructure work ahead to enable your agents to see and interact with the world. We'll usher in the promise of agents into reality. 🚀 Honored to be supported by @rsanghvi @adityaag @finn_meeks @gopalkraman!

The story of how one Chipotle burrito 🌯 led me to co-found a $6M+ ARR AI startup: ​ I started my freshman year and met @Davidjpark96 because we were both looking for roommates. We met up at a Chipotle near campus. ​ We made our order (no guac), sat down, and talked about our interests. David told me about the T-shirt brand he built during high school. I shared about the computer games and Minecraft mods I built. ​ As we ate our burritos, an idea struck us: wouldn't it be cool if I built an app, and David marketed it? ​ Imagine how many $10 Chipotle burritos 🌯 we could earn! ​ So we started building. ​ 2015: Social sharing app 👬 We spent our first year in college building a digital business card that shared your contacts with a single swipe. Like most naive founders, we didn't talk to users and built in stealth. We made an app nobody wanted and failed after launch. ​ We made 0 Chipotles 🌯 ​ 2016: Dating app ❤️ Our second product was a dating app that leveraged AI. At that time, I was taking machine learning classes and became interested in neural networks. ​ We thought: let's optimize finding long-term relationships using AI. ​ This time we finally got some users! But we just couldn’t figure out a monetization strategy. A profitable dating app is one that needs high retention and shouldn’t actually solve long-term relationships. ​ We still made 0 Chipotles 🌯 ​ 2017: Crypto trading 📈 Bitcoin was becoming mainstream, and we had to jump into this hype. ​ I thought we could leverage AI to trade crypto and profit. David raised a $10K angel check for this idea by pitching to a stranger over a 3-hour flight. He framed the check on the wall in our living room. ​ We raised 1000 Chipotles 🌯 ​ I bought NVIDIA Titan V GPUs and had it running in my closet while trying to train a model that could trade cryptos, but I couldn't get the trading algorithm to profit. ​ We lost 500 Chipotles 🌯 ​ 2018: AI Platform 🤖 I joined the PhD program at UCSD and researched music generation and language models with neural networks. ​ We thought: what if we could create an AI platform to let people upload their data, and we'll automatically train a deep learning model for them? AutoML was all the rage back then. ​ We spoke to some startups and even got an LOI. But those startups shut down a few months later and we lost all our customers. ​ We made 0 Chipotles 🌯 ​ 2019: Jenni AI 🖊️ When GPT-2 was released, we were excited about shipping something with language models. I dropped out of my PhD program and we both committed to our startup full-time. We started from our college basement with ~$5K starting capital. ​ Jenni started as a writing assistant for SEO content writing. We wanted to help SEO content marketers produce content faster using GPT-2. ​ To learn how SEO worked, we initially ran Jenni as a content mill and dogfooded our product. After a year, we scrapped the agency and launched our product as a SaaS. ​ We hovered at 100 Chipotles/mo 🌯 ​ By then, COVID was at its peak, and marketing agencies were cutting back their spend, so it was hard to acquire new customers to fund ourselves. We applied to YC21 and got rejected. To extend our runway, we booked a one-way ticket to Malaysia to reduce our burn. ​ We built many iterations of Jenni but struggled with product market fit. One of our big SEO customers emailed us saying they wanted to churn from Jenni and use our competitor instead. We were losing in both distribution and product quality. ​ At that point, we were close to throwing down the towel. But there was one thing we noticed during these months that kept us going. ​ There was a group of users using Jenni actively despite us not marketing to them: college students 🎓. After interviewing these students, we discovered that they had pain points that nobody catered to, so we decided to pivot from SEO to academic writing. ​ With this pivot in mind, I aggressively pruned features in our product and optimized a single killer feature that delighted students. David experimented with many distribution strategies and cracked a formula for going viral on TikTok. ​ One morning, I was woken up at 5 AM by my co-founder calling me on the phone and telling me Jenni had gone down. Turns out we hit our OpenAI API limit because there were too many students signing up. ​ After 2 years of struggle, we finally had our first sign of PMF. ​ The Chipotles 🌯 were flowing in. ​ From that moment, we focused on scaling and strengthening our PMF. We continued to optimize the product, add features, and tune Jenni to solve more problems within an academic's research workflow. ​ Over the last 2 years, Jenni has grown 15% month-over-month to 250K monthly active users. ​ Now we make 60,000 Chipotles/mo 🌯 ​ With guac 😎

Next Step: Train a model that will predict when you tab, to do the tabbing for you

I'm incredibly excited to be joining forces with @kamathematic as co-founders to build the new agentic internet, one where tool calls are the new clicks!

I'm hiring founding engineers to join Smithery to build the AI gateway of the future. We're currently a team of 2 and the leading MCP platform. Our usage grew by 20x over the last few months. If you're obsessed with connecting agents to everything, let's chat!👇

Watched @karpathy's talk on the State of GPT. The most interesting slide was how LLMs learn a spectrum of performances, and you need to ask LLMs for good performance. Feels similar to Upside down RL (arxiv.org/abs/1912.02875) where you sample trajectories from desired rewards.

Thanks for backing us @jxnlco!

Elo rating is better for so many other things: - hiring candidates - sprint task point estimates Things are more often relative rather than absolute.

Proud to have Smithery integrating with @raycastapp to bring thousands of MCPs to your Mac! You can now install Smithery MCPs into Raycast from our website or from Raycast's spotlight search. @thomaspaulmann

The missing key here is the non-linearity (e.g., ReLU). Without non-linearity deep learning is just linear regression. With just a tiny bit of non-linearity, you can approximate many things.

I'll be hosting the MCP Track tomorrow at @aiDotEngineer @swyx There's an awesome lineup of speakers - come learn about MCPs!

Thanks to @MenloVentures and @neondatabase for having me at MCP party tonight to talk about Smithery and the future of agents! PC @CCgong

We've just been accepted to @vercel's AI accelerator! At @whoisjenniai, we've been using Vercel since day one (~4 years ago) to ship our Next.js front-end back when it was called Zeit. Great to see how fast Vercel has grown and how it's supporting the AI builder community.

We won 2nd place at @OpenAI x @GovTechSG Singapore hackathon! @gabrielchua @ivanleomk @jetnew_sg @LimYiFan @markchen90

1. Making the features as simple as possible with good UX. Focused quality is better than quantity. 2. Serving a market that's unpopular

This shows us there are a lot of opportunities to potentially self-train from the model's own data. For example, GPT could self-train to be more consistent by sampling A -> B, B -> A, and asking the model to resolve the conflict, then using that as new training data.

Smithery is partnered with @ExaAILabs! For a limited time, your agent will be able to get the smartest web searches for free.

Just wrote my first angel check! I'm glad to support @samarth_shyam & @agrimsingh, who are building Niyam, an AI workflow to design hardware 10x faster. These founders are cracked and disrupting an old but critical industry. If you're investing in AI, connect with them!

Every time I visit SF, I improve my taste. Every time I go to Singapore, I become locked in and ship faster.

That's where Protobuf shines!

Really excited to say that I'll be working at Blizzard over the summer on Overwatch! :D

Great example: smithery.ai/server/@21st-dev…

Similar reason why I dropped out of my PhD program in 2019. Industry seems to be leading and there are too many opportunities. Academia can't compete with industry budgets, especially when training those giant models!

Just did a quick test of the new Claude 3.5 Sonnet on @arcprize benchmark: It's about 10% higher in (relative) accuracy compared to the old 3.5 Sonnet with the same prompt/setup.

It's crazy that I can just vibe code on my phone in the car by writing product specs using @linear and @SlackHQ mobile apps with Devin @cognition_labs integration. Feature shipped by the time I got out of the car.

The new Browserbase MCP update is out and available to use via Smithery! 👇 Thanks for all the work from @kylejeong @alexdphan @pk_iv

2019: we started Jenni in our college basement fine-tuning GPT-2 2020: all our pilot users churned due to COVID 2021: we got rejected by YC W21 2022: we found PMF after many pivots 2023: we reached $100K MRR Now passing $400K MRR (2x in 5 months)!

The meta lesson here is: optimize your agents a bit but don't over optimize. Agent design is still the wild west and the next GPT will distill your tricks into the model, causing you to overhaul your design. Evals are the only model invariant.

I'm all in on MCPs

We're releasing MCP OAuth support 🔐 on Smithery! You can now drop a Smithery URL in any OAuth MCP client, and it just works. No API keys in URLs. We're now pushing for more clients to adopt OAuth to improve their UX. Thanks to MCP auth group @aaronpk for making this possible!

No more websockets! We've been pushing hard the last week and finally migrated thousands of Smithery-hosted MCPs to the new Streamable HTTP standard. This will bring in future features like resumability, OAuth, and more! Docs to use MCPs via HTTP 👇

I tried 4 different VS Code plugins for AI chat: - @SourcegraphCody - @Cursor - @CodiumAI - @GitHubCopilot So far, Cody's chat has the slickest UX, and you can use Claude Sonnet 3.5 as the underlying model, which is way better than Github Copilot's chat model.

Just realized how startup-pilled I've become after sending several of our wedding invitations via Slack 😅 @TheLiyenTan

Introducing narrative citation generation in Jenni. Jenni @whoisjenniai can seamlessly weave in citations in your writing in a more natural way instead of only adding them at the end. Took a lot of work to get this working. Coming very soon!

At least VS Code never ghosts. 😂

APIs & data models are the artifacts derived from a clear thesis. Building one is a forcing function to codify your thesis in its most precise form. It's less about the API but more about having a sharp understanding of your users & product.

Live demo: Using @cursor_ai and @OpenAI o1 to generate code to steer a drone.

Thanks to @speakeasydev and @mintlify for having me at MCP Night to talk about @SmitheryDotAI!

Just tried to vibe code Rust with Sonnet 3.7 and it's not great. In many ways, vibe coding entrenches existing frameworks and languages while making it hard to adopt new ones. Hopefully, we're not sleepwalking ourselves into a local minimum.

Just optimized Smithery's MCP deployment experience. Before: Takes ~10 steps to configure and deploy an MCP server. Now: Takes 2 steps. Our agent figures out how to configure your server based on your GitHub repository.

I'm launching Smithery, a registry where you can find tools to integrate with your LLMs/agents. You can now integrate with 51+ open-source tools using one line of code via @AnthropicAI's Model Context Protocol without the hassle of JSON schemas. 👇

Singapore OpenAI DevDay! 👇

Had a packed month visiting SF and meeting all types of AI founders/startups! The AI software tool stack is being defined this year: - Vector databases are solving the limited context memory of LLMs, like @trychroma and @pinecone - Data source connectors like @carbon__ai are bridging retrieval with files stored in different places - Search engines like @metaphorsystems are making it easier to add knowledge to LLMs - @replicate making it easier to train and deploy custom models - making it possible to run models locally via @Ollama_ai - and that's just the tip of the iceberg... All these efforts are making it easier to build end-user applications.

Flight booking AI demos smell like the classic "solution looking for a problem"

Excited for Smithery to be sponsoring Mastra's AI agent hackathon! Time to see the most creative ways people are using MCPs within their agents!

Recap for last night's AI Tinkerers Singapore Meetup: • @lewistham: Introduced alBERT, a desktop launcher that learns from Chrome activity to generate tool calls for AI agents. • Scott Albin: Showcased Groq’s ultra-fast AI inference technology and its expansion into Asia. • @rachpradhan: Demo-ed Bhumi, a Rust-based AI inference client that significantly speeds up model processing. • @kennyxchong: Presented “Chat with your Strava data”, using multi-agent workflows to translate natural language into SQL for Strava analytics. • @alexchenzl: Built NanoBrowser, an open-source alternative to OpenAI’s Operator for AI-driven web automation. • Bui Quang Huy: Answered “Can we make robots dance?” with a demo using computer vision to control robots via a laptop camera. • @leonguertler: Explored @TextArena, competitive text-based games measuring LLMs’ soft skills. Thanks to @airwallex & Hawkr ( @TheLiyenTan ) for sponsoring the venue and food!

I've been doing a deep dive on AI agents lately and will be learning/building in public on this topic for the next few weeks! To start, I drafted an open source package to easily connect your LLM agents to many Model Context Protocols so it can use tools without extra config. It currently supports @ExaAILabs and @e2b_dev. github.com/unroute/typescrip… Feedback welcome as I'm still adjusting the API design!

Had a blast hosting Singapore's first AI Tinkerers meetup! Shoutout to the awesome demos we saw last night: @hellogabbo: run local open-source models with Jan & Cortex Jia Qi: speech separation @erniesg: RAGgaeton, Scaling AI-Augmented Writing for Any Content Type @jetnew_sg: Atlas Knowledge Graphs @PradyuPrasad: Milton AI Agent for economic data analysis @zeyanliu: Using stable diffusion to transform your photos into professional portraits Naginder Singh: Finding market trends & pandemic impacts using LLMs @hrishioa: Mandark coding assistant @0xSage: Llama 3.1 multimodal speech understanding Stay tuned for the next one! PC: @marcastbury1 @TheLiyenTan

Happy to announce that our startup @whoisjenniai will be joining & funded by AIGrant (@natfriedman @danielgross @Hersh_Desai)! I'll be flying back to San Fransisco to attend the summit at the end of Sept. Let me know if you want to grab a coffee in person!

The most interesting thing about MCP isn't the protocol itself, but rather a paradigm shift where services will focus on building the best "agent experience", rather than just "user experience". We're only seeing the beginnings of this new ecosystem, and it's already promising!

Been using @wandb since 2018 in my PhD program. Back then it was an obscure experiment tracking tool that helped collaboration between people in my lab. Now they've got a full conference with many attendees. Crazy how fast AI has exploded in the last few years.

I'm running an LLM based startup @whoisjenniai. While ChatGPT is great general purpose tool, there are many user needs that are not covered. Focusing on those pains is key. Just because there's Google search doesn't meant there's no need for Wolfram Alpha.

New Smithery docs, open source on Github. Built using @mintlify ( @handotdev @hahnbeelee )! smithery.ai/docs

Another incredible evening at the AI Tinkerers Singapore 4th Meetup! Here’s a recap of the amazing lightning talks and demos we had: - Vivek Gudapuri: Showcased how to optimize multi-agent workflows using the LangDB gateway, focusing on cost and time efficiencies. - @huoru_tarnish: Introduced Hika AI, an intelligent search engine designed to provide deeper insights and personalized knowledge. - @calclavia: Explained Model Context Protocols and demonstrated Smithery AI, a registry for installing MCPs to enhance LLM capabilities. - @rachpradhan: Showcase Kew - An async Redis-backed task queue - Nguyen Quoc Anh: Presented a system using zero-shot learning with foundational models for animal recognition in camera trap videos. - @ivanleomk: Introduced Open Clio, a tool to cluster and visualize chat history with Claude, offering insights into user patterns while ensuring privacy. - Bryan Zhao: Showcased the OneService Hotline, an AI assistant for elderly users, highlighting their hackathon success. - Nicholas Chen: Explored cognitive architectures as a structured approach to building AI agents. - Kai Yuan Neo: Shared his experience building a camera-based travel assistant app using AI, focusing on practical integration and coding challenges. - @joentze: Presented Onword AI, a platform enabling cross-application actions, emphasizing its potential for everyday users. Thanks to @awscloud for providing the venue and @DataStax for sponsoring the delicious Hawkr meal boxes ( @TheLiyenTan ).

We're launching Smithery Playground 🛝 - the most ergonomic place for you to explore, test, and debug MCPs! 1. See how models interact with multiple MCPs 2. Drill down and inspect MCP traces in Dev Console 3. Connect to MCPs running on localhost 👇

Why AI copilots are better than AI agents: As we scaled Jenni to 250K+ monthly active users, we tested many designs for our AI features. Some designs help users complete tasks autonomously while others assist users via human-in-the-loop. We found that the copilot experience tends to win. Why? To recap the definition: - an AI copilot is when the AI assists you in a task with frequent feedback/input - an AI agent does your task autonomously without much intervention Error propagation Agents, by definition, try to accomplish larger goals autonomously through a sequence of steps. The problem with agents is when one step fails, it propagates errors toward subsequent steps. It's like giving an intern a 3-month waterfall project and never checking in until the end. In many cases, LLMs cannot recover from these errors even with revision-style prompts. This yields garbage output that the user has to manually fix, costing more time than doing it manually. Copilots sidestep this by prompting users for feedback at each step in their existing workflow. How about when LLMs improve? Even if we had better LLMs, the fundamental issue of having non-deterministic programs makes it not suitable for high-stakes tasks - especially for tasks like writing code. You don't want your compiler to compile your code correctly 50% or 90% of the time, it must be correct 100% of the time - and do so with proof. There will be some exceptions for tasks where quality matters less. But if your users still need to check how the sausage gets made, then they'll be wasting rather than saving time. Copilots win today, Agents can win tomorrow To qualify, I'm claiming that AI in the next ~3 years is not ready for agents. Today's LLMs are best used as copilots to solve focused problems seamlessly while keeping you in the driver's seat. Creating a good copilot that doesn't fight against the user is hard. For example, our writing assistance feature at Jenni took a year to refine with user feedback and we're still improving it. Tomorrow's agents will be built from data collected from today's copilots. --- TLDR: We should learn to walk before we try to run.

Fake artificial intelligence = real intelligence?

Smithery is Google for the new internet.

Want to build an MCP but not sure what people want? We've created a GitHub discussion community where users are posting about MCPs they need but may not exist 👇

"Tool calls are the new clicks." A live demo of Smithery at @southpkcommons 👀

Looks like I won't be leaving Minecraft soon. Officially got a job that's related to MC Modding. :)

Want to win up to $6K in prize credits building competitive AI agents? I'm hosting an MCP Hackathon on March 8 in Singapore, backed by @AnthropicAI and @AWS. Registration is now 50% filled and we're 2 weeks from the event! Be sure to sign up early and form your team 👇

Smithery now supports MCPs hosted from anywhere! 🌎 No matter where you're hosting, you can now list on Smithery and get a storefront for your MCP. Submit a URL, and we generate your landing page in 30 seconds.

Moved into our new place in Singapore 🇸🇬! @TheLiyenTan The best thing about SF is the tight-knit startup community that lives in close proximity. We're transplanting the energy from those Victorian-style houses into this renovated heritage shophouse with 16 empty studios.

Sampling from the Llama 70B model locally on an M4 Max MacBook Pro using @jandotai: 9.72 tokens per second. Pretty usable but also warms up the laptop ☀️

How we optimized our servers by 20x 💸 @ 250K monthly active users ​ Several people have been asking why we use Rust, so here's the full story: ​ Back in July 2023 📅, we were serving 80K monthly active users. About 4 requests per second were hitting our servers at any given time from Jenni users. ​ Like many AI startups, we built our initial backend server with Python. Python was a good language to prototype. It had convenient AI/ML libraries (Numpy, OpenAI SDK) and Jupyter notebooks were great for tinkering. ​ Our backend server is monolithic and handles many operations on Jenni, including prompting, RAG pipeline, etc. It's deployed on serverless Cloud Run. ​ The Problem 🐍 By mid-2024 📅, we scaled to 250K active users and were serving 10+ requests per second. ​ As the complexity of our product scaled over time, 2 main problems stood out: ​ 1. High Resource Consumption Python was consuming a lot of CPU and memory, and it was hard to debug. Sometimes containers took a minute to start. ​ 2. Increased Bugs One source of bugs was the lack of strictness in Python as a dynamically typed language. For example, you could pass an object into a prompt string, and no compiler would warn you something went wrong. ​ These silent errors were costly and something we couldn't afford. ​ And we've tried Python typings and linting - the tool stack really sucked. ​ Getting Rusty 🦀 After hearing success stories from friends who tried Rust and surveying other developers' opinions, I decided to prototype a simple API endpoint in Rust. ​ We deployed this prototype to production and it ran extremely fast. More importantly, the strict compiler ensured a class of errors that cannot happen. ​ Being 100% confident about something is much more valuable than creating many unit tests that lead to 99% confidence. The best tests are the ones you never had to write. ​ Outcome Within a few months, our team got the hang of Rust and was able to rewrite our Python server into Rust. ​ Results: • Cost Optimization: Without further optimization work, our server cost was optimized by ~20x. • Faster Container Starts: Our containers start within ~0.6 seconds - an order of magnitude faster. • Less Compute Usage: ~60x less compute to serve the same traffic. ​ Rust helped us scale up by scaling down and doing more with less. ​ Although Rust does have a learning curve and lacks an abundance of libraries like Python, the tradeoffs made sense for our team and scale. ​ We still prototype in Python and use it for research, but Rust is a clear winner for reliable production code.

Now that you've posted this tweet, it'll be in the training set of the next GPT and get memorized.

We had an incredible night at our 3rd AI Tinkerers Meetup in Singapore! A huge thank you to our speakers: @leonguertler - TextArena: Introduced a benchmark for evaluating language models through text-based games. @SherryJiang - Peek: Showcased an AI agent for finance assistant. @vivekkalyansk - Cartograph: Presented a tool that generates interactive visualization and documentation from codebases. @hellogabbo & Jia Qi - Ichigo: Showcased a voice AI that improves multi-turn conversations and context handling. @joentze - Onword AI: Showcased the future of app orchestration with AI agents. @arjunkmrm - Promptspace: Introduced a tool for visualizing complex prompt chains. @ivanleomk - Discussed creating branching stories through iterative generation and DAGs. @ThienTran - Explored leveraging INT8 Tensor Cores in NVIDIA GPUs to speed up transformer model training. @gabrielchua - Discussed the application of GPT vision finetuning to chemistry diagrams Thanks to our sponsors: @awscloud @stripe @TheLiyenTan Stay tuned for our next meetup!

Happy New Year! 🎉 My 2024 wrapped: - Exited from my first startup after 6 years of hustling - Met many awesome builders in San Francisco in early 2024 and Singapore in late 2024 - Started Singapore's @AITinkerers chapter - Made my first angel investment - Launched Smithery AI as part of my exploration with agents and reasoning

Visas can be stressful, especially when I hear stories from friends trying to work in the USA or get an O1. ​ Fortunately, my experience getting a Singapore visa as an American has been relatively frictionless. ​ Applications are all online, streamlined, and there's no dealing with multiple departments. Rules are clear and things got done within a month. ​ Above all, when you walk into the immigration center, instead of referring to you as an applicant, immigrant, or alien, they call you a customer. ​ The framing is subtle but significant - an example countries should learn from if they wish to retain foreign talent. @SGglobalnetwork

Spending the Sunday at a GPT-5 hackathon

Remote work was neither the revolution they promised nor the catastrophe they now claim. After 6 years building a remote startup, I developed a habit of writing things down - a culture I'm bringing to @SmitheryDotAI despite us being in-person half the time. Many in-person teams run on tribal knowledge. Quick hallway chats, whiteboard sessions, decisions made over lunch. It quickly aligns people, but lacks depth. Writing is useful because it forces knowledge compression. You have to turn your fuzzy thinking into tokens. You can't document a decision or RFC without understanding it. Written communication also prevents team members from thrashing. People can have deep work blocks instead of constant interruptions. This creates space for divergent thinking - quieter team members can fully develop ideas before meetings, leading to genuinely different approaches. Writing also scales better. You write once, and all new hires onboard faster; 2-year-old decisions become searchable. Now, with MCPs that integrate with all your work context, like @Notion, having years of written context becomes an unfair advantage. Agents can leverage your writing to scale your work even further.

I was recently asked: "Why gamble and build a second startup?" Perhaps contrarian in the valley: I actually refrain from any "casino gambling" or poker. 😅 But building a startup is fundamentally different. Gambling is zero-sum. Someone wins, someone loses. Building is positive-sum. Every risk becomes a challenge. Every challenge forces growth. Every unit of growth makes your craft closer to meeting the challenge. And this itch for raising the bar - to constantly create, discover new insights, and invent new things within a team - is why I'll keep building.

Officially graduated. Got to throw those fancy hats. :)

Use it for Claude Desktop, @windsurf_ai, @cline and other clients: smithery.ai/server/@smithery… (credits to @arjunkmrm, @saradu, @zirayhao for making this epic demo possible!)

Had a great time in San Fransisco! Hosting my last founder dinner this Friday before heading back to Singapore. There'll be co-working, dinner and board games! Details:

Intelligence and simulations go hand in hand. Strong intelligence is likely an efficient environment simulator. We probably need more sensors/embodiment if empirically bound.

One thing I'm increasingly doing is switching between o3 and Sonnet 3.7 in my @windsurf_ai. I need a hot key for this. Use o3 with "chat-only" mode to plan, and Sonnet 3.7 with write-mode to implement. Much better than trying to get o3 to tool call directly.

Congrats @WilliamBryk, @jeffzwang! Rooting for Exa since the Metaphor days 🔥

Evolution is noisy gradient decent. If it worked for us humans, it's plausible for machines.

Happy to see our academic copilot Jenni help researchers get published and cited!

Forget the plots. I can't believe we're not putting batch on dim=0! 😅

I haven't found remote hackathons to be effective. The whole point of a hackathon is to have a high but concentrated dose of exploration. It's espresso engineering.

This is a revolution. EDX now runs on both MC 1.7 and 1.8. Same mod file. Thanks to @asiekierka for inspiring us to make NOVA. :)

Visited a startup's office @menloresearch where they might eventually build an MCP so LLMs can prompt their robot!

Went to @sandersted's OpenAI talk at @wandb Fully Connected conference last week. Here's my main takeaway + personal experience working with GPT-4: 👇

Doing some end of year traveling before moving to San Francisco next year!

Inspired by the good vibes from the last SF AI Tinkerers I attended (hosted by @mickeyxfriedman), I'm hosting Singapore's first AI Tinkerers event on August 2nd! This is your chance to showcase your latest AI and LLM technical demos, projects and features to other builders! 👇

When brute force O(n) doesn't work, break the problem down into a hierarchical binary tree O(log(n))!

What actually happens when you put lava in a furnace in Minecraft: vimeo.com/104105171

Seems like this will turn into a generative adversial game in the near future...

Graduating soon, then moving to the United States. This is going to be a new adventure!