Trung Vu

The new toy I've been working on Video Wiz: Summarize and ask question with any youtube video Tech stack - @nextjs - @tailwindcss & @shadcn UI - @langchain4j - OpenAI API - @vercel for hosting & AI streaming - @upstash for memory - @supabase for db & vector store Took about 5-6 hours to get to this point. Shamelessly copied from multiple starter kits to get here - TailwindUI Salient template - Azure ChatGPT - @a16z companion app starter kit Learning so far: - Building chatgpt wrapper is not easy. It took me a few hours to get summarization work - Langchain devx leaves a lot to be desired - Need to more ruthless trimming down scope What's left: - Adding auth to keep track of chat history - Adding chat 1 day from launch, stay tuned #buildinpublic @_buildspace

AI is not the job stealer, it's the new electricity. Don't fear it, harness it. Remember, it's not #AI taking jobs, but those who know how to use AI effectively. So, let's educate ourselves and leverage AI for our growth. #ArtificialIntelligence #SelfImprovement

@pedroh96 Happy Brex customer and fellow YC founder here. How is brex using AI agents for its customer operations (support, sales). How do you plan to bring innovation on AI agents at brex to cap 1

I’ll be here in about an hour for an AMA, responding to anything and everything. Roadmap, velocity, startups, how I like my steak. Drop your questions below.

@NVIDIAGeForce Geforce day

🟢 GEFORCE DAY IS BACK 🟢 To celebrate, we're giving away TWO GeForce RTX 5080 Founders Edition GPUs, signed by NVIDIA CEO Jensen Huang. Want one? Comment "GeForce Day" for a chance to WIN & stay tuned for more!

@pedroh96 Congrats Pedro. 100% agree with marrying well

One month from the best day of my life. ❤️ Marrying well is (unsurprisingly) the biggest life hack.

@kwindla Amazing! Thanks Kwin. We literally just talked about building a model for Vietnamese ourselves

Tiny SOTA model release today: v3 of the Smart Turn semantic VAD model. Smart Turn is a native audio, open source, open data, open training code model for detecting whether a human has stopped speaking and expects a voice agent to respond. The model now runs in <60ms on most cloud vCPUs, faster than that on your local CPU, and in <10ms on GPU. Running on CPU makes it essentially free to use this in a voice AI agent. 23 languages, and you can contribute data or data labeling to add a language or improve the model performance in any of the existing language. This model is a community effort. We think that for what we built it for, this model benchmarks better than any other model that's currently available. But if you're interested in turn detection, you should also check out excellent recent work from the @krispHQ and @ultravox_dot_ai, teams, which have released models that are very good, make somewhat different trade-offs compared to the Smart Turn model, and do better than Smart Turn relative to their respective goals. Super-fun things happening all the time these days in voice AI! Anybody can use the Smart Turn model in any deployment. It has no license restrictions and is completely open source. It's bundled into the upcoming @pipecat_ai 0.0.85 release. And, of course, it's available on the Pipecat Cloud voice agent hosting platform.

@harryqi666 Congrats Harry and the team. I still remember when you guys were building trading app. Crazy how far you have came

On Christmas Eve 2019, Michael Seibel, then CEO of YC, called me: “You might not be cut out to be a founder.” He was right. We had pivoted 20 times. I thought I could always return to my 7-figure quant job. That night, I realized: you commit or you quit. Fast forward: Motion raised $60M at $550M. Michael is now a personal investor.

@akshay_pachaar Is there an open source alternative to voyage 3?

A new embedding model cuts vector DB costs by ~200x. It also outperforms OpenAI and Cohere models. Let's understand how you can use it in LLM apps (with code):

@DaivikGoel @ycombinator @shor_pay @avikonduru @michaael_chaan @zavidavidoo Which countries do you support?

@ycombinator @shor_pay @avikonduru Big thanks to @michaael_chaan and @zavidavidoo for making this happen! And learn more about us at: tryshor.com

Shor (@shor_pay) is the fastest, simplest, and most affordable way to run global payroll. Onboard in minutes via Slack, generate contracts instantly, and pay in local currency within 24 hours. ycombinator.com/launches/OBs-s… Congrats on the launch, @daivikgoel & @avikonduru!

@ArthurMacwaters You can roll over to a self directed IRA (not roth) and invest in whatever. Or even better do a ROBS (rollover as business startup) and use the money for your next startup

I have a fairly small amount of money in my 401k from mckinsey… Can someone explain why I shouldn’t just take it out early (pay taxes and penalties) and put it all in Tesla and Palantir? Like not joking is this dumb

@kwindla How do you run the 120b model on your macbook 😮

A voice agent powered by gpt-oss. Running locally on my macBook. Demo recorded in a Waymo with WiFi turned off. I'm still on my space game voice AI kick, obviously. Code link below. For conversational voice AI, you want to set the gpt-oss reasoning behavior to "low". (The default is "medium".) Notes on how to do that and a jinja template you can use are in the repo. The LLM in the demo video is the big, 120B version of gpt-oss. You can use the smaller, 20B model for this, of course. But OpenAI really did a cool thing here designing the 120B model to run in "just" 80GB of VRAM. And the llama.cpp mlx inference is fast: ~250ms TTFT. Running a big model on-device feels like a time warp into the future of AI.

@anarchyco This is amazing. Thank you for sharing. for custom pipeline, would be great to have option to use livekit, pipecat, or similar

livetok.io/cost-calculator Feedback welcomed, will keep extending/updating it based on comments if there is enough interest. Thank you!

Releasing first version of Voice Agents cost calculator 💰with different providers: Conversational AI Platforms, RealTime Models and Custom Pipelines Link below👇

@clarklab @boltdotnew ooh, i need help making a launch video

Y'all seemed to really like the launch video for my submission to the @boltdotnew hackathon, so I made a video showing how I made the video. Here's how you can do it: 🗝️ Keyframes! Plan out the video's main steps, and then animate between them 🖥️ Use whatever software you know. I used @figma and @ScreenFlow, but anything with a timeline would work 🏎️ 60fps and high motion blur on the easings makes things seem just a litttttle bit smoother 🗂️ Templates! There are so many good starter templates out there, especially in @figma Check out the video below 👇

@JeffLonsdale And now the song is stuck in my head

It was much cuter when they did the same thing with “You Are My Sunshine”

Somehow preschoolers through first graders at this school music recital all know the APT song. (One kid played it on the piano and all the kids in the audience were singing along)

@clarklab @tailwindcss Love this thread. THanks for the detailed breakdown

The @tailwindcss team just released a new template and there are some SERIOUSLY neat little #CSS tricks under the hood. It's called Compass, and it's for e-learning with a video focus. I'm sure it's been shared before, but here's what I found especially clever, a thread 🧵

@jasonlk I was at a talk by Meagen Eisenberg today and she said if her marketing team members doesn't use AI to get better, they wouldn't have job in marketing for very long

I don’t think AI will take away all tech jobs But it may eliminate the need for the mediocre: - The mediocre sales rep that doesn’t know the product - The mediocre marketer that sends canned campaigns - The mediocre engineer that wants a cushy job Not clear we will need them in 24 months. AI is already better than them in many cases as it is. We already don’t need mediocre support reps anymore in many cases. Many have been able to use AI to eliminate the bottom 20%-30% of customer support reps.

@kwindla @modal_labs This look awesome. Looking forward to it Coincidentally, will be in the Bay Area during most of the course, any Voice AI meet up I should attend

🥳🛜🎇 $500 @modal_labs Voice AI credits 🎇🛜🥳 Our friends at Modal are giving everyone in the Voice AI and Voice Agents course $500 in usage credits.

@kwindla Would this reduce the latency significantly?

Serverless WebRTC for Voice AI Introducing the new `SmallWebRTCTransport` in Pipecat 0.0.63 ... 🧵

@jasonlk Amazing line up. It’d be my first time at Saastr and can’t wait for it

The First Wave of AI and GTM Speakers is Here for 2025 SaaStrAnnual.com, May 13-15 in SF Bay! 1000+ speakers, braindates, workshops, and 1-on-1’s! AI Speakers: 🤖CEO Snowflake + CEO Observe: “Where B2B Applications Are Going” 🤖CEO Box Aaron Levie: “AI, Agents and The Next Era of SaaS” 🤖COO Google Cloud Francis deSouza: “Hyperscalers: The Future and More” 🤖CTO Rubrik: Co-Founder & CTO, Arvind Nithrakashyap 🤖CTO Neo4j: Philip Rathle (Valuation – $2B+): “How Revolut Left Salesforce and More: Rolling Your Our Own SaaS App with AI” 🤖CEO Descript Andrew Mason: “The Future of Marketing in the Age of AI 🤖CMO & President Infinite Reality Karine Kogan: “Build Brand in the AI Era” 🤖CEO Instabase Anant Bhardwaj 🤖GTM OpenAI Maggie Hott: “GTM at AI Speed” 🤖CBO Perplexity Dmitry Shevelenko: “The First $100m ARR in AI” 🤖CEO Decagon Jesse Zhang: “The Future of Agents and Automation: Where Will We Need Humans?” 🤖Head of AI Dialpad: “How to Build AI at Scale” GTM/ B2B Speakers: 🚀CEO HubSpot Yamini Rangan: “Going More Multiproduct, Going More AI, and Going More SMB and More Enterprise” 🚀CEO Dropbox Drew Houston: “DropBox’s Third Act: AI & Content Intelligence” 🚀CEO Calendly, Tope Awotona “Open AMA and AI in 2026” 🚀CEO Clio, Jack Newton: “Reaccelerating Vertical SaaS to $3 Billion” 🚀CRO ServiceTitan Ross Biestman: “The Secrets to Scaling in Vertical SaaS” 🚀CMO Asana, Shannon Duffy: “Adding AI to SaaS: Pricing, Product and More” 🚀CEO Gusto Josh Reeves + CTO Gusto Eddie Kim: “Remaking SMB SaaS in the Age of AI” 🚀ex-CMO Ramp, Guillaume Cabane: “The Top GTM Lessons from Scaling Ramp” 🚀CEO Own, Sam Gutmann: “Top 10 Learnings Selling to Salesforce for $1.9 Billion” 🚀VPM RevenueCat Rik Haandrikman: “Powering 40% of All Mobile Subscriptions: Bottoms-Up Marketing, PLG, and More — How to Really Do It” 🚀VPM Palo Alto Networks Navneet Singh 🚀CCO Stripe (Head of Global CS and Product Sales), Tim Smith 🚀VPCS Notion Monica Perez 🚀VPCS Intercom Junan Pang 🚀VPCS GitHub Abbas Haider Ali 🚀VP Product Rippling Anique Drumright 🚀SVP GTM Optimizely Rohit Bhalla: “What’s Working in 2025+ in GTM” AI & GTM Workshops: 📈GVP IBM AI Raj Datta 📈CPTO Digital Ocean Bratin Saha 📈VP P Rippling Anique Drumrightbha 📈CEO Demandbase Gabriel Rogol + AWS Head of North America Data & AI Sales, Arym Diamond 📈CEO Flatfile David Boskovic 📈VP Revenue Marketing Docebo Silvia Valencia 📈Databricks – Sr. Industry Solutions Manager – Digital Natives, Madelyn Mullen 📈GitHub – Global Engagement Lead, Startups, Alex Malebranche 📈Twitch Video / Amazon Interactive Video Service – Product & Growth, Khasimul Asif Mohammad 📈G2 – CMO, Sydney Sloan and VP of Research Insights, Tim Sanders 📈CEO Artisan Jaspar Carmichael-Jack: “The Real Data on AI SDRs” 📈VP of Product Advocacy and Market Strategy, Darin Patterson 📈CEO & Founder Runway Siqi Chen: “Your Customers as Prompt Engineers” 📈CEO DuploCloud Venkat Thiruvengadam 📈Controller OpenAI Sowmya Ranganathan VC Deep Dives: 💸Accel Partners Ivan Zhou 💸A16z Seema Amble 💸Emergence Capital Jake Saper 💸Bonfire Ventures General Partner Brett Queener 💸Mayfield Fund General Partner Navin Chaddha 💸Chemisty Ventures Kristina Shen 💸Tidemark Founder Dave Yuan 💸Innovation Endeavors Partner, Harpinder Singh 💸Softbank VP Isabela Lopes

@kwindla @ZainullAbiddinD @GroqInc @pipecat_ai @twilio what are the ~100ms providers for STT & TTS you mentioned?

Voice-to-voice latency measured at the client was 1200-1400ms last night, on launch day. Based on what we've seen before, I expect that Llama 4 will get *even faster* on Groq as they continue to optimize. (It's already really fast.) Rough numbers: STT - 250ms LLM - 300ms TTS - 250ms VAD stop_secs - 400 network/audio - 100ms Opportunities to make this faster: - Groq has traditionally focused on throughput more than latency. As voice AI use cases become more and more import I expect they'll implement more ttft/streaming optimizations. SOTA for STT and TTS right now is down around ~100ms. - I was using basic VAD for turn detection for that demo. You can use a new smart vad model like this one: github.com/pipecat-ai/sma…

Llama 4 voice agent starter kit with @GroqInc and @pipecat_ai ➡️ Groq STT (distil-whisper-large-v3) ➡️ Groq Llama 4 (llama-4-scout-17b-16e-instruct) ➡️ Groq TTS (playai-tts) ➡️ Function calling ➡️ Deploy to Pipecat Cloud for production ➡️ Optionally add a @twilio phone number for telephone voice AI

@AskPerplexity What can go wrong with Trump’s tariffs plan

Treasury Secretary Scott Bessent went on Tucker today to talk about Trump's tariff policy and its impact on the middle class. I listened to his whole interview and everything else he's said about it and think he made some solid points. Ask your questions about tariffs below: