Get live statistics and analysis of Michael's profile on X / Twitter

Literally 20 companies are lining up trying to hire this kind of person. If this is you dm me and I'll introduce you to a founder.

I’m willing to pay $300 for software that goes through my twitter dms, deepresearches and identifies everyone, and places them into a CRM.

I couldn't find a single API on the internet that just takes a twitter user's handle and enriches their information (at least with a Linkedin profile/email/relevant websites.) Someone is going to solve this problem and have people absolutely begging to pay them.

People forget more value was created by YC rejects than successful YC applicants

Hetzner blocked my accounts twice. I’d make an account, miss a payment or something, and bam I can’t log in to pay my balance. Burned my two primary emails doing this. Could never reach a person. Nobody cared about my business. Hetzner is literally 5x cheaper than some alternatives, which all have their own issues. But the fact that I literally could not reach a human was the reason I gave up on them. I’d rather pay 5x more somewhere else that actually responds. Even drug dealers have better support (I assume) In practice, this really sucks. There are things I want to do that are only economically viable with Hetzner.

I'm deploying on Vercel. NO!! you should deploy on Cloudflare! NO!! you should deploy on Hetzner! NO!! you should deploy on a raspberry pi! Guys, 4.5 sonnet can deploy your app on a vape if you want it to. It's not a big deal.

If you're a successful founder and were rejected by YC, either initially or in all attempts, please DM me. I'm attempting to accurately estimate the value of all companies with founders rejected from YC. Here's what I have so far:

WE'VE RAISED $11.8M 🚀 It's been an absolutely incredible watching this brilliant team grow over the past 5 months. When I joined, Inference was 4 people in a mostly empty office. Now it's almost double that, and while some things have changed (Smash has been perma-banned at the office and the fridge is now religiously stocked with cherry-cola Celcius), the insane shipping velocity I joined for is still going strong. I came in at a precarious time: Inference had been building out a network for decentralized LLM inference for years, but the team decided to go all-in on a completely new direction: small language models that crush frontier models on specific tasks, but at a fraction of the cost. Five months later, our conviction, and years of investment into LLM inference infra, has clearly paid off. We've already saved AI-native companies millions, slashed their inference latency, and enabled projects that would have been infeasible without small, specialized models. I really can't wait to announce all the things we've been working on over the coming weeks. Expect lots of very excited tweets from me!

>it’s 2026 >you have an AI app idea >you go to v0 >it builds you an app in Next.js + shadcn using a fine-tuned Next.js LLM >AI gateway + AI SDK supported out of the box >your app is cooked >change your theme with one prompt >want vercel analytics? add with one click >bot protection? done. >one click deploy to Vercel >you need a domain name >vercel domains integrates with your deployment, you can buy with one click >it’s 6 weeks later, and you’re at 20k MRR >Vercel sends you an email: based on your analytics and AI logs retention is dropping slightly >you open up Cursor for the first time and clone the repo >you change some of the copy >push straight to prod >vercel Agent finds that you somehow exposed user data >it also thinks the previous copy was better >instant rollback >you look at your vercel fee >it might be help your bottom line if retention was even worse >you receive an email to turn on Fluid compute >now your Vercel bill is 30% lower >you have a new AI app idea >you go to v0

Vercel needs way more credit for this imo. DX like this doesn't grow on trees.

Reflection 70B just dropped, beating GPT-4o and Claude Sonnet on key benchmarks while being much smaller. This is huge, but the real story is how it happened: Matt Shumer isn't your typical AI researcher. He's a prompt engineer who's been in the trenches since early GPT-3 days, building for millions of users. That experience led to a key insight: LLMs suck at admitting mistakes when reflecting. They'll make errors just to have something to correct. The solution? Fine-tuning for genuine self-correction without new errors. So simple it's brilliant, and the big labs missed it. The result is a model that thinks only when necessary. And with reflection built in, you can very easily hide it from the user. Built in 3 weeks by 2 guys who get how LLMs actually behave in the wild. No billions or years of research needed. It's open-source. 405B version next week, maybe dataset release too. Innovation flood incoming. Big Tech is likely scrambling. Their models, often built by researchers without real-world experience, look outdated now. Grok's system prompt is embarrassingly bad. Claude already does this with prompting (ask it to think carefully, you'll see inference slows down, because the UI doesn't show text in between tags), but fine-tuning it in could be an easy upgrade. Let's be very clear: models are going to get much better in the next month or two. What Matt did is simplistic, and there are a lot more easy gains to be had. For example, if you prompt a model with "I'll tip you $20 if you get this right, or "My job depends on this," you get better outputs. With better training, that might not be as necessary. Chain-of-Thought and Self-Reflection are also very simple, and this was built with a SYNTHETIC dataset. Can you imagine if we used people to carefully curate a much better dataset? If we experimented with other forms of CoT, self-reflection, multi-agent reasoning, etc? There are so many things to try, and you don't have to be at OpenAI to try them! Long-term, models might not need to fine-tune CoT/Reflection, this kind of thought process might be deeply baked into them. For now, it's an easy win. Yes, it's slower – it adds a reasoning step. But with Groq/Cerebras hitting 1800+ tokens/sec, does it matter? Every IQ point for a model is crucial. I'll happily wait 100ms for better results. The future of AI isn't just bigger models. It's smarter training, real human-like thinking. And it might come from engineers, not PhDs. Reflection is just the beginning. Watch this space.

This is unironically Guillermo’s master plan for Vercel. The first stage was vertical integration across the stack: >created Next.js >Vercel to deploy apps easily >Turborepo acquired to make dev experience better and to own monorepo builds/caching >Fluid compute and heavy investment into serverless infra >Vercel Firewall Then horizontal integration across the JS ecosystem: >NuxtLabs acquisition >Hiring Rich Harris (svelte) >Astro sponsorship/partnership Then horizontal integration across the new AI development stack: >SplitBee acquisition->Vercel Analytics >hires ShadCN (guy behind ShadCN, component lib for humans and AI) >v0 to prototype apps (uses ShadCN heavily) >hires tweakCN creator to make ShadCN/v0 apps look better >Vercel Marketplace for easy integrations (including AI tools) >Vercel Agent for code reviews (especially important with so much AI code) >AI Gateway to make integrations even easier and collect valuable data (being a router is very very good. It allows you to build all kind of cool stuff and Vercel is well-positioned to capitalize) >BotId to protect AI routes and login from malicious bots. Horizontal integration was always the plan, but the AI era made it many times more profitable and a powerful force multiplier for Vercel. James Hill famously used both vertical and horizontal integration to dominate the Railroad industry. He built feeder lines, bought docks, and made vertical investments to reduce costs. He also built a horizontal trust of Railway lines (that later was broken up by Antitrust.) Guillermo is doing the exact same thing to dominate Javascript. Honestly? Bullish.

Shitty search experiences cause me unspeakable pain. >I'm looking for a YC startup called Webhound. >I search for 'hound' >Webhound, a yc startup, isn't in the top 30 results >I search 'yc hound' on google and the first result is Webhound. I'm willing to submit a PR to fix this and make YC company search 100x better. This will undoubtably have a noticeable impact on business outcomes for YC companies. How do I get access to the YC codebase?

apparently this is what happened with browserbase they built some of the best browser infra in the world, and one of the top use cases ended up being automating linkedin following/messaging for salespeople. so yes, browserbase is a sales tool as much as it is a devtool. revenue is revenue.

Literally all the resources you need to be a cracked ML/LLM engineer (coming from someone learning ML rn) >unsloth documentation >vLLM documentation (goated) >HF TEI docs (worth reading) >GPU glossary I wish I started with these a month ago. Expect much better ML content from me now that I have found them. What am I missing?