lowkey goated uber driver

Two years ago, we built a custom file format for container images to achieve sub-second cold starts. When you run a serverless platform, your product lives and dies by performance. Image pulls are one of the main variables you need to optimize. One way to do this is by building a custom FUSE filesystem that's mounted on the host, and allow the container to lazy load the image as it runs. Initially, we tried to find off the shelf projects like AWS SOCI and Nydus, but they just weren't fast enough. We ended up building our own file format, called CLIP. A side benefit of controlling the interface is that we have programmatic hooks into the image format. We can easily distribute layers and cache layer content, which is extremely useful for cross-cloud workloads. It's an investment we made early on, and the performance gains pay off every single day.

One piece of advice we got during YC was to explain our company using verbs instead of nouns. Early on, I walked into a meeting and did the opposite: “We’re building a cloud platform for AI” No one knew that that meant, their eyes glazed over. Then I started saying this instead: “We containerize your code and run it on GPUs in the cloud so you don’t have to manage the infra yourself” That clicked way more. Our brains understand verbs because they’re more concrete. If you describe your company using nouns, you risk people not understanding you. And no one buys or invests in things they don’t understand.

construction on the Van Wyck en route to JFK is just part of the simulation as you travel farther from manhattan, the game designers added an intermediate loading state to upscale graphics to smoothly transition to the next scene

"we're a copilot shop"

Everyone agrees with this idea, but when you dig into the use cases for agentic file systems, people really want an entire environment for their agents filesystem sandbox, integrations, orchestrator The filesystem isn't going to be a point solution, but will be a component of a broader platform

People are using the cloud in completely different ways than they did even 5 years ago: - Reading 30Gi checkpoint files every time an API is invoked - Forking the state of a running VM - Spawning 1,000 new VMs and reloading that forked state across all of them Developers have always wanted an alternative to AWS with better DX. But these use cases go so far beyond ergonomics, they're new ways of using compute entirely.

You know a market is early when 100 products claim to do what you want but none of them actually solve your problem

I understand why @UseCorgi is a $2B company

pov: you didn’t escape the permanent underclass

By far one of the coolest AI products on the market right now

Many people aren’t stuck because they aren’t doing enough cold outreach, they’re stuck because they aren’t building real relationships with people. I get hundreds of emails a week that all sound the same: “Hey — quick one!” “Running a serverless GPU cloud with sub-second cold starts is seriously impressive!” “I’d love to connect and explore ideas with you!” If you’re doing this, just stop. Instead, try genuinely building relationships with people. You will get so much more in return than treating people like strangers at a trade show.

We almost got sued because of CUDA drivers. Biggest deal in our company's history, thousands of GPUs. But the drivers were wrong. Obscure errors everywhere. My co-founder was on his honeymoon. I wrote a letter to my lawyer asking if we’d get sued if we didn’t deliver in time. @llom2600 managed to get WiFi somewhere over the Atlantic Ocean and upgraded the drivers from the plane. The campaign launched that Monday at 2am. We set our alarms and watched the nodes spin up. It worked flawlessly. We went back to sleep.

We went through YC with a browser-based IDE with GPUs attached. Sort of like Replit but for ML. It looked super sick and investors loved it. But when we watched users interact with it, they had no idea how to use it. They'd click around the dashboard and leave. In a fit of frustration, we deleted the frontend and turned the entire experience into a Python SDK instead, and that's when it really clicked.

some reflections from solely using cloud agents this year: 1. every engineer should default to cloud. it completely changes how you view and use agents. if you run a company, it might be worth mandating everyone starts in cloud 2. cloud agent adoption has been much slower than i expected— e.g. looking at a ton of cursor profiles it’s clear majority cloud usage is still rare 3. getting your dx cloud agent ready still requires creative jiu jitsu. dev infra docs could be much better — “this is how to make our stuff accessible to agents/parallelizable.” luckily investments also benefit humans 4. it’s still a PITA to setup & manage cloud envs across cursor/devin etc. but i assume it’ll get bitter lessoned and we don’t need conventions for setup scripts etc. 5. where are the labs?! would love to see codex et al. invest more in their cloud experience. i know they can do it :) 6. it’s strange that cursor/devin’s investment in mobile apps lags behind their investment in cloud agents. they should go hand in hand. the ability to start agents from slack mobile isn’t enough! 7. a cloud agent spinning up other cloud agents (middle manager pattern) is goated. e.g. nice to go for a run, yap for twenty minutes, and end up with parallel agents. only devin supports this well 8. the uis of ADEs have somewhat adapted for cloud agents. but ui patterns for upcoming long running *and* proactive agents are understudied. super excited to see more experiments here (and will contribute) overall: i freaking love cloud agents. you’ll dissappoint me personally if next month you still spin up more local agents than cloud. very grateful for cursor and devin for making this technology so easy to use!

My ideal agent harness: - Works on my iPhone - Connects to GitHub - Runs on a schedule - Sends me UI screenshots with changes it made - Can set system prompts for each project - One long-running chat with a “manager” agent that dispatches subagents for each task - Can use both Claude and Opus models - Smart context management; chats should compact automatically What should I be using?

If you're selling to developers, live chat is one of the best sales tools. Self-serve users aren't going to join a call, but as soon as they get blocked, they'll reach out over chat. So far, someone asking thoughtful, pointed technical questions in chat are the #1 leading indicator that they're going to convert. When we see that signal, we hustle super hard to make sure they're successful.

The GPU market is wild. You can't get anything on-demand from hyperscalers now, but many random neoclouds have spare capacity

claude haiku is the boost mobile of LLMs

what a time to be alive