Peter Steinberger on "Fun Is Velocity"

1. Key Themes

Annoyance as Product Discovery Engine

Peter's entire company was born from being annoyed that he couldn't check on his agents from his phone while raiding the kitchen. Every major product decision traces back to a personal friction point, not market research or VC thesis.

"My source of inspiration is usually being annoyed... there was still no good way to just send a prompt from my phone to my computer so some agent could check out how my agents are actually doing." 00:01:17

Personal Brand as the Only Unforkable Asset

Peter draws a sharp distinction between products (which can be cloned overnight) and personal reputation (which cannot). He explicitly frames this as the most important career investment a builder can make.

"Everything that you can build can be forked or cloned. But your name cannot. So your personal brand is way more important than any single product that you ever will. Start working on that before you need it." 00:12:33

Your Dependency's Business Model Is Your Business Model

OpenClaw was critically wounded not by competition but by Anthropic disabling Claude subscriptions with ~24 hours notice. The project had been over-optimized for a single model, and when the underlying platform changed its terms, there was no time to adapt.

"Your dependencies business model is your business model... when they pinged me with around 24 hours notice that they are going to disable the subscription for everyone, there was not really enough time to change course." 00:18:20

Fun as a Leading Indicator of Product Quality

Peter makes a causal claim — not a correlation — that the weeks he personally enjoyed building, the product visibly improved. The weeks it stopped being fun, the team shipped config options (i.e., bloat).

"Fun is velocity. The weeks I enjoyed building, the product got visibly better. The weeks I didn't, we shipped config options." 00:22:47

Feature Creep Is an Existential Risk in Open Source

With nearly 3,000 contributors, OpenClaw accumulated roughly 9,500 configuration permutations, making comprehensive testing mathematically impossible. The root cause was Peter's reluctance to say no to free labor.

"At our highest, I had to count that we ended up with around 9,500 configuration options if you count all the permutations. You can write all the tests you want. It is impossible to cover all of these and not break things from time to time." 00:16:26

The Agentic Shift: From Reactive Prompting to Proactive Orchestration

Peter describes a fundamental workflow evolution — from "asking an agent to do a task" to "designing a system where agents proactively bring you fully reviewed, tested pull requests." The bottleneck has moved from generation to review and decision.

"I try to make the agent the more proactive work for me. So when I shift my attention to something, I don't want to read issues. I want to see fully reviewed and tested PRs." 00:27:38

Open Source Requires a Written Vision Document

Without a stated vision, open source maintainers default to saying yes to every interesting PR. Peter now writes a vision.md for every new project as a forcing function for himself and contributors.

"These days I, when I do new open source, I write a vision.md file where I kind of explain what it is now, where I see this to be... it is always tempting to add this one feature that looks really cool. But what you rarely think about is when you merge this feature, it really means here's this pile of code that I don't need." 00:36:39

The "Hard and Boring" Category Is the Right Startup Target

In an era where anyone can prompt a fun product into existence in a weekend, the defensible category is the one that is hard and unglamorous — because it naturally filters out casual competitors.

"Maybe I would pick something again that's in the category hard and boring. Because that's usually a category that is a little bit easier to actually find people that will appreciate when you solve something. If you pick something that is fun, even if it's hard, you're going to have a very tough time, especially in a time where people can just prompt things into existence." 00:40:36

Perpetually Running Agents Are a Token Economics Problem, Not a Technology Problem

The vision of always-on, proactive agents is already technically possible. The blocker is cost per token, not capability — specifically the brutal economics of large context windows being re-sent after KV cache expiration.

"That's not so much a tech problem. It's more a token problem... my early system of like heartbeats was too static... one hour later, you call like a heartbeat checkup on everything. And that means you send 600,000 tokens back to the server after the KV cache cleared and you pay a stupid amount of money for a lot of not useful work." 00:37:04

2. Contrarian Perspectives

Security Researchers Are Often a Net Drain, Not a Net Benefit

Counter to the conventional wisdom that engaging bug reporters makes software safer, Peter argues that most security reports he received were AI-generated without manual testing, and the pressure they created cost him months of development time and strategic focus.

"I would be less stressed out about security researchers. They are really good at making you feel really bad... most of them really sent reports that their agent produced without actually even testing it. I would take a stronger sense of explaining what are the parts that we guarantee. And what are the parts that will not be fixed because that's not our security boundary." 00:33:46

Press Coverage of a Security Threat Is Irreversible — Corrections Are Worthless

When press reported 20% of OpenClaw skills were malicious (the actual audited number was 0.3%), Peter published a paper with real data. It didn't matter. He draws a structural conclusion about media dynamics that most founders underestimate.

"A correction never travels as far as a scare." 00:14:27

Stopping Reading Your Own Code Is a Legitimate Engineering Strategy

Peter explicitly does not read all the code in his own project and frames this not as negligence but as rational risk management — only scrutinizing changes that touch scary systems or take unexpectedly long.

"I see code review more as a risk management... Do I really care how the UI is built, if it looks correct? No... part of code review is simply like observing, looking how big it changes and trusting your gut." 00:31:25

Being Declared Dead Is a Better Marketing Event Than a Launch

OpenClaw peaked at 4.7 million weekly downloads in June — after being declared dead by the press — versus bottoming at 835,000 in May when it was fully operational. Controversy and obituaries drive more downloads than feature releases.

"We bottomed out at around 835,000 weekly downloads in May. And then after being declared death in June, we peaked at 4.7 million. The highest ever. Both of those are true at the same time. A hype is like the weather. You might see it coming, but you can't control it." 00:18:51

Asking VCs to Take Money After Going Viral Is the Wrong Move — Lab Partnerships Are Better

Despite being flooded with VC term sheets after OpenClaw went viral, Peter chose to partner with OpenAI directly rather than raise traditional venture capital, framing lab partnerships as a qualitatively different and more interesting path for certain founders.

"My inbox was full of VCs begging me to take their money... when the big labs came knocking, and suddenly I was on the phone with Mark, Sam, a few others, this seemed like a much more interesting path." 00:12:06

3. Companies Identified

OpenClaw

Open-source AI agent framework that lets users run Claude and other models locally across any device. Why mentioned: Peter's own creation — the central subject of the talk. Nearly 3,000 contributors, peaked at 4.7M weekly downloads, Jensen Huang called it "the most successful open source project in the history of humanity."

"Almost 3,000 people have commits in the repo... Jensen called it the most successful open source project in the history of humanity." 00:09:12

OpenAI

AI lab; Peter joined as an employee/partner after OpenClaw went viral. Why mentioned: Peter chose OpenAI over competing lab offers and VC money; he operates within their ecosystem.

"I was on the phone with Mark, Sam, a few others... my gut told me that I like OpenAI the most." 00:12:33

Anthropic

AI safety lab, creator of Claude models. Why mentioned: Critical dependency failure story — Anthropic disabled Claude subscriptions with ~24 hours notice, nearly killing OpenClaw at a pivotal moment.

"When they pinged me with around 24 hours notice that they are going to disable the subscription for everyone, there was not really enough time to change course." 00:17:51

NVIDIA

Semiconductor and AI infrastructure company. Why mentioned: Gave early, substantive support — specifically sent engineers to take over much of OpenClaw's security work, which was a turning point in stabilizing the project.

"I need to give NVIDIA a special shout out. Because they were very early and they simply asked me what I need. And then they sent people to take over much of the security work." 00:21:23

Microsoft

Enterprise software and cloud company. Why mentioned: Satya Nadella used "enterprise-grade claws" in a Microsoft keynote, validating that "claw" had become a generic noun in the industry.

"Satya says, enterprise-grade claws in Microsoft's keynote. There are 33,000 claw-named repositories." 00:24:13

4. People Identified

Jensen Huang

CEO of NVIDIA. Why mentioned: Publicly called OpenClaw "the most successful open source project in the history of humanity," a significant endorsement that shaped public perception of the project.

"Jensen called it the most successful open source project in the history of humanity." 00:08:42

Satya Nadella

CEO of Microsoft. Why mentioned: Used the phrase "enterprise-grade claws" in a Microsoft keynote, demonstrating that OpenClaw's terminology had penetrated mainstream enterprise vocabulary.

"Satya says, enterprise-grade claws in Microsoft's keynote." 00:24:13

Andrej Karpathy

AI researcher and former OpenAI/Tesla executive. Why mentioned: Peter credits Karpathy's thesis — "your agent, your machine, your life" — as the philosophical foundation for why OpenClaw exists as a privacy-first, local-first alternative to lab-hosted agents.

"That's not even my thesis, that's Gary's. We just shipped it first." 00:25:42

Boris

Appears to be a speaker at the same event (Lightcone conference). Why mentioned: Peter references Boris's talk from the previous day about letting models run.

"Yesterday Boris told you to let the model cook. I'm the guy who found out what happened when tens of thousands of people let it cook at the same time." 00:00:32

Mario

Contributor to the OpenClaw project. Why mentioned: Built a high-quality compaction implementation in Python that replaced Peter's hacky early version, enabling the Discord expansion.

"I added a very hacky version of compaction. So Mario could make a good one in Pi." 00:05:48

5. Operating Insights

Use Sub-Agents for QA by Having Them Stress-Test Each Other's Work

Peter describes a concrete multi-agent testing architecture: spin up an orchestrator (he used Codex), use 12 sub-agents to understand a project and decompose it into features, then have each sub-agent stress-test or code-review a specific feature and report findings back to the parent session.

"I spin up Codex. Use 12 sub-agents. Understand my project, break it down into features. And then each sub-agent would like stress test the feature or code review feature and then inform the other session where to focus testing on." 00:30:26

Gate Human Attention Behind Agent Pre-Validation of Ideas

Rather than letting teammates pitch feature ideas directly, Peter requires ideas to be prototyped by an agent first — with screenshots and a working demo — before any human review. This eliminates most bad ideas before they consume meeting time.

"If you come to me and you tell me this feature idea, I'm going to get mad at you. Like it's so easy that you just discuss the feature idea with an agent, you build it, you make screenshots, you let me play with it. That way, if it's good, we can like immediately enter in it. But most of the time people will figure out why it's not good and don't even come to me." 00:28:07

Give Agents Their Own Dedicated Machines to Avoid Human-Agent Resource Contention

Agents that use computer-use features will fight the human for mouse control and CPU. Peter's solution is to give agents their own always-on machines so work continues when his laptop is closed and there is no contention.

"I do a lot of Mac software, agents love to take over my screen and click around. And if you put that in their own, if you give them their own machine, it will not bother you. Otherwise you'll fight with the agent for the mouse cursor." 00:39:11

Use Time-as-a-Signal for Code Review Risk Assessment

Peter developed a heuristic: estimate how long a change should take, and only read closely when the actual time diverges significantly from expectation. This converts code review from a uniform tax into a targeted risk filter.

"You develop a little bit of a feeling how long something should take. So if I do a small tweak that should change how dragging works and it takes three hours, I know something's wrong. I'll look closely." 00:31:53

6. Overlooked Insights

Mac/Windows Sandboxed Test Environments for Agents Is an Unserved, High-Value Infrastructure Market

Peter throws out a product request in a single sentence at the very end: cloud-hosted, fast, cheap Mac OS and Windows test environments for agents. He notes Linux boxes are trivially available, but Mac and Windows sandboxes that actually work are "unreasonably hard." Given that the entire agentic software industry needs to run QA on non-Linux platforms and current tooling fails at this, the infrastructure gap is significant — and Peter, with direct influence over one of the largest agent ecosystems, is explicitly asking someone to build it.

"I feel it's very easy to have, to get a test box for Linux. It is unreasonably hard to get one for Mac that works really well. And all the stuff that I got for Windows is also quite annoying. I haven't found a really good provider yet that does all of that fast and cheap." 00:41:07

Compute Fleet Orchestration for Local Agents Is the Deepest Unsolved Infrastructure Problem

Buried in a question about bottlenecks, Peter describes a problem that nobody has cleanly solved: when multiple agent sessions run compute-intensive tasks simultaneously on local or semi-local machines, there is no good system to manage resource allocation, handle timeouts gracefully, or move workloads between machines. He is currently screen-sharing into multiple computers manually to distribute load — an absurd workaround that signals a genuine infrastructure vacuum at the intersection of agentic AI and compute orchestration.

"There's no really good system of like managing all of that... if it's something that needs other stuff that's on my computer, same. And we haven't really built something yet where things could easily move from here to here without issues... right now, I use too many systems. I screen share into some computers to like distribute the load. I shouldn't be doing that." 00:35:09