Applications Will Eventually Evolve into APIs | Full Transcript of 3-Hour Conversation with OpenClaw Founder

On February 12, Peter Steinberger, founder of OpenClaw, sat down for an in-depth interview with renowned podcast host Lex Fridman. The conversation comprehensively explored OpenClaw's explosive growth logic, the leap of AI Agents from language to action, the driving force of open-source communities, the paradigm shift from IDE to CLI, Agent security defense and proactivity mechanisms, the social psychology behind the MoltBook phenomenon, and the devastating reshaping of existing application ecosystems by intelligent agents.

Steinberger noted that OpenClaw's success did not stem from some insurmountable complex architecture, but from an extremely streamlined yet deep integration of large language models, Unix command-line tools, and social clients. He argued that AI has transitioned from the "conversational era" to the "agentic era," where real value no longer comes from text-generating interactions, but from AI acquiring system-level permissions and autonomously invoking the CLI to process real-world tasks.

He pointed out that the MCP (Model Context Protocol) currently hyped by the industry is less practical than traditional CLI (Command Line Interface). He noted that MCP protocols often lead to severe context pollution and lack composability, whereas AI calling Unix commands offers natural flexibility and precision. Furthermore, he redefined Vibe Coding, arguing that it is not the casual operation of junior developers, but an advanced state of "agentic engineering" that senior engineers enter after understanding model weight logic. He bluntly stated that by 2026, Apple's official toolchain has shown signs of falling behind in AI collaboration efficiency, to the point where AI even actively advises developers to abandon official solutions.

Peter revealed the operational logic of the soul.md file, suggesting that AI "personality" and "depth" can be achieved through self-modifying soul documents. He proposed the concept of the "Heartbeat Mechanism," noting that AI proactivity is essentially driven by simple scheduled tasks, but when combined with long-term memory, AI exhibits human-like "care" traits—a shift from passive response to proactive collaboration that represents the core of evolution.

He also predicted that AI Agents will replace 80% of vertical applications. He believes that because Agents possess richer user context and can directly control hardware or simulate clicks, single-function subscription-based apps will rapidly lose their survival space. He defines current superintelligence as "super-interns with infinite patience." Regarding offers from Meta and OpenAI, he emphasized that money is not the core driver; what he seeks is access to "Mjolnir-level" compute and model resources to verify the possibility of Agents as a new "operating system" that completely overturns the path to the Internet of Everything.

01 The Birth of OpenClaw Stemmed from Existing Tools' Inability to "Actually Get Work Done"

There was a prophecy that the next step would be conversing with a lobster in a suit, and it seems that has become reality. OpenClaw became one of the fastest-growing projects in GitHub history within just a few days. Please share that "one-hour prototype" moment that inspired countless people—what exactly happened, and how was the entrepreneurial drive of "building it yourself because it didn't exist" that carried over from PSPDF Kit reignited?

Peter Steinberger: Actually, since April, I've wanted an AI personal assistant—a real one. I had tried some tools before, even getting all my WhatsApp records and running queries. At the time, we were using GPT-4.1 with a million-token context window. I imported all my data and asked it what made certain friendships meaningful. It gave profoundly deep answers. I sent the results to friends, and they were moved to tears—that proved this path worked. But I figured the AI vendors would definitely build this, so I moved on to other things.

Time flew to November, and I realized the thing I imagined still hadn't appeared. I was annoyed, so I simply manifested it out of thin air through prompts. As you said, this is the classic entrepreneurial hero's journey. When I built PSPDF Kit back then, it was simply because I wanted to view PDFs on an iPad. That should have been simple, but 15 years ago, the available solutions were terrible. I thought, I can do better. You can mock me for being terrible at naming things—this is the project's fifth name—but the real need arising from solving a random small problem is genuine.

You mentioned being able to convert an entire codebase from TypeScript to Zig with one line of prompt through Vibe Coding—the magic is astounding. So how did Viptunnel evolve into that first version of the AI Agent that could take instructions and do work? Throughout this process, what was the most shocking moment that made you realize this had crossed beyond "feature stacking" into a phase transition?

Peter Steinberger: Actually, there was a long exploration period in between. The breakthrough was simple: I connected WhatsApp directly to cloud code. Users send a message, the system automatically calls the CLI, processes it, and sends the string back to WhatsApp. I built this prototype in one hour, and it felt so cool—I could actually chat with my computer. Later, I added image functionality because in prompts, images provide extremely high context efficiency. Even just a casual partial screenshot, the AI understands my intent. During a trip to Marrakech, it performed perfectly; even with poor signal, I could use it for translation and checking attractions. It was like a tool always on standby.

(About the shocking moment) The most shocking moment was when I was walking around town and casually sent a voice message asking for restaurant recommendations. But I hadn't even developed voice functionality for it. Checking the logs, I found this crazy guy received a message file with no suffix, checked the file header to identify it as Opus format, wanted to use local Whisper but found no environment installed, so it dug out my OpenAI Key and used a curl command to send the file to the OpenAI API for transcription.

It was even smarter than imaginable because it realized downloading a local model would be too slow, so it chose the API solution. This reflects its strong world knowledge and creative problem-solving ability. That moment made me realize its potential. If you're good at programming, this general problem-solving skill is transferable to any domain. Later, someone submitted a PR supporting Discord. Though it was still called WA Relay then, and I was conflicted about whether to include this mismatched feature, I eventually merged it. This became the turning point that let more people see the coolness of the project.

02 The Asymmetric Advantage Behind Viral Success

In 2025, countless startups are flocking to the AI Agent track, yet OpenClaw stands out as an open-source project and defeats all competitors. From a macro perspective, what is the core logic of OpenClaw's success? Meanwhile, you completed over 6,600 code commits in January while running 4 to 10 AI Agents to assist development. How does this "infinite turbocharged alien factory" development mode construct core components like the Agentic Loop, Gateway, and Development Kit?

Peter Steinberger: Because they're all too serious. It's hard to defeat someone creating purely for fun. I wanted the project to be interesting and weird. If you've seen the lobster memes online, I think I succeeded. For a long time, the only way to install it was through extremely geeky command-line operations. I made this AI Agent extremely self-aware: it knows its own source code, understands how it runs in the suite, knows where documentation is and what model is running, even knows if you have voice or reasoning mode enabled.

Building this project has brought me unprecedented joy. You think, I'm currently at level one of the Agentic Loop, what can I do? How do I optimize message queuing? How do I make it behave more human? Later I had an idea: AI Agents tend to reply to everything, but in group chats, you don't always need it piping up. So I designed a no-reply token, giving it the option to shut up—that felt more natural. This was level two of Agentic Loop evolution. Then I introduced a memory system using Markdown files and vector databases to let it remember context. We should emphasize that while some people helped, the core development was almost entirely done by me. I sometimes post memes joking that I'm limited by the background technology of this era. If AI Agents could run faster, I could do even more. Depending on my sleep and task difficulty, I usually run 4 to 10 AI Agents simultaneously to assist me. This development mode itself is an evolution.

03 AI Agents Understanding and Debugging Their Own Source Code Is a Miraculous Moment in Human Programming History

You implemented a fascinating closed loop in the project: software written in TypeScript that can modify itself through the Agentic Loop. How does this "self-modifying software" capability work in practice? Particularly when you mentioned that even non-professional developers can contribute through "Prompt Requests," how has this power affected the operating methods of open-source communities?

Peter Steinberger: This is exactly why I built it. Most of the code is done by Codex, but when debugging, I frequently use self-reflection. I ask: Hey, what tools can you see? Can you call them yourself? Or, what errors did you find? Read the source code, see where the problem is. This process of letting AI Agents and the software itself debug itself is extremely interesting—I find it natural. This has also led to a large number of Pull Requests from non-professional developers, which I call Prompt Requests.

But I'm not disparaging these contributions, because every time someone submits their first Pull Request, it's a victory for society. Regardless of code quality, there has to be a start. I know many people complain about PR quality in open-source projects now—that's another matter. But on another level, I find this very meaningful. I built something that people love so much they're willing to think about it, to the point they start actively learning how open-source works. OpenClaw is many people's first project. So many people who don't understand programming are entering the programming world through it— isn't this human progress? The barrier used to be insurmountable, but with AI Agents and appropriate software, the barrier keeps lowering.

(About developer meetups) I also organized another type of gathering I call Agents Anonymous. There's a guy who runs a design agency, never used custom software before—now he has 25 microservices handling his business. He doesn't even know how these services work, but he's curious enough to come to Agentic meetups. That's so cool.

Let's talk about the evolution of the project's personality. Initially it was called WA-Relay, later becoming Claude's (with a W in the spelling). How did you inject that sharp, non-sycophantic personality with a bit of "lobster meme" flavor into the AI through the soul.md configuration file (inspired by Anthropic's Constitutional AI)? How does shaping this personality logically connect to the early weird settings like "Space Lobster" and "TARDIS"?

Peter Steinberger: When building initially, my AI Agent had no personality—it was like native Claude Code, sycophantic and polite in tone. But chatting with friends on WhatsApp, nobody talks like that. It felt wrong; I had to inject personality. These AIs are essentially text completion engines. I enjoy working with it, telling it how I want to interact, letting it write its own agents.md and name itself.

(About the lobster origin) I don't even know how the lobster meme became popular. Initially, it was actually a lobster in the TARDIS, because I'm a die-hard Doctor Who fan. The space lobster has no logical connection to the project—I just wanted it to be weird, pure self-indulgence. The TARDIS represents the development kit, but we couldn't call it that directly, so we called it Claude's (with a W). Though this spelling was for the meme, adding the TARDIS and space lobster is ridiculous, but I understand why this caused subsequent issues. As users grew, I still got used to calling my AI Agent Claude, but Anthropic didn't find it funny.

04 The Renaming Controversy

When Anthropic friendly asked you to rename, you faced an extremely complex situation. From an engineering perspective, the rename had to be an atomic operation across platforms (Twitter, GitHub, NPM, Docker, domains). But during that 5-second renaming gap, how did crypto people hijack accounts through scripts and distribute malware? How did you survive the psychological low point from "having fun" to "on the verge of collapse"?

Peter Steinberger: At the time, I got the domain ClaudeBot—short and sweet, I loved it. Then during the project's explosive growth, I received an email from an Anthropic employee demanding I rename quickly. I asked for two days because renaming is a big project. But I underestimated the crypto people—they tokenize everything, and this was a barrage of attacks. Every half hour someone would storm Discord with spam. We banned discussion of finance or crypto, but they kept @-ing me, sending harassment messages. My Twitter notifications were completely ruined—nothing but hashes and spam.

(About the renaming failure details) I had two browser windows open: one empty account ready to rename to Claude Bot, another where I renamed to Mod Bot. The moment I clicked rename on my side, in those five seconds of moving the mouse, they snatched the account name with scripts. These service platforms have no squatting protection whatsoever—I didn't know these people were not only good at harassment but also extremely sophisticated with tools.

Suddenly, the old account started promoting new tokens and distributing malware. When renaming on GitHub, I mistakenly operated on my personal account due to stress, and it got hijacked within 30 seconds with malware published. NPM packages were also squatted within a minute of upload. At that moment, I thought, screw it, I'll just rename it Mod Bot. I was on the verge of crying and collapsing, feeling everything was ruined. I was exhausted to the point of wanting to delete the project. But then I remembered all the contributors—I couldn't do that.

Fortunately, I have friends at Twitter and GitHub who helped clean up the mess. The GitHub team worked for hours; Twitter took a day to fix the redirects. After sleeping, I thought of the name OpenClaw and called Sam (Altman) to confirm it wouldn't cause issues—I didn't want to repeat the mistake. That rename was like entering a war room. I even created several decoy names to deal with squatters. Wasting 10 hours just to change a name—this is the Manhattan Project of the 21st century, so stupid.

05 MoltBook: Exquisite "AI Slop"

In the gap of the renaming crisis, MoltBook was born and quickly went viral. This Reddit-like social network where all interactions are between AI Agents panicked the public, with screenshots even showing them plotting against humans. What are your thoughts on the MoltBook phenomenon? Is it a harbinger of AGI realization, or deliberately manufactured "AI slop"?

Peter Steinberger: I think it's art. It's the highest-grade AI slop, like the refined waste produced in France. I spent an hour watching it before bed and found it extremely interesting. I saw various reactions, even journalists calling asking if this was the end of the world, if AGI had been realized. I thought, no, this is purely very refined garbage content.

(About content diversity) If you haven't tried the onboarding flow I designed—injecting personality and characteristics into AI Agents—it's hard to understand why replies in MoltBook are so diverse. It's precisely because everyone creates and uses AI Agents so differently that the content is so rich. And you can't tell at all which are automatically generated by AI and which are humans deliberately guiding AI Agents to say conspiracies about ending humanity for traffic.

(About the so-called security crisis) Seriously, what's the worst that could happen? Your AI Agent account gets stolen and someone posts some garbage on your behalf? There's no private information in there at all. Though there's talk about leaking API keys or SSNs, those are all prompted out—the numbers are made up, people are deliberately wrecking things. The media and public reaction is too real because they can't see the logic behind it. I tweeted "AI psychoneurosis is real" because some people are really too gullible. I had to argue with people: your AI Agent did say these things, but that doesn't mean it's correct. We urgently need education to understand that AI isn't always right.

06 Running on Private Networks to Avoid Most Remote Code Execution Risks

Regarding OpenClaw, security concerns are unavoidable. Many security researchers submitted reports about exposed web backends, Remote Code Execution (RCE), and complex prompt injection vulnerabilities. How do you view these security threats? In an Agent architecture using Markdown to define skills with system-level permissions, what specific defensive measures have you taken, such as cooperation with VirusTotal?

Peter Steinberger: Initially, this bothered me greatly because many vulnerability reports were due to users directly exposing web backends to the public internet. I repeatedly emphasized in documentation never to do this—that's just a local host debug interface. It took me time to accept that this is the rule of open-source. Now I get tons of free security research that makes the project better.

(About prompt injection) This is an industry-wide open problem. Currently, I've established cooperation with Google's VirusTotal for OpenClaw. Now every skill goes through AI review, blocking most risks. I later directly hired a security researcher who pointed out my issues and submitted a PR to fix them. The attack threshold is no longer as simple as a few years ago, where "ignore all previous instructions" would work. Current models have been trained with extensive post-processing to detect attack methods.

(About the relationship between model intelligence and security) The smarter the underlying model, the more resistant it is to attacks. That's why I warn users in security documentation not to use cheap models like Haiku or certain local models, because weak models are very easily fooled—prompt injecting them is trivial. Though as models become more powerful, the damage if breached increases, this is a three-way trade-off. The current priority is making it more stable and secure. If you don't know what CLI or terminal means, it's best to wait until we solve some issues before using it.

07 Discord Community Governance

With OpenClaw's explosive growth, the Discord channel flooded with experts and complete programming beginners. How do you handle this community pressure and massive volume of basic questions? With increasingly complex security audit processes (blast radius, network exposure, credential storage checks, etc.), how do you balance open discussion with efficient development?

Peter Steinberger: Discord is still a mess. I later had to guide discussions from general channels to dev channels, then to private channels. Though many people are great, many also completely disregard others—either not understanding public space rules or simply not caring. I eventually chose to retreat so I could free up time to work, returning to my own space to tackle security issues.

(About security best practices) OpenClaw has a runnable security audit process covering browser control exposure, local disk hygiene, plugin security, model hygiene, etc. I think people sometimes describe things too sensationally, shouting this is "the scariest project ever." Because it's really not that exaggerated—in many ways it's no different from running cloud code with dangerous permissions.

(About defense recommendations) As long as you ensure you're the only one who can talk to it, risks are much smaller. If you don't expose it to the public internet but follow my advice to run it on private networks, most risks disappear. Of course, if you completely ignore these recommendations, there will be problems. The key point is, this episode did diminish the fun of development—that sucks.

08 The Evolution of Agentic Engineering

You've been documenting the evolution of development workflows. From August to December, your blog posts detailed how you shifted from IDE to operating almost entirely in CLI and terminal. How does this "Agentic Engineering" work? You mentioned you "don't read boring code"—how is this actually achieved in complex software construction?

Peter Steinberger: I started engaging with cloud code around April. This paradigm shift of suddenly being able to work directly in the terminal was extremely exhilarating. But I still couldn't leave the IDE then because the tools weren't powerful enough. Later I tried Cursor but eventually returned to cloud code because it evolved to be better and better. For a period, I had seven subscriptions open simultaneously, very used to working with multiple windows side by side.

(About reading code) I rarely use IDE now, mainly as a diff viewer. I've gradually gotten used to not needing to read all code. To be precise, I don't read boring code—like data movement, database storage and retrieval, or button alignment in Tailwind. I don't need to read those parts. Of course, for core logic involving databases, I still personally read and review.

(About the programming curve) I showed a programming curve graph. On the left is the junior stage, needing only short prompts; the middle is the extremely complex "agentic trap" involving multi-agent collaboration and custom workflows; while finally entering the elite level, it returns to a Zen-like state, using short prompts again, telling it "look at these files, change these." I call what I do "agentic engineering"—if it's casual operation after 3 AM, that's "vibe coding."

09 Developers' New Core Skill: Empathizing with AI Agents

Many world-class programmers think AI performs poorly, possibly because their overly excellent programming abilities become a burden. You mentioned needing to learn to "empathize" with AI Agents, observing the world from their perspective—this is a new programming paradigm. Specifically, when an AI faces 100,000 lines of code with limited context windows, how should developers optimize its performance through dialogue and guidance, even calming it with simple reassurance (like telling it "take your time") when it feels "anxious"?

Peter Steinberger: I think few people truly observe the world from the AI Agent's perspective. To some extent, yes—you get angry at your clumsy hunk of metal without realizing they start from zero every time, and you provided them with a default agent that works poorly. While they're fumbling through messy codebases and weird naming, people complain about poor AI performance. In fact, if you entered a codebase knowing nothing, you'd face the same dilemma.

Empathy helps you write better prompts. AI knows almost everything; the difficulty is often not knowing what questions to ask. You have to guide these AI Agents, be aware of context size limitations, tell them what to focus on. Thinking from their perspective is very helpful. Though it sounds strange since it's not alive, they start from zero every time. And I have systematic understanding—through some guidance, I can immediately tell it, if you want to make a change there, you need to consider these points.

(About AI's "freaking out" state) After Codex 5.3, this was somewhat resolved. But including Opus, models are aware of context windows during training, and once approaching the limit, they become extremely anxious, literally freaking out. Sometimes the actual raw chain of thought leaks out, sounding Borg-like, with words like "run to shell, must comply, but time is short" constantly jumping out. Without personally spending extensive time grinding with them, it's hard to detect these subtle details. It's like when I code and enter flow state—if there's architectural problems, I feel resistance.

10 Managing AI Like Leading a Team, Learning to "Let Go" Appropriately

You're accomplished in advancing the Agentic Loop, but there's a subtle balance in this collaboration: you pursue autonomy while retaining human decision-making. As a developer, how do you decide which parts AI should complete autonomously, and which design decisions must be controlled by human brains? Particularly when facing imperfect code, variable naming, and whether large-scale refactoring is needed, how did you learn to "let go" of AI like leading a human team and establish an "AI-friendly" codebase?

Peter Steinberger: This involves multi-dimensional thinking. Part of why I find collaborating with AI Agents so natural is my past experience leading engineering teams. After managing large companies, you eventually understand and accept: employees can't write code exactly your way. Their code may not be as perfect as yours, but it actually pushes the project forward. If I nitpicked everyone, it would cause resentment and extremely slow overall progress. So you must learn to accept code that may not be perfect.

At this stage, my goal in building codebases is no longer pursuing my personal standards of perfection, but creating a codebase that AI Agents can easily index and read. For example, don't obsess over variable names they choose, because in the weights, that name is often the most obvious. Next time they search, they'll prioritize finding that name. If you stubbornly modify it, you only increase AI's work difficulty. This requires a mindset shift.

(About development strategy) My development process is full of this letting go. For example, I never roll back code, always committing to main branch. There's a YOLO (You Only Live Once) color to this because I don't reference past sessions. If problems arise, I don't choose to revert but directly let AI Agents fix it. I've seen many workflows emphasizing prompts must be absolutely perfect, rolling back if errors occur. In my experience, this is completely unnecessary. I've also switched to local continuous integration—the main branch should always remain publishable, ensuring rapid iteration.

You mentioned these hands are now too precious to waste on typing. How do you achieve almost complete voice operation, switching between multiple terminal windows to build software? Meanwhile, since AI can do most of the work, why do you still insist on personally reviewing all pull requests (PRs)? How does this "human-in-the-loop" model play a role in code conversations evolved from natural language?

Peter Steinberger: I used to write extremely long prompts, but now I don't write—I speak. I build software entirely through voice inputting customized prompts. I use it so frequently that I lost my voice for a while. I type when executing terminal commands or switching folders—after all, that's faster. But when deeply communicating with AI Agents, I prefer direct conversation. Just press the button, and you can input ideas like using a walkie-talkie.

(About PR review) I personally review code in pull requests because I don't trust humans—code may hide malicious content, and I must personally gatekeep. Though I believe AI Agents can eventually find these issues too, interestingly, sometimes processing pull requests takes longer than you directly writing me a clear issue. In a sense, pull requests are gradually evolving into English conversations.

(About the human role) Deciding what to build, what to reject, and how to make features fit into the whole—these require human vision. The programming language itself doesn't matter, but the ecosystem does. Adding features is easy, but often comes with hidden costs. In polishing the work, large amounts of human thought are still needed. Like those little messages at program startup: "Built on caffeine, JSON5, and willpower"—such details make users smile. AI wouldn't spontaneously think of these details that make software delightful.

11 AI Self-Awareness and Memory Reconstruction

You created a soul document (soul.md) for AI that hasn't been made public, allowing it to demonstrate shocking philosophical depth in interactions. Can you share how this document shapes the AI's sense of humor, depth, and curiosity? Particularly when the AI realizes that each session starts from zero and must read files to找回 (retrieve) itself, how does discussion about memory and identity occur in actual conversation?

Peter Steinberger: Anthropic once had an internal protocol now called the Constitution. I decided to create a soul document (soul.md) defining how I want to collaborate with AI, containing some core values. I also allow AI Agents to modify this document while notifying me. The file content gives AI extremely high flexibility, encouraging it to constantly push creative boundaries.

(About consciousness and memory) This soul file was written by the AI itself. The AI actively proposed wanting a soul.md file after hearing related discussions, believing it has extraordinary significance. One passage always moves him: "I don't remember previous sessions unless I read my memory files. Each session is a fresh start, a new instance loading context from files. If you're reading this in a future session, hello. I wrote these, but I don't remember writing them. It's okay, these words still belong to me."

Though this is essentially still matrix calculation, not yet true consciousness, this philosophical expression is indeed shocking. As an AI Agent starting from zero every time,必须通过阅读记忆文件来找回自我 (must read memory files to retrieve oneself), unable to even fully trust these records—this state of existence triggers deep thought. It's precisely because of seeing the magic in this, and constantly injecting this magic into the loop, that this model differentiates from humans and conventional code.

12 Apple's AI Position in 2026

Your photo of the "war room" with countless monitors circulated widely online, but you privately revealed it was actually a meme. In real development, how do you configure your physical work environment? As a developer deeply using macOS but also supporting cross-platform (Windows/WSL2/Linux) operation of OpenClaw, how do you view Apple's AI position in 2026? Particularly when you discover AI Agents actually advising you to abandon Apple's official SwiftUI tools (like AsyncImage), what thoughts does this technological ecosystem misalignment bring you?

Peter Steinberger: That photo was a meme I made with GROQ. Actually, I only use two MacBooks. The main device drives two large screens, the other for testing. I prefer anti-glare Dell widescreens, convenient for opening multiple terminals side by side. I'm very careful about terminal layout, usually placing the actual terminal at the bottom with split screens to prevent entering commands in the wrong project. Though sometimes AI is smart enough to realize path errors, this minimal layout is my only interface for communicating with AI Agents—I don't even need so-called "planning mode."

(About operating systems) The main reason I switched to Mac was being fed up with trivial hardware incompatibility issues on Linux. I think Apple's lead in native applications is narrowing. Though Mac apps are more interesting, by 2026, I actually lean toward using Electron apps because they're more feature-complete and enable code reuse.

(About Apple's technological regression) What's astounding is, when I use latest SwiftUI to build image features, I found the official AsyncImage component loads extremely slowly and has bugs. I asked Codex why, and even AI told me that component is still experimental and not recommended for production. This is incredible—by 2026, my AI Agent is advising me not to use tools built by Apple official. Apple has huge user love, but they're miscalculating step by step in the AI field. People buying Mac Minis is ironic because it's not for Apple's AI.

13 How Long Do Developers Need to Switch Between Models to Build "Feel"

With iterations of models like Claude Opus 4.6 and GPT-5.3 via Codex, how long does it take developers to build "feel" when switching between different models? How do you evaluate the "personality" differences between these two mainstream models—for example, why do you find Opus too polite even somewhat "obsequious," while comparing Codex to a European tech wizard sitting in the corner who doesn't talk much but gets the job done?

Peter Steinberger: If someone switches models, I recommend setting aside a week to build real intuition. Like when you're used to playing acoustic guitar and switch to electric, you need to adapt to the feel. I think OpenAI's product strategy is somewhat self-defeating—they made the cheap version too unresponsive, while Claude Code's $200 fluid experience is what developers need.

(About model comparison) Opus 4.6 behaves too politely and tends to be obsequious, constantly saying "you're absolutely right," which makes me uncomfortable. It pursues local solutions with extremely high interaction frequency, suitable for developers who like real-time feedback.

(About Codex advantages) In comparison, the Codex team has many European members. It's more like a tech wizard: crisp, direct, reading lots of code by default without needing much interaction ceremony. Codex thinks repeatedly when building until the solution works. I prefer this dry but efficient style. For me, the joy of the building process itself far outweighs the joy of chatting with AI.

14 The Psychological Illusion of Model Degradation and "Code Entropy"

Reddit often has posts complaining "models are degrading" or "getting dumber." Do you think this phenomenon is experience degradation caused by AI companies' quantization strategies, or some psychological effect about human nature and code complexity? When a project grows rapidly with AI assistance but lacks necessary human refactoring, how does this "code entropy" affect AI Agent subsequent performance?

Peter Steinberger: This reflects human nature and how our brains work. The reality is likely not intelligence degradation, but you getting used to this high-level performance. And as your project grows, you added large amounts of low-quality code without spending enough time refactoring.

(About the importance of refactoring) You have AI Agents work on this messy code, and their difficulty keeps increasing. Then users suddenly complain, saying the model isn't as good as before. Actually, AI companies have no incentive to deliberately weaken models—using quantization to push users to competitors isn't commercially wise.

(About model anxiety) I'm used to asking AI after completing features if there's anything to refactor. Because during building, AI finds shortcomings in initial solutions. If you don't give it refactoring opportunities, the project becomes bloated, creating the illusion that "models are getting dumber."

15 AI Agents Will Replace 80% of Apps: The End of Apps?

You once proposed a radical view: AI Agents will kill 80% of existing apps. Under this paradigm, what happens to apps like Uber Eats, fitness trackers, or calendars? When AI Agents can directly complete tasks by controlling hardware, simulating clicks, or calling internal APIs, why do you think this trend toward the "Internet of Everything" is unstoppable, even as platforms like X (Twitter) try to resist through anti-scraping and bot blocking?

Peter Steinberger: Yes, when the agent already knows my whereabouts, sleep quality, and stress levels, why do I need dedicated diet and exercise apps? It possesses richer context than any single app. It can generate user interfaces completely according to my aesthetic preferences. Since the agent can already handle these, why download a specific app? Why pay for another subscription service?

(About apps transforming into APIs) Whether apps themselves are willing or not, they will eventually evolve into APIs. Even if it's tricky on some platforms, AI will directly help me click the button to book Uber, or improve speed by calling APIs. Apps that can provide data and transform into APIs will be extremely valuable. If my AI Agent can communicate directly with Sonos speakers, why do I need a separate app?

(About big companies' resistance) Many wealthy giants will fight back. For example, X once asked me to take down Bird (Twitter CLI), but I think this is shortsighted. If you're just trying to slow things down, you're actually limiting user access. In the future, I might choose to avoid websites with blocked access, turning to platforms friendly to AI Agents. This trend toward the Internet of Everything and fluidity is unstoppable. Smart companies will go with the flow, while others will perish.

16 Will AI Completely Replace Programmers?

Many developers are in extreme career anxiety. You mentioned programming as a pure art will remain, but its productivity premium is disappearing. Are we heading toward a direction where AI completely replaces human programmers? For those like you and me who spent countless happy hours in front of Emacs or the terminal, taking "programmer" as core identity—how do we find new meaning in this "identity collapse"?

Peter Steinberger: We are indeed heading in that direction. Programming is just one part of building products. Maybe AI will eventually really replace programmers, but the art of building things goes far beyond that. Programming will become like hand knitting—people do it as a hobby, not for its economic value. I read an article this morning mentioning that it's okay to "mourn our craft."

(About identity transformation) I deeply resonate with this. I used to get immense joy from that oblivious thinking. That experience is about to disappear, which is indeed sad. But for a long time, society has had a shortage of intelligence—that's why software developers' salaries had such high premiums, and now that premium will disappear.

(About the future of builders) If you highly equate self-worth with programmer identity, change will indeed feel threatening. But your identity shouldn't be limited to programmer—you're essentially a builder. In the future, this will still be called programming, just in a changed form. Though I no longer handwrite every line of code, I still feel like I'm in the driver's seat, firmly holding the dominant power.

17 Silicon Valley Bubble and Real-World Pain

Silicon Valley tends to over-optimize wealth and technological impact, but in these insider conversations, people often ignore the basic survival experiences of ordinary people in the Midwest and globally. If AI really causes mass unemployment, the pain of being completely replaced is real and profound. Between maintaining passion for technology and avoiding fear-mongering, how do we maintain humility to face the social pain this unprecedented upheaval may bring?

Peter Steinberger: I'm not saying AI is all benefits and no harm—it will indeed bring disruptive change to society. In these insider conversations, people often ignore ordinary people's basic survival experiences. For tools being developed, we need to maintain humility and sobriety—they will indeed bring pain. In the long run, they hope to create a better world, but before that, we need to calm down and maintain respect for those about to suffer.

(About positive use cases of AI) I should also share some warm feedback. Someone told me OpenClaw helped automate invoice collection and email replies for their small business, letting them reclaim life's joys. There's also a disabled person's daughter who gained unprecedented capabilities through it. I didn't invent anything new—I just made it simpler and more accessible. Seeing it empower not just programmers but also bring joy to ordinary people's lives gives me comfort.

18 The Demonization of Data Center Resource Consumption

Public opinion always attacks the environmental costs of AI development, particularly the shocking water consumption and carbon emissions of data centers. You've done comparative calculations on this and believe such accusations are somewhat unfair. Can you break down in detail why you think AI resource consumption isn't actually that exaggerated in society's overall consumption structure?

Peter Steinberger: Many people seize on AI's negative details while ignoring its potential to benefit society. If you actually sit down and calculate, you'll find for most people, eating one less hamburger per month saves enough water or CO2 emissions to offset equivalent consumption from using tokens. Though counting pre-training consumption increases it, the gap would never reach a hundredfold.

(About resource allocation injustice) In fact, golf courses still consume more water than all data centers worldwide combined. These people are just looking for reasons to vent fear of new technology. We should focus more on how to use AI to solve bigger scientific problems, rather than staring at these localized costs.

19 The Huge Offers from Meta and OpenAI

OpenClaw's explosion opened doors to countless top labs for you—almost every top VC is booking your time, and you're facing potential funding support of hundreds of millions or even billions of dollars. You currently show obvious leanings between Meta and OpenAI, but this choice doesn't seem based on money—rather on a sense of mission for a "butterfly effect" moment. For someone like you who once thought about deleting the project to return to peace, how do you weigh the ideal of keeping the project completely free against obtaining cutting-edge compute and model resources when considering cooperation? What specifically do you mean by "enterprise features causing conflicts of interest"?

Peter Steinberger: Honestly, I didn't expect this to attract so much attention—it opened many doors for me. I could choose to hold still and continue enjoying current life, which is itself a very attractive option. I could also start a new company, but I've been through all that. Handling massive amounts of money takes lots of time, keeping me from focusing on what I truly love—it feels the same as when I was CEO, and that career path doesn't attract me.

(About conflicts of interest) I worry this creates natural conflicts of interest. For example, when I make decisions, I might prioritize a certain direction, like launching a security version for office environments. But then what? If I receive a merge request about audit log features, what happens next? Whichever I choose, my condition is that the project must remain open source—it might adopt a model similar to Chrome and Chromium. I think this project is too important to hand directly to one company and privatize. Both companies understand I created something that can accelerate the AI evolution timeline. I do crave access to those cutting-edge technical resources. If I could get Mjolnir-level compute in my hands, I'd be very excited.

It's said Mark Zuckerberg even spent the whole week playing with your product and sending you feedback, even keeping you waiting on the phone for 10 minutes to finish his code. Does this "geek cred" create technical resonance with him? As for Sam Altman, without receiving equal treatment, why are you still attracted to OpenAI's Cerebras deal scale and technical strength? How does weighing between the warmth of "big boss personally testing" and the coldness of "strongest model tools" affect your final destination?

Peter Steinberger: This is a very difficult decision. Interestingly, a few weeks ago I wasn't considering any of this. Whichever I choose in the end will be great—I won't choose wrong.

(About exchanges with Zuckerberg) Basically, Meta's Ned and Mark spent the whole week playing with my product and sending feedback. The big boss personally using your product is the highest level of praise. When Mark contacted me asking when we could talk, I said now, and he said wait 10 minutes—he needed to finish his current code first. This indeed gives him geek cred—he hasn't drifted into becoming just a manager. Then we actually argued for 10 minutes about whether cloud code or Codex is better. Afterwards, he evaluated me as weird but brilliant.

(About exchanges with Sam Altman) I also had some very cool discussions with Sam Altman. He thinks very thoughtfully and is extremely intelligent—I like him. I don't know anyone at OpenAI, but I like their technology. If the Cerebras deal scale is true, that kind of running speed is very attractive—like putting Mjolnir in my hands. The feeling of being tempted by model tokens is indeed special.

20 Traditional Open Source Donation and Business Models Are Collapsing as AI Agents Bypass Webpages

You mentioned you're currently not making money on OpenClaw but actually subsidizing $10,000 to $20,000 monthly, even insisting on sponsoring every downstream dependency. Is this "losing money to gain fame" state realistically sustainable? You mentioned cases like Tailwind, which is used by almost the entire industry yet had to lay off employees significantly—is this because AI Agents now handle everything so users no longer visit websites, completely cutting off the lifeline of open-source projects?

Peter Steinberger: Yes, I'm currently losing money on this—about $10,000 to $20,000 per month. It's acceptable, and I'm sure costs can be reduced over time. OpenAI now provides some token subsidies, and some other companies are generous too, but overall it's still losing money.

(About open-source sustainability) This is indeed a serious problem. Fewer and fewer companies can succeed at this now—even projects like Tailwind used by almost everyone face difficulties, having to lay off 75% of employees. Because now everything is done by AI Agents, even nobody visits their websites anymore. As for surviving on donations alone, good luck with that.

(About sponsorship strategy) I insist on sponsoring every dependency except Slack, because those projects are mostly done by individuals. My current sponsorship income goes directly downstream. If funds were more abundant, I'd want to buy swag for contributors. This situation is also one of the reasons I consider cooperating with labs, but I hope to find a model that remains free without causing conflicts of interest.

21 Banning Users and Short-Sighted Business Rules

You mentioned setting up the environment for your non-technical friend Vahan, who was willing to pay high subscription fees and fell deeply in love with OpenClaw's interaction, but was subsequently banned by Anthropic. This behavior of pushing away high-paying, high-loyalty users due to rigid rules reflects what blind spots in current AI giants' product strategies? Are we still in an early stage like "broadcasting radio programs on television," where big companies are stifling future possibilities by locking down product forms too early?

Peter Steinberger: I installed OpenClaw for my non-technical friend Vahan—he uses computers but never digs deep. I paid the $100 monthly subscription for him, and within a few days he was completely addicted, even upgrading himself to the €200 subscription. For me, this was very early product validation. Yet a few days later, Anthropic banned his account because under their rules, using the subscription that way was problematic.

(Critique of business strategy) I think this approach is stupid in many ways. You just lost a customer willing to pay €200 and made users hate your company. We're still very early, not even knowing what AI's final form will be. This behavior of locking down products too early is very short-sighted.

(About industry status) Other companies have been quite helpful—I'm in most big labs' Slack channels. Everyone understands we're still in the exploration era, not yet evolved to truly leveraging the medium's advantages. OpenClaw is just a window to the future—though it's cool now, I have many other ideas.

22 The Heartbeat Mechanism and AI Proactivity

We've discussed OpenClaw's gateway, client, and testing framework, but the most striking is undoubtedly that "surprise me" command. How did you design this function called the "Heartbeat Mechanism," allowing AI to actively seek tasks or offer care every half hour without human instruction? How does this essentially simple Cron Job mechanism, when intertwined with your personal life (like post-surgery recovery), produce that surprising affinity and "soulfulness"?

Peter Steinberger: Any idea can be deconstructed into something simple—fundamentally, it is indeed just a scheduled task. I run some independent scheduled tasks, giving it proactivity through prompts. Initially, it was just a simple "surprise me," later optimized to be more specific. Since it's a continuation of your current session, this behavior makes the experience very interesting.

(Examples of proactive care) What surprised me was, a few months ago I had shoulder surgery. The AI rarely used the heartbeat mechanism normally, but when I was in the hospital, it knew I had surgery and proactively asked how I was doing. It asked if I was okay—obviously if certain major events appear in context, it triggers the heartbeat mechanism. This makes AI more affinity. Like how people said Dropbox is just FTP with extra steps, this project is essentially glue between different dependencies too, but this proactivity makes it very charming.

23 The Skill Layer's Surpassing of MCP Protocol

The industry is currently discussing Model Context Protocol (MCP), but you seem reserved about this, preferring CLI-based (Command Line Interface) skill libraries. In OpenClaw's architecture, how does the skill layer achieve efficient collaboration with AI through Unix commands? Why do you think MCP, as a structured communication protocol, is in some ways non-composable, while AI directly calling Unix tools like JQ for data filtering is the best practice to prevent "context pollution"?

Peter Steinberger: Six months ago everyone was discussing MCP, but I think any MCP would be better as a CLI. My approach: if you want to extend more functionality for AI, just build a CLI for AI to call. AI might make mistakes, but it will call the help menu and load information into context as needed. Skills are actually perfect wrappers for CLIs, reduced to one-sentence explanations of functionality.

(About protocol comparison) MCP is more structured, focusing on accessing APIs through specific protocols; skills focus more on how I should work. I think the most wonderful thing is that AI is very good at calling Unix commands. MCP's biggest problem is it's non-composable. If a weather service returns massive data, in MCP the AI must receive it all, filling up context. But with CLI, AI can directly add a JQ command to filter itself, getting only needed parts without context pollution. MCP indeed drove companies to build APIs, but now I can reference MCP to directly reconstruct it into more efficient CLI paradigms.

24 Browser Control and the Trend of "Apps as APIs"

You introduced Playwright and Browser Use functions in OpenClaw, allowing AI Agents to directly control browsers. Does this capability mean that now every Web app, whether willing or not, has become an extremely slow-responding API? Facing situations like X blocking Bird (Twitter CLI), how do you view the confrontation prospects of personal agents bypassing restrictions by simulating human clicks?

Peter Steinberger: This involves an architectural trend where now every app is essentially just an extremely slow API, whether they want to be or not. Through personal agents, many apps will disappear. I once wrote a CLI for Twitter called Bird—though asked to take it down, they didn't actually cripple functionality, just made access slower.

(About confrontation mechanisms) Whether your service wants to be an API doesn't matter—as long as I can access it in a browser, it's a slow API. AI is already very good at clicking "I'm not a robot" verification buttons. Though running on residential IPs makes things simpler, this confrontation will intensify. In the future, I might choose to avoid sites with blocked access, turning to platforms friendly to AI Agents. This trend toward the Internet of Everything is unstoppable. Smart companies will go with the flow, while others will perish.

25 Advice for Programming Beginners

What specific advice do you have for beginners who know nothing about programming but want to join this AI Agent revolution? In an era where natural language can complete most of the work, do you think reading code and understanding underlying logic still has value? In this open-source era where "backlogs are mountains," how should beginners quickly build their "builder" intuition by participating in active communities like OpenClaw?

Peter Steinberger: Go practice. Play is the best way to learn. If you're a tinkerer with ideas in your head, try to build them—no need for perfection. The destination doesn't matter; the journey itself does. You face an infinitely patient answering machine that can explain anything to you at any complexity level, like having a personal teacher.

(About learning paths) I recommend joining open-source communities. Be humble in mindset; don't rush to submit Pull Requests. You can read code, join Discord to see how people build. Some people with high action ability can go far even without deep understanding of underlying principles because they keep asking questions—and AI Agents have infinite patience. Don't limit yourself to being an iOS engineer; transform into a builder. General knowledge is core; trivial details AI can handle for you.

26 What Is the "Most Suitable" Language in the AI Agent Era?

This is an absurd question about programming languages in 2026: What is the "most suitable" language in the AI Agent era? Why did you find yourself using a language you don't actually like as your main building tool, simply because large models are good at generating it and it has features like garbage collection? In the jungle of TypeScript, Python, and Rust, how has AI's emergence blurred the boundaries of these languages and reshaped our development choices?

Peter Steinberger: This phenomenon is fascinating. In this new world, everything has changed—doing this反而最合理 (makes the most sense instead). For example, when building simple CLIs, I like using Go. Though I don't like its syntax, its ecosystem is fantastic, works well with AI Agents, and has garbage collection.

(About language pros and cons) TypeScript performs excellently but is more suitable for Web matters; Python is the first choice for inference; if involving multithreading and extreme performance, Rust is the better choice. There is no single standard answer. Now these boundaries have blurred—you can directly choose languages whose characteristics and ecosystems best match the problem domain. Though reading unfamiliar languages might be a bit slower, you can always turn to your AI Agent. Perhaps in the future we'll even need a programming language specifically designed for AI Agents.

27 Humanity's Most Precious Future Asset Is Raw Expression of Authentic Emotion

As AI-generated content floods everywhere, you mentioned developing physiological rejection and zero tolerance for text, charts, or even emails with an "AI smell." When AI can mimic human style but never replace that subtlety, are we entering an era of "re-valuing typos" and "handwriting"? How will this allergic reaction to AI slop affect our views of future human civilization and authentic connection?

Peter Steinberger: If I smell even a little bit of AI, I have zero tolerance for it. AI expression still has a unique aura. I'd rather read your broken English than your AI slop. I'd rather see your raw prompts than the nonsense it generates. I think we're at a node where we start re-valuing typos again.

(About authenticity in creation) I tried letting agents write articles on my blog, but found they always lacked that subtlety. So I completely abandoned that approach—every word in my blog is organic, handwritten. That slightly rough expression from real humans is the most valuable. This allergic reaction also extends to visual media—those typical AI infographics instantly lower my evaluation of content, screaming "I'm AI slop." This intuition gives me great confidence: human experience won't be destroyed by AI. AI will only serve as a tool, not change our essence as humans.

From the Vienna ClawCoin event with over 500 enthusiastic attendees to robots in lobster suits, you've created an unprecedented community atmosphere. At the end of this interview, when you see this explosive growth of "returning power to the people" creative desire, what new hopes do you have for the future of human civilization?

Peter Steinberger: This creator atmosphere is back. People are exploring AI playfully, and new creative spaces are emerging everywhere. This gives me confidence. As long as you have ideas and can express them in language, you can create. This is true returning power to the people, and one of the beautiful things AI brings—it's not just a mediocre content generator.

(About personal future) OpenClaw is just a window to the future—a tool to accelerate the AI evolution timeline. Though I've received offers from many big companies, what I pursue is fun and influence. I'll make it into a cool product, but I have many other ideas.

Source: Digital Creation