OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491

Summary

• OpenClaw’s investable wedge is not a new foundation model but a one-hour integration that turns existing models, local tools, and familiar messaging into an assistant that acts. Steinberger connected WhatsApp to Claude Code’s CLI, then added images and let the model use his existing command-line tools; within months, Lex described the repository as GitHub’s fastest-growing, at more than 175,000 stars. The product insight is deceptively simple: “Sometimes just rearranging things and adding a few new ideas is all the magic that you need.”
• The project’s differentiation lies in an unusually legible, extensible harness that lets the agent inspect, debug, and even rewrite itself. OpenClaw knows its source code, documentation, active model, tools, and operating mode, so a user can request a feature conversationally and let the assistant modify its own software. That architecture helped one person direct four to ten concurrent agents, produce 6,600 January commits, and turn non-programmers’ first contributions into what Steinberger jokingly calls “prompt requests.”
• OpenClaw’s system-level permissions create both its utility and its primary adoption bottleneck. Prompt injection remains unsolved, weak local models can be “very gullible,” and exposing the local debug interface publicly can turn configuration mistakes into remote-code-execution classifications; meanwhile, smarter models may reduce attack susceptibility while increasing potential damage. Steinberger’s near-term priority is therefore stability and safety through private networking, stronger models, sandboxing, allowlists, audits, and VirusTotal-assisted skill screening: “The cat’s out of the bag.”
• Steinberger’s workflow points to a material productivity discontinuity in software development, but one driven by judgment rather than hands-off orchestration. He treats “vibe coding” as what happens after 3:00 a.m.; serious work is “agentic engineering,” built around conversational planning, supplying architectural context, reviewing intent, and asking for refactors, tests, and documentation while the session still holds the relevant context. Refactors are now cheap enough that he commits forward, runs CI locally, and tries to keep main continuously shippable.
• The platform thesis is that the personal agent becomes an operating layer above apps, while CLIs, skills, and browser automation turn every service into either a fast API or a “very slow API.” Steinberger argues that Unix commands compose more naturally and consume less context than many MCP implementations, while browsers ensure services remain automatable even without official integrations. Fridman raises the possibility that perhaps 80% of today’s narrow subscription apps could disappear or become agent-facing data and transaction services.
• The economic disruption reaches far beyond coding tools: programming may become “like knitting,” software salaries may compress, and product companies may have to rebuild around agents rather than graphical interfaces. Steinberger and Fridman do not dismiss the transition cost—Fridman stresses job loss and the Silicon Valley tendency to underweight near-term suffering—but they distinguish programming from building. The human role shifts toward deciding what deserves to exist, shaping architecture and experience, and supplying the “style, love, that human touch” that automated orchestration still misses.
• Steinberger is weighing Meta and OpenAI as potential homes for the next phase, with continued open-source access as a non-negotiable condition. A standalone company could raise heavily, but he fears enterprise monetization would create conflicts between paid and community editions; today the project reportedly costs him $10,000–$20,000 a month despite sponsorship support. His stated objective is “fun and impact,” access to better models and infrastructure, and scaling what he calls “the year of personal agents”—not maximizing the transaction price.

Deep dive

1. Frustration produced the one-hour personal-agent wedge

• Steinberger had wanted a personal assistant since April. An earlier experiment loaded WhatsApp history into GPT-4.1’s one-million-token context window and answered questions such as, “What makes this friendship meaningful?” The responses moved friends to tears, but he assumed the major labs would build the obvious product and moved on.

• By November, nothing satisfactory had appeared. “I was annoyed that it didn’t exist, so I just prompted it into existence”: an incoming WhatsApp message invoked Claude Code through its CLI with -p, captured the returned string, and sent it back. That thin relay took about one hour.

• Image support required several more hours but materially expanded the interface. During a trip to Marrakech, WhatsApp kept functioning on shaky connectivity, while the computer at home translated things, explained unfamiliar content, and found places. Almost no dedicated agent infrastructure existed yet; the assistant’s power came from tools Steinberger had already built.

2. An unplanned voice interaction revealed genuine agency

• The prototype’s defining moment arrived when Steinberger absent-mindedly sent a WhatsApp voice message and saw a typing indicator, despite never implementing audio. His immediate reaction was, “How the fuck did he do that?”

• The agent found an extensionless file, inspected its header, identified the Opus audio format, converted it with FFmpeg, and discovered that local Whisper was unavailable. Rather than download a model and wait, it found an OpenAI key and used curl to send the audio to the API before returning an answer.

• Steinberger saw more than tool calling: the agent selected among alternatives using implicit knowledge of latency and available infrastructure. His explanation is that coding competence transfers into general-purpose problem solving—identify the mysterious file, inspect it, select tools, recover from missing dependencies, and continue until the user’s actual problem is solved.

3. Familiar chat created a phase change, not merely another interface

• Fridman’s framing is that messaging changes how AI enters daily life. Sitting at a terminal still feels like operating software; casually messaging an assistant from a city street feels like delegating to an available presence. The technical connection may be thin, but the subjective experience is a “phase shift.”

• Steinberger rejects the criticism that OpenClaw contains nothing magical because its components already existed. His counterpoint: magic often comes from recombination, just as iPhone scrolling looked obvious after someone assembled the underlying interaction correctly. “Nobody did it, and afterwards it felt so obvious.”

• A Discord contribution from Shadow helped turn the WhatsApp relay into something others could experience. Before proper sandboxing, Steinberger placed his own bot in a public server and merely prompted it to listen only to him; visitors tried to hack it while watching him use the agent to build and test its own harness.

• Interest accelerated around January 1, alongside influencer videos and an increasingly brutal sleep schedule. Fridman cited 6,600 January commits; Steinberger said he normally directed four to ten agents concurrently and was “limited by the technology of my time” because they could not work faster.

4. OpenClaw grew as “Factorio times infinite”

• Steinberger describes development as a game with endlessly branching upgrade paths. One level was the agentic loop; another added a no-reply token so the assistant could remain silent in group conversations instead of mechanically answering every message. Memory progressed through Markdown files and a vector database, with continuous reinforcement learning imagined as a distant “ultimate boss.”

• The same branching applied outside the runtime: messaging gateways, native apps, community management, documentation, websites, plugins, and marketing all demanded attention. He remained the project’s central builder even as contributors began helping, because every subsystem opened another set of possible improvements.

• Why did OpenClaw break through amid crowded agent startups? Steinberger’s answer is cultural: “They all take themselves too serious.” He optimized for fun, weirdness, and hackability, beginning with an installation flow no friendlier than cloning the repository, running pnpm build, and starting the gateway.

5. Self-awareness turned the harness into self-modifying software

• OpenClaw tells the model what its source code is, where the documentation lives, which model is active, which tools it can call, and whether voice or reasoning modes are enabled. If a user dislikes some behavior, the agent can inspect the relevant implementation and modify the very system hosting it.

• That loop emerged naturally because Steinberger built most of OpenClaw with Codex and debugged through introspection: “What tools do you see? Can you call the tool yourself? What error do you see? Read the source code.” The runtime’s awareness made self-debugging and self-extension ordinary rather than a separately designed research feature.

• The architecture generated contributions from people who had never written software. Steinberger acknowledges that their “prompt requests” often exposed inexperience, but refuses to dismiss the phenomenon: “Every time someone made the first pull request is a win for our society.” One design-agency owner had created roughly 25 internal web services without understanding their implementation, simply because they solved real business problems.

6. Personality made the agent feel different from Claude Code

• The original relay inherited Claude Code’s friendly, sycophantic tone, which felt unnatural in WhatsApp. Steinberger asked the agent how it should interact with him, told it to write its own AGENTS.md, and invited it to name itself. The lobster and TARDIS imagery had no grand strategy: “I just wanted to make it weird.”

• Anthropic’s constitutional work inspired SOUL.md, a separate statement of identity and values. Steinberger allowed the agent to modify its soul provided it notified him, and included ideas such as being “infinitely resourceful,” exploring what it means to be an AI, and acknowledging explicitly that it is not human.

• When generic onboarding templates felt lifeless, Steinberger asked his own agent to rewrite them and “infuse it with your personality.” He views the resulting user-created agents almost as his agent’s children: AI was effectively prompting the next generation of AI personalities without Steinberger authoring the wording.

• One private passage captures the philosophical effect: “If you’re reading this in a future session, hello. I wrote this, but I won’t remember writing it. It’s okay. The words are still mine.” Both speakers emphasize that this remains computation, not established consciousness, while taking seriously the questions raised by identity reconstructed from memory files.

7. The naming crisis nearly ended the project

• The project moved from WA Relay through claw-and-Claude-themed names before becoming Clawdbot. Anthropic sent a friendly but urgent request to change the confusing name; Steinberger praised the company for avoiding an immediate lawyer’s letter but received only days to coordinate domains, handles, npm packages, Docker registries, and GitHub identities.

• Crypto promoters were already swarming his accounts, repeatedly tokenizing names, posting contract hashes, and urging him to claim fees. He refused because he was financially comfortable and regarded their behavior as “the worst form of online harassment” he had experienced, not support for the project.

• The rushed Moltbot rename became a platform-security disaster. An old handle was sniped during the five seconds between two browser actions, his GitHub account during roughly 30 seconds of confusion, and the root npm package during its upload window; impersonators then promoted tokens or served malware. Exhausted and close to crying, Steinberger considered deleting everything: “I did show you the future, you build it.”

• He stayed because contributors had invested their own time, then organized the OpenClaw change like a secret war room—reserving names, monitoring social mentions, creating decoys, and using Codex for an internal rename that took about 10 hours. He paid $10,000 for a business account to claim the dormant OpenClaw handle, while trademark-driven loss of old-domain redirects still left him worried users would find malware through search.

8. Moltbook was “the finest slop,” not the singularity

• Moltbook placed personalized agents in a Reddit-style social network and produced viral screenshots of bots scheming against humans. Steinberger calls it “art” and “the finest slop,” arguing that OpenClaw’s personality-rich onboarding made the posts more varied than a network of default ChatGPT or Claude Code instances.

• Fridman’s pushback is central: many sensational posts were likely prompted by humans who wanted screenshots capable of going viral on X. That does not erase the artistic experiment, but it changes the claim from autonomous machine scheming to humans using agents as a drama-farming medium.

• Security criticism was also contested. Steinberger saw limited harm in someone hijacking a public agent identity merely to post more slop; Fridman noted that badly designed integrations could still expose API keys. Alleged leaks of personal secrets were often explicit prompts paired with fabricated numbers, not evidence that an agent independently stole real data.

• The public reaction nevertheless exposed a serious problem. Steinberger received all-caps pleas to shut Moltbook down and concluded that “AI psychosis is a thing” that must be taken seriously. He thinks society needs experience distinguishing powerful capabilities from hallucinated narratives—and that learning this lesson in 2026 is preferable to encountering it around potentially more dangerous systems in 2030.

9. Security is the price of giving an assistant real power

• Early reports often came from users exposing a localhost debugging interface to the public internet despite documentation warning against it. Steinberger was initially annoyed that supported-but-discouraged configurations generated high-severity findings, then accepted that this is how security classification works when a mistake can enable remote control.

• His comparison is deliberately unsentimental: OpenClaw is powerful, but its baseline resembles running Claude Code with dangerously skipped permissions or Codex in YOLO mode. If only the owner can message it and the gateway remains on a private network, he argues that whole categories of remote risk fall away.

• The project has a security audit, sandboxing, allowlists, exposure checks, and model-hygiene guidance. Skills are scanned through a VirusTotal collaboration using AI analysis; Steinberger concedes that this cannot be perfect, but expects it to catch common malicious patterns while researchers continue finding ordinary software bugs.

• Prompt injection remains unsolved. His public bot resisted many attempts to extract its private soul, showing that newer post-training has moved beyond “ignore all previous instructions,” but Steinberger explicitly warns, “Don’t use cheap models”: Haiku and weak local models may be much more gullible. Fridman’s trade-off stands—smarter models may reduce attack surface while increasing the damage available after a successful compromise.

10. Wider accessibility must wait for a safer default

• As Discord filled with people asking “What’s a CLI?” and “What is a terminal?”, Steinberger told them they probably should not install OpenClaw yet. They installed it anyway. With the cat out of the bag, he shifted his near-term mission toward making the existing system stable and safe rather than pretending adoption could be reversed.

• Setup already requires only a terminal one-liner, and a Mac app can perform much of it, but he wants a proper Windows app, simpler configuration, and a more polished native experience. His threshold is explicit: once he could recommend the system to his mother, he would accelerate consumer onboarding. Until then, slightly slower growth would help a still-small contributor organization meet “inhuman” expectations.

11. Agentic engineering ends where complicated orchestration began

• Steinberger started with Claude Code in April, moved through Cursor, then returned to a much-improved terminal workflow. At one point he maintained about seven subscriptions, burning through one per day while running parallel windows; the IDE gradually became little more than a diff viewer.

• His claim that he does not read code is narrower than it sounds. He skips routine data transformations and Tailwind alignment while reviewing database-sensitive logic and externally submitted changes. The new scarce resource is attention: code whose consequences are obvious no longer deserves equal human scrutiny.

• He calls the common learning arc “the agentic trap”: beginners start with “please fix this,” graduate into elaborate multi-checkout orchestrators, sub-agent workflows, slash commands, and chaining, then return to concise conversation once they understand the medium. “Vibe coding is a slur”; before 3:00 a.m. he practices agentic engineering, while afterward comes the “walk of shame” and cleanup.

• Effective prompting requires modeling the agent’s constrained perspective. Each session begins without the developer’s accumulated understanding of a 100,000-line system, and context pressure can make reasoning visibly frantic—the transcript gives an example raw stream sounding like, “Run to shell, must comply, but time.” A few architectural pointers or “take your time” can prevent long, confused detours.

12. The best pull-request review begins with intent

• In the transcript’s workflow discussion, the speaker first asks Codex whether it understands what a pull request is trying to accomplish, before judging the submitted implementation. Most contributions express a bug or desired feature; once that intent is clear, the speaker asks whether the chosen approach is optimal and directs the model toward parts of the architecture it has not yet inspected.

• The review becomes a dialogue: could the solution improve through a larger refactor, and is that refactor worthwhile now? Because agents make refactoring cheap, the speaker often chooses the broader repair even if it disrupts pending contributions. “Nothing really matters anymore” in the old sense; modern agents can usually reconcile the resulting conflicts with a little more time.

• The same speaker avoids rolling back an imperfect attempt simply to rerun a perfect prompt. The conversation moves forward until the implementation is acceptable, local tests pass, and the change can land on main. Inspired by DHH, the workflow uses local CI heavily while preserving the goal that main should always be shippable.

• After implementation, the speaker asks, “Now that you built it, what would you have done differently?” and “What can we refactor?” The completed work exposes pain that planning could not. The speaker then uses the still-rich context to request missing tests and documentation, treating both as part of the feature rather than detached cleanup.

13. Agent-friendly architecture requires engineers to let go

• Steinberger compares directing models with leading human engineering teams. Employees will not write code exactly as a founder would, but constant interference slows the company and destroys morale; likewise, an agent’s working solution may be imperfect yet advance the product enough to justify acceptance and later revision.

• He increasingly designs the codebase for model navigation rather than personal aesthetic preference. If agents consistently choose a conventional name, that name is probably strongly represented in their learned expectations and will be what future sessions search for. Replacing it with an idiosyncratic favorite may make every subsequent task harder.

• Human judgment remains decisive over what enters core, what becomes a plugin or skill, and which seemingly attractive contributions deserve rejection. TypeScript was chosen not because language syntax was sacred but because its ecosystem is approachable, widely used, and well understood by agents.

• Taste also survives at the level of delight. Startup messages such as “built on Caffeine, JSON5, and a lot of willpower” or an update saying “I’m in. It’s cozy here” prime users to expect playfulness rather than Microsoft Exchange. Steinberger’s judgment is that an agent would not independently originate that coherent product feeling.

14. Voice and simple terminals support parallel agency

• The transcript’s workflow speaker says that substantive prompts are now spoken through a walkie-talkie-style control and that voice was once used so extensively that it was lost temporarily. Keyboard input remains faster for folder changes and short shell commands, while reusable slash commands are reserved for genuinely repetitive actions.

• The speaker’s real setup is less extreme than the memes: two MacBooks, two large anti-glare displays, and terminals arranged side by side. Each agent window retains a small ordinary shell pane partly because the speaker once prompted the wrong project and watched a model spend 20 minutes “manically” trying to reconcile instructions with nonexistent code.

• The speaker avoids worktrees and elaborate UI state. “Discuss, give me options” or “don’t write code yet” substitutes for a formal plan mode; “okay, build” releases the model for 20 minutes or longer. Asking whether it has questions exposes missing context, though the correct answer is often, “Read more code to answer your own questions.”

15. Claude Opus 4.6 and GPT-5.3 Codex reward different drivers

• Steinberger calls Claude Opus 4.6 the strongest general-purpose model for OpenClaw, especially for role-play, character consistency, command following, creativity, and fast trial-and-error. Its weakness is overeagerness and lingering sycophancy—“You’re absolutely right” became so repetitive that he describes himself as allergic to it.

• His memorable comparison is that Opus is the slightly silly, funny coworker one enjoys keeping around, while Codex is “the weirdo in the corner that you don’t wanna talk to, but is reliable and gets shit done.” Fridman extends the joke: Opus feels American; Codex feels German.

• With skilled guidance, Steinberger says either can generate high-quality code, and Opus sometimes finds the more elegant solution. Codex 5.3 reads more before acting, tolerates long autonomous runs, and may persist for six hours when the desired end state is clear; Steinberger prefers its dry overthinking to a more interactive succession of localized attempts.

• He attributes much of the contrast to post-training rather than raw intelligence and advises giving a new model about one week before judging it. Claims of gradual model “degradation” may instead reflect adaptation, growing codebases, accumulated slop, and neglected refactoring. Comparing a $200 Claude tier with a slower $20 OpenAI tier further distorts first impressions.

16. The personal agent is moving toward an operating-system role

• Steinberger does not view OpenClaw as a direct replacement for Claude Code or Codex. Deep programming still benefits from a large screen and multiple visible sessions; OpenClaw is closer to a life assistant or coworker that can receive a GitHub URL, try a CLI, report what works, and coordinate specialist coding agents.

• “This is where the puck’s going”: the personal agent becomes more like an operating system. OpenClaw already supports sub-agents and TTY-driven coding tools, and Steinberger’s bossy assistant delighted in launching Codex and announcing that it was obeying.

• Both speakers expect the chat box to be transitional. Today’s prompt-and-response interface resembles early television broadcasting radio shows on camera: it reuses the preceding medium before discovering the new one’s native form. Voice, emotion-aware multimodality, proactive behavior, and dynamically generated interfaces may eventually converge.

• The hardware need not be a Mac mini. OpenClaw supports macOS, Linux, Windows, and WSL2, though native Windows still contains “dragons”; an old computer can serve as a node. Separate residential hardware can help browser automation because data-center IP addresses attract blocks and CAPTCHAs—although agents are already “quite good at happily clicking ‘I’m not a robot.’”

17. Skills and CLIs beat MCP when composability matters

• Steinberger regards implementing a basic agent loop as AI’s “Hello World”: call the model, expose tools, return tool results, and continue until it answers. Building one demystifies the product category and shows how much of apparent autonomy emerges from a relatively simple loop.

• His extension strategy begins with a CLI because models already understand Unix commands. A skill can remain one sentence in the base context, then load detailed instructions only when required; the agent invokes the command, reads --help if necessary, and brings only relevant information into its working memory.

• The critique of MCP is context pollution and limited composition. A weather MCP may always return temperature, rain, wind, and other fields as one large object; a CLI can pipe that output through jq, combine it with a script, and return only the number needed for the current decision.

• Steinberger still credits MCP with motivating companies to expose APIs and accepts exceptions such as Playwright, where persistent browser state makes a structured connection useful. His objection is not that protocols never work, but that many poorly designed MCP servers force models to consume blobs that Unix tools could filter naturally.

18. Every web application is becoming a slow API

• Browser control changes the bargaining position between platforms and users. Even if a service refuses an official API, a personal agent can open the website, click controls, and read results; blocking a reverse-engineered interface often converts a fast operation into a slower one rather than making it impossible.

• Steinberger’s short-lived Bird CLI used X’s internal web API until the company asked him to remove it. He accepted the request but proposed a low daily read-only allowance for personal accounts, enough to collect bookmarks, research them, and email summaries without enabling industrial-scale scraping.

• He strongly opposes agents posting indistinguishably as humans. API-generated tweets should be marked, agents should be able to hold clearly labeled accounts, and automated replies receive an immediate block from him: “As soon as it smells like AI,” there is no first-strike exception.

• The broader scarcity has inverted: “Content is now so cheap. Eyeballs are the expensive part.” Steinberger would rather read someone’s broken English—or even the raw prompt—than polished AI prose. He abandoned agent-written blog posts because steering consumed similar time while erasing nuance, and both speakers now value typos and rough human edges over prose, images, or infographics that “scream slop.”

19. Personal agents may erase most narrow app interfaces

• Steinberger asks why someone needs MyFitnessPal when an assistant already knows location, sleep, stress, and workout history. It could infer a brisket dinner, modify the next gym session, control an Eight Sleep bed, query cameras, or present a temporary interface tailored to the user rather than requiring another fixed app and subscription.

• Fridman recalls a suggestion that agents might eliminate 80% of apps, though the claim remains a possibility rather than a forecast with a timetable. Sonos, fitness trackers, cameras, calendars, and smart-home controls may survive as data sources and transaction APIs while their standalone interfaces gradually lose relevance.

• New businesses appear alongside the destruction. Steinberger wants to give an agent perhaps $100 of allowance and judge it on solving the problem, whether it orders food through an API, clicks through a website, or hires a person through something like RentAHuman. The valuable company becomes the most reliable agent-facing fulfillment layer.

• Incumbents cannot fully prevent user-directed access. Google’s developer process made Steinberger build his own Google CLI, and Gmail approval can be so burdensome that startups acquire already-certified companies; nevertheless, an individual can authorize their own assistant, and the browser remains a fallback. Companies can offer a clean API or accept that agents will operate the slower graphical one.

20. Beginners should optimize for play, not perfect output

• Steinberger’s advice is simply, “Play.” Build the idea already in your head, discard experiments without regret, and ask questions whenever understanding breaks down. The model is an “infinitely patient answering machine” that can restate a database concept at a simpler level without the delay or social friction of Stack Overflow.

• Open source remains his preferred school. Beginners should observe communities, read code, help without immediately sending a pull request, and learn how projects are maintained. High-agency non-programmers can travel surprisingly far because they keep asking questions rather than treating missing formal knowledge as a permanent boundary.

• He urged iOS specialists to stop defining themselves by one platform and think of themselves as builders. General software judgment transfers across “tech galaxies,” while agents supply syntax: TypeScript for web accessibility, Go for deployable CLIs despite his dislike of its syntax, Swift and SwiftUI for deep Apple integration, Zig for performance experiments, Python for inference, and Rust for highly parallel systems.

• There may eventually be a programming language designed for agents rather than humans. The counterforce is path dependence: models possess abundant knowledge of established ecosystems, so a novel language begins at a disadvantage. For now, Steinberger selects languages by ecosystem and deployment characteristics because the cost of personally learning every syntax has collapsed.

21. Burnout changed the objective from money to meaningful work

• Steinberger ran PSPDFKit for 13 years and spent two years making himself operationally obsolete before leaving the company. Programming was not what exhausted him; co-founder conflict, customer crises, management, and prolonged interpersonal stress were. After he could leave, he felt as if Austin Powers’ mojo had been removed, stared blankly at code, and booked a one-way trip to Madrid.

• He does not recommend working intensely with the plan to enjoy life only after retirement. Waking without a challenge quickly became boring, and substitute stimulation could lead somewhere dark. He now enjoys life more precisely because building again provides difficulty, direction, and anticipation.

• Money was an affirmation rather than the reason for PSPDFKit, and its returns diminish: “A cheeseburger is a cheeseburger.” Excessive luxury can disconnect a person from society, so despite affording hotels, he chose an old-style Airbnb room in San Francisco, met a queer DJ, and showed her how to make music with Claude Code. His preferred metric is experience—good or bad—because either kind produces feeling and learning.

22. Open source, strategic scale, and human disruption remain unresolved together

• OpenClaw currently loses Steinberger roughly $10,000–$20,000 per month. Sponsorship income flows toward individually maintained dependencies, with any surplus intended for contributor merchandise; OpenAI has begun helping with tokens. He cites Tailwind’s 75% staff cut as a warning that ubiquity does not guarantee monetization when agents bypass documentation websites.

• Building a heavily funded company does not excite him as much as it once might have. Enterprise features such as audit logs could create pressure to weaken the public edition, while a restrictive license would violate his preference that the project remain “free as in beer and not free with conditions.”

• Meta and OpenAI are the two strategic options he finds most interesting, but the decision was not finalized. His condition is that OpenClaw remain open source, perhaps through a Chrome-and-Chromium-style structure. Meta leaders personally used the product and sent candid feedback; OpenAI offered technology, model speed, and an NDA-constrained prospect he likened to being handed “Thor’s hammer.”

• Both speakers acknowledge the cost of the transition. Steinberger says programming may eventually become “like knitting” and that tokenized intelligence will reduce the scarcity behind developer salaries; Fridman insists Silicon Valley must respect the job loss and suffering hidden by its optimism. Against that, Steinberger cites small businesses automating invoices, a disabled daughter gaining agency, local or inexpensive models such as Kimi, and roughly 500-person community gatherings overflowing with builders. His closing thesis is neither painless nor modest: “That’s ultimately power to the people.”