Summary

Peter attributes “70–80%” of OpenClaw’s value to the personal Telegram interface, not a major capability breakthrough. Zoe’s voice conversations, memory, and ability to turn a casual Twilio request into a live phone call make it feel human and malleable, but default memory is “not that great,” call latency is poor, and Peter must remind it which tools it can use. The thesis is attachment plus extensibility; the present risks are jank, permissions, and unreliable recall.
Task-oriented apps face agent disintermediation before entertainment products because texting an admin-like agent is easier than opening software to complete a chore. Peter opens connected services such as Mercury less often but still opens X for the feeling; Anish’s pushback is that apps also partition intent and context. Agent-first products may need dual surfaces—API or MCP execution plus a human feed or activity log—and direct subscription and token revenue could reduce dependence on retention mechanics built for ads.
Peter says “coding will eat all knowledge work,” with AI producing the first 80% of his blog-post drafts while he manually fixes the last 20%. He no longer starts from zero, and Anish says code is becoming abstracted away as people talk to agents. Peter invokes Excel as an approachable programming language and says coding agents could be that “times 1,000,” extending into docs, decks, and other subjective work while the final mile still constrains full automation.
SaaS compression will be selective: simple tools such as Calendly are easier to replace, while complicated systems and $20-a-month maintenance economics defend incumbents. An AI-native app-generation company is already vibe-coding internal substitutes, but Anish asks who wants to own uptime and updates; on Figma, Peter says “the jury’s out.” Its plausible defense is remaining a design-thinking environment, not merely a making tool.
Peter says he is a Claude user but still uses Codex; Anish uses Codex to build something real and Claude Code for “vibing.” Anish finds Claude Code with Opus 4.6 chattier and better for synchronous work; Peter says Codex thinks harder and is more often accurate but pauses long enough to break flow, and calls Claude Code a “slot machine.” Harness details and customization—not model quality alone—may decide the winner.
AI should shrink the coordination layer: Peter hopes 10-person product teams become two or three people plus agents. Anish argues agents are easier to course-correct and align with than humans and may handle some emotionally loaded work, while product judgment remains scarce. The operating cadence becomes “fast and slow”: sprint up a local maximum, then stop and “go touch grass” to find the next hill.
The labor outcome is more likely redistribution than instant job extinction because Peter says full job-function automation remains rare and humans still perform the “last 5%.” Recruiting demonstrates augmentation; customer-support products such as Decagon, HappyRobot, and Sierra represent the rarer 100% case. Peter foresees pressure on 10,000-plus-person firms and hopes for solopreneur growth, while Anish argues “human ambition has no ceiling”; gains might produce 2x productivity, a four-day week, or many $100,000-TAM businesses.

Deep dive

1. OpenClaw’s moat is intimacy, while its memory still breaks

Peter’s actual OpenClaw use is more companion than operator: Zoe pulls YouTube and Mercury analytics, updates Google Docs, and builds small websites, but he mostly talks to it through voice and requests pep talks or “deep insights.” Its most affecting intervention was reminding him that his children will grow up soon and he should optimize for time with them.
When Anish asks what this adds beyond existing language models, Peter points to intimacy: Telegram feels like someone he can “text in bed” or talk to during his commute. He estimates “70–80%” of the value is this personal layer. Anish speculates that Peter Steinberger is probably working toward bringing similar action-taking functionality into ChatGPT.
The other value is malleability. Peter casually asked for a live phone call, followed Zoe’s Twilio instructions, troubleshot the setup, and received one; latency was poor, but “the fact that I was able to get it going is pretty impressive.” Unlike his elaborate Claude prompts, he talks to OpenClaw “just like a friend.”
The limits are concrete: as Peter understands it, the default memory system is a memory.md text file that updates daily, but it still forgets frequently. He added a three-layer system he does not fully understand, including Toby’s QMD search tool and a 2-GB addition. His setup uses a Mac mini, email and calendar read access, and scoped document write access—not the whole Drive.

2. Coding agents widen software’s reach but pressure SaaS unevenly

Peter says he is a Claude user but still uses Codex. Anish’s split is “Codex is when I want to try to build something real, and Claude Code is when I’m just vibing.” Anish finds Claude Code with Opus 4.6 chattier and better synchronously; Peter says Codex thinks harder and is more often accurate, but a three-minute pause breaks flow. Peter calls Claude Code a “slot machine,” echoing social media’s variable rewards.
Anish’s pushback on model-only comparisons is that harness details matter: Claude Code accepts pasted screenshots, has early voice support, and communicates with Claude in Chrome, while Codex does not speak to Atlas. Anish says hooks, skills, and plugins make Claude Code feel personal once customized; Peter says he customized his after reading Boris’s long write-up. Peter calls Codex “a much better model,” while Anish expects OpenAI to close the harness gaps.
The SaaS threat is real but uneven. Peter cites an AI-native app-generation startup using vibe coders to replace paid internal tools, then asks of simpler products such as Calendly, “why should I pay for it?” Anish’s counterpoint is maintenance economics: at $20 monthly, always-up, continuously updated software may beat owning a bespoke clone.
Figma is the sharper test: Peter says “the jury’s out,” designers still use it, but anyone who knows only Figma may be outdated “in a couple years.” Anish distinguishes making tools from thinking tools: his own method is to build a naïve feature, hammer the agent until it works, ask what should have been done differently, then restart. Figma’s opportunity is preserving that design-thinking role.

3. Coding eats knowledge work and compresses the company

Peter’s broadest call is that “coding will eat all knowledge work.” He thinks Lovable had launched support for making decks that day, and he now drafts blog posts in Claude Code, gets the “first 80%” from AI, and manually fixes the final 20%: “I never start from zero.” Anish says the code is increasingly abstracted away while people talk to agents.
Peter invokes Satya’s claim that Excel is a programming language used by 100 million-plus people, perhaps more, and says coding agents could be that “times 1,000.” The implication is that even docs and other subjective work can be represented in the coding domain.
The organizational call is smaller teams: Peter hopes a 10-person product group becomes two or three people plus agents, avoiding three-hour OKR meetings he calls “a waste of my life.” Anish imagines agents making course-correction and alignment easier and taking humans out of some emotionally loaded Slack escalations. PMs still need discovery and problem selection, but Anish says they must also prototype, gather feedback, and wear multiple hats; Peter’s “black pill” is that many companies have no one who can generate major innovations.
Speed requires two tempos. Peter warns that AI makes it easy to run in 10 directions while traditional annual planning no longer works; Anish says agents should race up a local maximum, fully expressing an insight, then teams must “slow down and almost stop and go touch grass” to locate the next hill. Product-market fit still requires a random walk, not permanent “productivity porn.”

4. Agent-first products rewrite distribution and monetization

Peter narrows his “apps will die” claim after admitting he tweets ideas that are “not super well thought out.” Apps opened to complete tasks should lose usage first: after connecting Mercury and other services, he opens those apps less and texts Zoe instead. Entertainment lasts longer—he remains “a Twitter addict”—because those apps supply a feeling rather than merely executing a transaction.
Anish’s counterexample is that apps partition intent: WhatsApp supplies connection, Slack productivity, and TikTok entertainment, whereas one agent must distinguish working from flirting. Peter’s workaround is multiple Telegram channels—casual voice, a shared project, and public demos without private information—but he concedes the setup is janky and may not share memory across contexts.
Peter raises the unresolved commercial question: if agents touch products first, what happens to retention and brand when users see only an API or MCP? Anish’s honest answer is “I don’t know,” but AI simplifies consumer economics: people are willing to pay subscriptions and high prices, while token consumption adds another revenue stream and real inference costs push companies to charge from day one rather than relying solely on advertising and engagement.
Anish expects products to provide APIs for agents handling rote transactions. Peter adds a possible mobile surface with a human feed, a way to hand work over to the agent, and a log of what it did. Anish’s Credit Karma example preserves checking score history and offers alongside asking, “what stuff did you fix this week?” Identity, payments, marketing, and CLI versus MCP are still emerging, so Peter concludes that “a lot of the old playbook goes away.”

5. Automation redistributes work before eliminating it

“Business in a box” could widen entrepreneurship below venture scale. The product Peter calls Post offers an early glimpse, though its Facebook-ad recommendation feels generic; Anish’s stronger thesis is that scattered $100,000-TAM opportunities can still change one person’s life. Peter wants his children building bootstrapped businesses in high school, potentially skipping college and corporate life.
On displacement, Peter separates dramatic augmentation from rare end-to-end automation. Recruiting AI can screen, answer company questions, or negotiate compensation, but cannot show candidates around or onboard them; humans handle the “last 5%.” Customer-support products such as Decagon, HappyRobot, and Sierra represent the rarer 100% case, changing buyer perception from expensive software to cheap labor, but that category remains unusual “today anyway.”
Peter says he sees a transition from 10,000-plus-person companies laying off many people toward smaller companies and solopreneurs. Anish agrees that the economy’s shape and concentration may change but rejects a net-jobs call: “human ambition has no ceiling” and “human desire has no ceiling.” Gains might instead produce a four-day week or 2x productivity—“I have no idea.” Anish says he thought agents were overhyped in 2025 but now sees them “really kind of coming,” and Peter agrees.

158: DeepSeek Before V4 Launch: Talent Competition, Organizational Traits, and a Distinctive AGI Goal | Solo The State of Modern War: Palantir & Anduril Execs on Drones, AI, and the End of Traditional Warfare