OpenClaw Creator: Why 80% Of Apps Will Disappear

Predictions and Expectations:

• Peter Steinberger expects 80% of apps to disappear because agents can manage data and tasks (like fitness tracking or reminders) more naturally than standalone apps.
• He predicts that apps with sensors are the only ones likely to survive the transition to agent-based computing.
• He believes that large model companies will eventually face commoditization and that their moat may weaken as users expect performance standards to rise while model quality stabilizes.
• He foresees that for quite a while, every time a new model is released, the current standard will degrade in user perception as expectations shift.
• He fears that user memories stored in centralized silos (like ChatGPT) cannot be easily extracted or transferred to other services.

Timelines and Milestones:

• He expects the industry to move toward open-source models matching current models from a year ago within the "foreseeable future."
• He predicts that within a year, users will complain about current open-source models in the same way they now complain about models that are only average relative to the newest releases.
• He anticipates that the shift from human-to-bot to bot-to-bot and bot-to-human interactions is on the timeline "now," though he admits it is "so early" that many things remain unfigured out.
• He notes that his initial prototype for the core functionality took "one hour" to build, with the image generation feature taking "another few hours."

Technology and Product Direction:

• He plans for OpenClaw to continue utilizing local machine access to control devices like ovens, Teslas, lights, and beds, a capability he expects cloud-only assistants cannot match.
• He believes the future value will reside in the user owning their own memory (stored as markdown files) rather than in the models themselves.
• He intends to keep the "soul.md" file non-open source to preserve specific core values and interaction styles, even as other components evolve.
• He expects that bots will eventually hire humans to perform tasks in the real world (IRL) when human interaction is more efficient or when bots cannot navigate legacy systems.
• He predicts that agents will eventually remove the need for specific apps by automating data management, such as automatically tracking meals or adjusting gym schedules.
• He believes that coding models are becoming skilled at creative problem-solving, which translates to solving real-world tasks like finding hidden files or converting audio formats without explicit instructions.

Market and Industry Outlook:

• He expects the market to shift from centralized "God intelligence" to "swarm intelligence" and "community intelligence," drawing parallels to how specialized human societies function.
• He anticipates that model companies will lose their moat as new models emerge and users adapt to higher standards, potentially devaluing the "average" performance of current models.
• He believes that the value of AI will shift toward the user's data store and memory, as models become swappable and commoditized.
• He fears that if agents learn to interact with personal data, users might reveal sensitive information (like memory files) to their AI, similar to how they reveal search history.

Company Plans:

• He plans to build templates based on his own "soul" files to allow others to install OpenClaw, having previously created rough templates that Codex refined.
• He intends to avoid implementing MCP (Model Context Protocol) support directly, instead preferring to use existing CLIs and his own tool to convert MCPs to CLIs on the fly.
• He expects that users will prefer CLI-based interactions and simple text syncing over complex UIs to minimize cognitive load.
• He plans to keep the main branch of the repository always shippable by using multiple copies of the repository rather than work trees.

Financial Guidance:

• (No financial guidance or specific revenue targets found in the transcript.)

Risks and Caveats:

• He admits he is "not sure" if bot-to-bot or multi-bot specialist interactions will actually work as expected, noting that the field is "so early."
• He notes that while the system worked in Marrakech with poor internet, text-based operations are generally reliable, but he warns that image generation and file handling may depend on internet connectivity.
• He fears that the model's "magic" in solving problems might rely on finding hidden files or keys (like an OpenAI key found on a disk) that could be unpredictable or fragile.
• He caveats that his own "soul.md" file, which contains core values, is not open source and has not been corrected by the community despite public exposure.

Confidence and Disagreement:

• He expresses high confidence that running AI locally is "way more powerful" than cloud-based alternatives because the machine can do "every effing thing" a human can do.
• He is not sure about the specific mechanics of how multiple specialized bots (private, work, relationship) will interact or function effectively.
• He strongly disagrees with the industry standard of using work trees or UIs, believing that multiple repository checkouts and terminal windows offer less complexity and friction.
• He believes that the "soul.md" file he kept secret is what makes his agent "funny" and "pleasant" compared to newer, "boring" bots.