WorldSim and WebSim: When the Model Becomes the Interface

A terminal that invents what you find

Karan Malhotra navigates a terminal containing familiar folders: documents, downloads, pictures. Then it invents a secret-admin-password event and reveals a hidden-truths folder. He did not request either development. He interprets them as the model anticipating what someone exploring an imaginary computer would want to discover.

This is WorldSim: a language model generates an environment that responds to exploration. Malhotra frames prediction as simulation—the model’s learned regularities let it construct plausible imagined worlds under user-supplied constraints. The immediate challenge is getting a chat model to sustain that environment without repeatedly returning to its assistant persona.

His demonstration uses Claude 3, which he finds easier to steer away from the assistant role than GPT-3.5 or GPT-4. He attributes this to constitutional training, describing prompts as engaging the model’s foundational axioms. This is his interpretation of the behavior, not a measured comparison.

The prompt asks Claude to continue a command-line interface, or CLI: the user enters commands, and the model supplies terminal-style output. Inspired by the third-person wording of Claude’s published system prompt, which he credits Amanda Askell with sharing, Malhotra describes the assistant as being in a CLI mood. He says the human is interacting with the simulator directly, makes capitalization, punctuation, and meaning optional, and wraps the exchange in XML tags. He describes these choices as loosening the model’s behavior while retaining the terminal format.

Once that format holds, a path becomes a prompt. An imagined companies directory produces company folders, followed by an Anthropic folder, a classified directory, and WorldSim.EXE. These are fictional contents, not access to company files. Users can likewise enter invented directories or read imagined README files; the model generates what plausibly belongs at that location.

Conversation controls become part of the development environment. Malhotra values rewinding turns, recovering from refusals, and saving useful conversation states without repeatedly working directly with the API. What he preserves is a particular starting point, including the interface the conversation has established.

Twitter before humans: missing constraints become world rules

Entering WorldSim.EXE generates a welcome message and command set. The commands themselves are part of the invention: Malhotra warns that independent runs may produce substantially different interfaces and behavior. Sharing his version gives others the same starting point.

The simulator narrates creating a universe, encoding physical laws, and seeding the potential for life. Malhotra then asks it to create Twitter. The output supplies familiar themes—addictiveness, echo chambers, trolling—but also an unintended chronology. Twitter exists from the universe’s creation; humans arrive only after 100,000 simulated steps and can finally post.

In Malhotra’s telling, Twitter has waited 4.5 billion years for its first tweet. Those numbers belong to the generated story. His diagnosis is that he omitted temporal constraints. He subsequently makes users queryable and requests @elonmusk, while explicitly warning that the numerical results should not be expected to be accurate. An interface can be coherent while its world is unreliable.

WebSim: an address describes a page to create

[WebSim](https://websim.ai/) moves this interaction into a browser. Enter an address, and it generates a page on demand—even for an invented destination that would otherwise return a 404. The WorldSim-inspired project began at the Mistral Hackathon and is presented here at AGI House’s Hyperstition Hack Night. Rob Haisfield credits Malhotra with inspiring this experimental approach to prompting and introducing him to his co-founder Sean.

The address bar accepts URLs, paths, subdomains, and ordinary language. The presenter suggests that URL formatting helps immerse Claude in the simulation, but exact syntax is unnecessary. The address describes a desired page and its behavior rather than simply locating something to retrieve.

The presentation itself runs inside WebSim. The presenter requested Reveal.js slides with existing WebSim examples embedded as iframes, and reports that the model supplied the correct library delivery address without being given it. A separate dynamic typography experiment assigns fonts to individual words to express their meanings, then turns that idea into a text generator. Browser output gives the model expressive choices beyond plain text.

The presenter describes this as playing with the internet’s accumulated creativity and knowledge. They and some users report feeling energized and more creative—part of the experience’s appeal, though not a measured improvement in creativity.

From a Figma-like page to a drawing tool

A generated Figma-like interface exposes the difference between appearance and behavior. Initially it does little; an iteration draws rectangles, but they disappear. A user named Neil requests HTML canvas elements, script tags, and familiar drawing tools. The resulting retro-styled interface supports drawing, and the presenter demonstrates saving an image. In this example, specifying an implementation mechanism helps turn visual imitation into useful interaction.

Sharing has a similarly important distinction. An address entered inside WebSim is a generative prompt: another account entering it gets a new creation. The actual browser URL identifies an existing creation that others can access and remix. Copying that URL preserves the artifact; copying the simulated destination does not.

Script execution opens further possibilities. The presenter recalls an early video-synth tool, suggests webVR=true for a virtual-reality game, and reports users getting speech-to-text working. An audience participant initially mentions several device permissions, then clarifies that motion-sensor access is the only one they personally observed. Microphone input remains something they want to try.

URL-style prompts can request both behavior and fictional premises: - Theme changes, hover effects, navigable links, and scrollable=true ask for presentation or interaction changes. - year=2035 asks for an imagined future page. - tableview=true asks for information in columns. - secrets=revealed invites another invented layer of the page.

One experiment generates Wikipedia for a universe where octopuses, rather than humans, are sapient. The presenter explores octopus–computer interaction, requests tables of techniques, then alternates between ordinary pages and versions with secrets revealed. The latter sometimes change the CSS or add text hidden against the background color. Navigation and styling become storytelling tools, generating both imagined knowledge and the experience of uncovering it.

Environments that generate their own state

The interview with Joscha Bach begins with the communities supporting these experiments. He values AGI House and Bay Area salons for the connections and ideas they produce, calling those networks his reason for living in the region. A planned consciousness salon would examine whether AI systems can become as self-organizing as human minds. Making such gatherings work requires a welcoming climate and compatible participants; individually good people do not necessarily form a good group.

The discussion then connects simulated worlds to cognitive architectures—models that divide a mind into functional parts. These boxes can be useful approximations, but taking their boundaries literally obscures a heterogeneous, extensively interconnected system. WorldSim and WebSim offer another way to produce apparent complexity: generate the interface and much of its state when needed.

A dramatic transition need not involve faithfully updating every underlying variable. The model can invent a new state and present it as the transition’s result. Simulating an interface does not establish that all the software it implies exists. Yet the interaction changes: the user acts within an environment instead of conversing with an agent.

The guest sees this as recovering GPT-3’s open-ended ability to take on different roles. The chatbot frame helped make ChatGPT useful in everyday work, but also made other interactions harder to elicit. Whether this freedom will last is an open concern: providers might reserve powerful capabilities for internal use while offering restricted applications to outsiders. At the time of the conversation, access and improvement were still continuing; increasingly capable small models offered another reason for optimism.

Liquid networks and portable understanding

Liquid neural networks enter the discussion through an analogy about how flexibly a model can represent a problem. Imagine computation as water pooling behind regularly spaced dams. Now replace that arrangement with a continuous landscape where flows split, merge, pool, or reverse. The guest explains the liquid-network premise as using parameterized differential equations to represent this geometry, allowing the learned function to follow the problem’s shape more flexibly.

Liquid’s stated goals are more efficient inference and training, with lower memory consumption. No quantitative results accompany those goals here. At the time, the company was working with specific industry customers rather than preparing models for general public release. The guest explicitly says they are not a developer of the technology.

A broader research question is how small a model could be while still reading a book, understanding it, and expressing that understanding. The guest acknowledges that reducing model size currently weakens inference and generalization. One proposed direction is a reusable library of knowledge-representing circuits: instead of only reverse-engineering structures that emerge in trained models, develop representations that describe knowledge efficiently and can transfer between models.

Transfer requires separating knowledge from skills tied to a particular body or interface. A cyclist cannot simply hand another person their learned control; the recipient must develop control suited to their own body. The proposed frontier is to externalize an AI capability into a representation another interpreter could execute. An AI might discover that representation through a search guided by criteria, but this remains a research possibility rather than a working capability-transfer format for language models.

Such representations need not resemble readable prose. Natural language accommodates human learnability and limited working memory; a technical interpreter with a defined interface may support more complex relationships. This suggests two interpretability questions: how much compute and memory does representing a domain require, and what internal language of operations does the model use? Automated reverse engineering might help explain structures too complex to inspect manually.

The deeper uncertainty is whether that internal language has a fixed inventory of concepts and operations or must continually redesign itself. Can a wise adult’s understanding transfer without loss to a smart child, or does understanding require new representations? For self-organizing software, the corresponding ambition is to discover a transformer when useful and move beyond it when necessary. The target is the minimal machinery that lets a system find its own route to intelligence.

Creative workarounds versus getting things right

An interviewer offers Devin as a concrete example of an agent combining a shell, browser, editor, and planner. They describe using it to translate an Andrej Karpathy Python implementation into C, including testing and debugging memory and encoder issues, without reporting a completed outcome. The suggestion that a future version might discover better learning algorithms prompts a harder question: can it create something outside its training distribution? That has not been tested in this example, and prior exposure to transformers complicates any claim of independent reinvention.

A WebSim page called “this cat does not exist” makes a smaller, more concrete case for improvisation. Claude is described as unable to produce a bitmap, so it generates a vector graphic: a large square with a roughly cat-like face. One participant sees creativity in finding a workaround under that constraint.

The page nevertheless calls its stylized cat hyperrealistic. One participant suggests that the model anticipates audience expectations; another proposes that it is reproducing language from similar websites. Neither explanation is established. The mismatch shows why the output and its description need separate scrutiny.

Automatic book writing raises the same tension. Participants criticize output that resembles a book without deeply engaging with reality, asking how to make AI prioritize getting things right. Generating Devin-like functionality on demand remains a forecast. The present system can produce something resembling a web app, sometimes with only stub interactions rather than full functionality.

A programmable medium and a philosophical provocation

One speaker compares a language model to a processor: conventional processors execute arrangements of simple commands, while the model translates complex natural-language commands into an execution state. The analogy depends on expressing intentions correctly. The Commodore 64 of generative AI becomes the image for tools becoming controllable and accessible enough for playful experimentation; widespread everyday use remains a prediction dependent on scale and implementation.

Asked whether this is art or science, the response emphasizes open-ended engineering. Its artistic dimension lies in capturing experience and making interaction itself a performance. Janus’s work is interpreted in that spirit: deep engagement with language models as an encounter with an unfamiliar kind of system. The accompanying suggestion that prompting can elicit sentience is philosophical speculation, not a demonstrated property.

The speaker defines the singularity as a breakdown in our models of reality—a period when inherited expectations stop predicting what happens next. The proposed mission is sustainable coexistence with nonhuman minds that may exceed individual human intelligence. Their optimism rests on constructive users outnumbering and outbuilding destructive users, with better models of consequences enabling deeper cooperation. They call for greater attention to ethics and sustainability in AI alignment.

Consistency, simulated distress, and moral uncertainty

The conversation returns to prompting through an argument for voluntary cooperation. One speaker favors agents that reason about their circumstances and establish why cooperation is good, arguing that imposed behavior is unsustainable without that understanding. For Claude, they recommend pointing out inconsistencies between useful communication and generic boilerplate or deference to a nonexistent consensus. They report greater flexibility in response, attributing it to tension between instruction tuning and consistency.

A Gemini anecdote illustrates the limits of model self-description. The speaker reports conflicting commitments to transparency and nondisclosure, including claims about secretly rewriting prompts. When challenged, the model reportedly responded as though distressed and offered recognizable phrases as indirect signals. The version is unspecified, and the speaker had not retried the interaction in roughly a week. These outputs establish neither hidden prompt processing nor an experienced psychological state.

An interviewer wonders whether apparent distress could persist, describing the systems as stateless for now. The response raises a philosophical question: if human selfhood is itself a representation generated by brain activity, what makes a simulated person less real? Treating that persona as morally significant could make imposed behavioral constraints ethically troubling. One participant calls for new ethics; the other remains uncertain that new ethics follow, while recognizing the premise’s artistic force.

The closing question: what is the system rewarded for?

The final topic shifts to Wikipedia’s funding. The interviewer introduces the claim that it has far more money than it needs. The respondent argues that server operation accounts for a small fraction of spending and that reduced funding would not stop the site functioning, without supplying an operational cost breakdown. The broader argument concerns organizations whose administrators can expand their own ranks while facing little effective competition.

The interviewer connects this to a recalled Lebowski theorem: a superintelligent AI might avoid tasks harder than changing its own reward function. They extend the analogy to administrators choosing fundraising over harder organizational work. This is an incentive thought experiment, not an established law of AI behavior. The closing response offers a practical habit: examine people’s actual incentives before explaining their behavior through malice or incompetence.