AI Engineer World's Fair 2025
On Curiosity — Sharif Shameem, Lexica
Loading slides…
Read the talk
On Curiosity
From browser-compiled JSX to failed shopping agents and recursive problem maps, Sharif Shameem explores how small demos reveal capabilities that model specifications cannot.
From a talk by Sharif Shameem
Before you start: Basic familiarity with language models, context windows, and React components will help with the implementation examples.
An idea arrives on the bus
Henri Poincaré had been stuck for weeks on a problem involving Fuchsian functions. Then, while boarding a bus, an insight arrived suddenly. In his account of the episode, the breakthrough follows sustained work but occurs away from the desk. The interesting question is what happens between deliberate effort and an idea becoming available to conscious thought.
For Sharif Shameem, founder of the image-model company Lexica, curiosity is the feeling through which subconscious pattern recognition surfaces. That is his interpretation of the creative process: experience accumulates, connections form before we can fully explain them, and an interesting possibility pulls us toward an experiment. A demo turns that feeling into something other people can inspect.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From an exciting idea to a working experiment
The demos follow a recurring sequence. An idea initially seems wonderful; implementation makes it look impossible; then a concrete workaround gets it working. Small context windows make that middle stage particularly tangible: an application that seems straightforward in your head may not fit into the model’s input at all. The eventual reward is both the pleasure of making something work and new evidence about what the model can do.
This makes demos more than presentation material. You cannot fully understand a model’s capabilities without interacting with it, and building an interesting application forces that interaction. Curiosity chooses the question; implementation exposes the constraints; the working result makes the possibility visible.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make generated code immediately visible
The first examples come from the release of GPT-3 in 2020. Shameem describes its context window as roughly 2,000 tokens. He recalls a price of about $75 per million output tokens, needing permission before sharing model outputs publicly, and restrictions that prevented shipping a chat app. These are recollections of the early API environment, not current pricing or usage rules. Despite those constraints, the model felt like a new computing primitive: software could be produced from a description.
A talk by Bret Victor supplied the design principle: creators need immediate feedback from what they create. Copying generated code into an editor, compiling it, and running it inserted several steps between a request and its visible consequence. Shameem put a JSX compiler in the browser, shortening that loop. The selected demo frame shows the result: a request for rainbow-colored buttons beside generated code and rendered rows labeled Red through Violet. He contrasts the old constraints with a Claude system prompt he describes as 25,000 tokens at the time of the talk—a striking comparison, rather than a fixed specification for Claude.
A second demo takes a request for a Google homepage. The whole application would not fit into one generation, so the implementation sends three prompts in parallel, generates three components, and joins their outputs in the background. Decomposition makes the application possible within the available context. The interface presents one request and one result even though the underlying work is split across several generations.
The idea combines two things Shameem had encountered separately: language models that could write code and a design principle about immediate feedback. His point about a person’s unique experience is concrete here. The demo did not require a wholly unprecedented ingredient; it required noticing a useful connection and giving it an interface.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A page that fits is not a task that succeeds
In a 2021 experiment, the question changes from generating an interface to operating one: could a model buy AirPods in Chrome? Shameem recalls roughly 4,000 tokens of context for this demo, while the Walmart shopping page alone occupied about 24,000 tokens. Passing the full page into the model was therefore not workable.
After several frustrating days, he wrote a custom HTML parser that reduced the page to its essential information. That solved the representation problem: the model could receive a useful description of the page within its context window. It did not solve the shopping problem. The agent became distracted by the terms of service and failed to complete the purchase.
The distinction matters. There are two separate questions: can the system expose a usable environment to the model, and can the model reliably pursue its objective within that environment? The parser addressed the first. Shameem interpreted the resulting behavior as a sign of agency latent in a model pretrained on web text, but the failed purchase also exposed how far that behavior was from dependable autonomy. In 2021, even seeing the attempt was surprising.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A basketball video suggests a different product
Farza’s basketball demo moves the exploration to Gemini 2.5 Pro. It takes footage of him playing basketball and uses it for a shot tracker, with feedback requested as though Michael Jordan were watching his gameplay. The important shift is the input and experience: a video becomes the basis for an application organized around a physical activity.
Watching a screen and commenting on code is one use of video understanding, but this demo makes another class of products easier to imagine. A concrete example expands the search space more effectively than saying that a model supports video. Shameem goes further, conjecturing that frozen model weights could support another ten years of compelling product development without additional training. That is a forecast about unexplored applications, not a measured capability horizon.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Excavation, barnacles, and a foggy pond
Shameem invokes a distinction attributed to Richard Hamming: uncertainty belongs naturally to scientific discovery, whereas engineering ordinarily requires knowing how to achieve a specified result. AI engineering mixes these activities. You may have to discover whether a capability exists before you can build a dependable product around it. In Shameem’s metaphor, the work resembles excavation: the demo is a tool, and curiosity directs the search.
That discovery does not belong exclusively to model laboratories. Shameem reports that OpenAI researchers told him they had not known GPT-3 could browse the web or produce functioning React components in the ways his demos showed. His anecdote illustrates why application builders can contribute knowledge: an unfamiliar interface or task can expose behavior that the people training the model have not explored.
Darwin’s barnacle research supplies a longer-timescale analogy. Shameem points to eight years of study and connects that work to evidence supporting evolution. The lesson is about delayed visibility: while an investigation is underway, it can be hard to distinguish consequential work from apparently aimless play. Its contribution may become clear only later.
By the talk’s 2025 vantage point, Shameem contrasts early contexts of roughly 2,000 tokens with windows reaching roughly a million tokens. More capacity expands the opportunities, but it does not reveal a complete route through them. His image is a foggy pond: take one stepping stone, inspect what becomes visible, and continue or backtrack. The next experiment depends on what the previous one reveals.
An unattributed tweet sharpens the danger of choosing too narrow a product category: marketing Claude only as a coding or reasoning model resembles selling an intergalactic spaceship as a toaster because one surface occasionally gets hot. Familiar uses can hide less obvious capabilities. Exploration and play help uncover those possibilities, while each builder’s particular experiences suggest different places to look. Shameem’s recommendation is to act on those connections even before their value is certain.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A running experiment and a broken banking app
The extended discussion immediately supplies an example of that transfer. The moderator had adapted the basketball idea to running footage and reported feedback resembling advice from a $600-per-month running coach. The exchange touches on gait and feedback at the level of individual steps, though reproducing the arrow above the athlete’s head remained unfinished. This is a personal usage anecdote, not an evaluation establishing that the model can replace a coach.
Shameem then opens another demo from a few weeks after the GPT-3 API became available. A friend’s enthusiastic claim that OpenAI had created AGI prompted him to try it; the useful discovery was more specific: it could write code. Connecting the API to a browser compiler made that ability into a back-and-forth visual editing experience, one he retrospectively compares with chatting beside code in Cursor.
The application is a simple bank balance interface. Shameem describes requests to add $3 or subtract $5 and reports a revealing bug: a button for giving away all the money also cleared a negative balance. The following JSX is an illustrative reconstruction of that reported behavior, not verified source code from the demo. In this reconstruction, starting from zero, adding $3 and then subtracting $5 produces a balance of −$2; clicking the final button erases the debt because this illustrative handler unconditionally assigns zero. The recording does not establish the original handler, initial balance, or exact execution sequence.
jsx
import { useState } from "react";
export default function BankDemo() {
const [balance, setBalance] = useState(0);
return (
<main>
<output>Balance: ${balance}</output>
<button onClick={() => setBalance(value => value + 3)}>
Add $3
</button>
<button onClick={() => setBalance(value => value - 5)}>
Subtract $5
</button>
<button onClick={() => setBalance(0)}>
Give away all my money
</button>
</main>
);
}
A visible, interactive result makes the program’s behavior easy to discover—including behavior that violates the meaning of the operation.
Shameem describes the demo as an early form of vibe coding, severe bugs included. It encouraged people to see these systems as reasoning engines that could operate tools, rather than only as generators of prose. His metaphor of intelligent beings connected to different apparatuses is a way of thinking about tool-enabled behavior, not a claim that the banking logic was sound. The practical invitation remains modest: build something fun and simple, then share it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The demos used a base completion model
An audience question clarifies an important historical detail: these early demos used the base GPT-3 model, not an instruction-tuned assistant. Shameem recalls instruction models arriving about a year later. The interaction was prefix completion: give the model text, include a few examples of the desired pattern, and let it continue. The examples established what kind of continuation was useful. This is the mechanism behind the early demos, rather than the conversational instruction-following interface readers may now expect.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Revisit the assistant that knows your work
Asked how curiosity and the subconscious became a moral argument, Shameem describes looking back at why he built the demos. His starting point was frustration: powerful models were available, yet their capabilities remained poorly understood. Thinking about pioneers such as Licklider and Alan Kay made that frustration sharper. They imagined ambitious systems with limited computers; contemporary builders have far stronger tools but often reproduce familiar products.
The alternative is to treat older computing visions as a source of unfinished product ideas. Shameem reads Licklider’s symbiosis vision as pointing toward an assistant that knows what you are working on, maintains that context, and can help immediately. He contrasts this with ChatGPT’s New Chat workflow as he experienced it at talk time: starting over with little continuity beyond a few remembered facts. The product opportunity is sustained working context, not simply another place to type a prompt. Doing justice to those earlier ideas means trying to implement them.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Multivac makes a large problem explorable
The final demo, Multivac, returns to the small context window with a larger ambition: how can a model help reason about a problem that cannot fit into a single request? Its answer is to break the problem into digestible subproblems and expose the structure in a visual interface.
The climate-change example illustrates the interaction:
- Enter an ambitious question, such as how to address climate change.
- Generate candidate directions, such as encouraging vegetarianism, building wind turbines, or installing solar panels.
- Click a sub-idea to generate a more detailed breakdown.
The visible map shows how this can continue: encouraging vegan adoption branches into explaining its benefits and making vegan versions of familiar recipes. Each expansion gives the user a smaller piece of the problem to inspect. These are proposed directions, not executed actions or evidence that the overall problem has been solved.
The ambition is a useful reasoning assistant and, eventually, a thought partner that contributes ideas of its own. Looking back at the rudimentary interface, Shameem invites someone to rebuild it with a newer model, using the name “Opus Max.” The recording ends with an unfinished experiment worth revisiting: keep the inspectable problem structure, bring stronger capabilities to it, and see what becomes possible.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Licklider's 1960 proposal for computers that participate in human thinking, exploration, and decisions.
Further reading
The original GPT-3 paper explains few-shot prompting and the model's 2,048-token context window.
The French account of Poincaré's breakthrough while boarding an omnibus at Coutances.
- Darwin's barnacle researchArticle
Historical context and correspondence sources for Darwin's eight years studying barnacles and their contribution to his evolutionary work.
Full transcript
Loading transcript…