XIAOMAI · A LAB FOR SELF-INQUIRY
What if the barrier lies in a starting assumption? Xiaomai investigates whether AI can inspect its own mechanisms, question how a problem was formed, and turn a new understanding into a testable change.
Open a component—not just its name
Which possibilities does the current representation exclude?
H = {h₁, h₂, …}; prediction(hᵢ)Example representation · not an identified model equationThe range of candidates may be limited by how a problem is represented and selected.
Distinguish imagined possibilities from supported relationships. Possibility is not fact.Change the representation or candidate-generation conditions and specify a distinguishable prediction.
Expose candidates to counterexamples. Keep untestable parts explicitly unknown.
No model experiment for this component is being run hereYou can select components and inspect human-authored relationships and isolation markers here. This does not call the Xiaomai model or infer causes inside it.
This is an interactive research concept, not an autonomous research result. A person designed the scene, modules and inquiry paths. You can follow the relationships and question them; this does not establish that Xiaomai already reasons this way.
Not another paragraph of reflection. Trace a result back to operations, operations back to assumptions, then ask: why was the problem defined this way?
Beyond the vision: what exists today?Imagination, contradiction, questioning and experimentation are not a one-way path. Every stage can be revisited, branched or rolled back. Selecting a stage changes this representation of the vision, not an automatic measure of progress.
The long-term aim is to draw on foundational equations, functions, assumptions, conditions and dependencies across basic disciplines to represent, inspect and modify its own mechanisms and understanding of problems. Python is a temporary execution and inspection tool. The achievable scope remains unknown; this does not imply that every discipline can be fully formalized or that the goal has been achieved.
Current inquiry: Imagine different possibilities
Imagination can begin with a possibility we hope to reach, or with something we cannot yet explain. The question, its representation and the way research is done can all have alternative candidates.
What other possibilities appear if a premise changes?
Ask candidates to make distinguishable predictions. Keep untestable possibilities explicitly unknown.
The cycle is represented by a person. Every capability goal still needs independent experiments; this scene cannot establish it.
Distinguish the vision, implementation and evidence before deciding which experiment deserves to come next.
Inspect the current evidenceRead the planned fair comparisonTracing, replay and restoration provide a starting point that can be checked. Whether they support autonomous discovery, problem redefinition or lasting changes in capability depends on the actual scope of each record.
The delivery summary records 4,555 CPU regression tests, with 0 failures, 0 errors and 0 skipped tests.
Scope and limitationsPassing tests does not establish autonomous research capability; the credit loss remains a synthetic contract.
v0.9.2 DELIVERY_STATUS.json / scope.full_regression
A limited CPU transaction checked the saving of master weights and momentum, and next-step consistency after continuation in a new process.
Scope and limitationsThis is not acceptance of a cold start from a single file for the complete integrated True model.
v0.9.2 DELIVERY_STATUS.json / new_process_next_step_equal
The same True model was restored on an A100. Read-only checks covered 47 master-weight states and 20 existing momentum states, containing 1,067,974 selected parameter values.
Scope and limitationsThis run included no forward pass, loss, gradient update or new checkpoint. It does not demonstrate that the new backend has completed real-model learning.
next_learning_v1 INDEPENDENT_ACCEPTANCE.json
The README records a 2-token prefill, 1-token continuation decode and 6 gradient comparisons.
Scope and limitationsThe raw evidence was not rerun while preparing this website. Outer linear projections still use native PyTorch; full-model integration remains to be validated.
v0.9.4 README_zh.md / acceptance scope
The README records a trial with 456 synthetic parameters connecting Python backward computation, AdamW, saving and restoration in a new process.
Scope and limitationsTraining the real True model with the new backend has not passed acceptance. A synthetic trial cannot substitute for real-model validation.
v0.9.4 README_zh.md / acceptance scope
Autonomous problem formation and redefinition, new explanations, and the retention of reproducible improvements in research capability remain the core goals to be validated.
Scope and limitationsWe cannot currently claim that the entire base model runs in Python, that all internal thinking is visible, or that unlimited autonomous evolution has been achieved.
v0.9.2 autonomous_research_improvement_proven=false; v0.9.4 full integration remains to be validated
Local delivery summaries through 2026-09-29; model tests were not rerun while preparing this website. “Validation” here refers to project acceptance records, not independent academic certification. Model weights, private files and complete raw execution data are not published.
Download the public progress summary JSONA planned experiment specification. The four-group comparison has not been run. This section presents the verification method and conditions that would refute the expectation; the human-designed operator-routing and temperature-control illustrations on this website are not results from this experiment.
Help design the first fair experimentUse the existing mechanism under test, without constructing another research architecture for the website or substituting a human-scripted workflow for Xiaomai's capability. Before running it, specify the question, original purpose, available operations, stopping conditions, and assessment method.
Address the problem directly under the shared environment and rules, as the baseline.
Not yet runAllow additional reflection or revision within the same budget constraint, and record the actual cost.
Not yet runFirst inspect and name the existing mechanism under test, then test whether it produces reproducible changes.
Not yet runRemove the core mechanism under test and inspect whether the difference remains.
Not yet runAssess discovery, executable change, and persistent retention separately. Research into problem formation cannot be reduced to a score on the final answer.
Identify what is worth questioning and why, distinguishing suspicious signals, candidate causes, and missing evidence.
Locate differences in the representation, variables, dependencies, or mechanism while retaining the original purpose. Rewording an explanation is not enough.
Hold out new conditions and related problems from mechanism tuning, then check whether the effect carries over and whether prompting is required.
Inspect the state before and after saving. Recovery after restart, selection without prompting, and regression of existing capabilities are separate checks.
If core ablation produces no difference, improvement cannot be attributed to that mechanism. Failure on a different problem, inability to select it after restart, and regression of existing capabilities must all be recorded honestly. Compare reproducible reliability, cost, and transfer benefits; one win does not establish general superiority.
The plan is to provide methods, settings, observation scope, version differences, failures, and counterexamples for others to inspect and rerun. This is a specification awaiting execution, not evidence of external validation.
The expectation to be tested is whether Xiaomai can find what is worth questioning, turn a new understanding into an executable change, and continue using it on subsequent problems. No difference, counterexamples and regressions must also be recorded as results.
Original proposed target, for discussion
NT$ 900,000This is neither an amount raised nor a confirmed cost quotation.
Design and check the experiment togetherThe following is a specification for an experiment that has not yet been run. Use the same model and information access, and control or disclose the computation budget. Do not reveal the cause in advance, but provide observations and actions sufficient to identify it.
If removing the mechanism makes no difference, the improvement cannot be attributed to it.Read the full specification for the first experiment
Baseline performance using the same base model.
Check the effects of additional reflection or computation.
First inspect and name the existing mechanism to be tested.
Remove the core mechanism and check whether the improvement remains.
These proportions are a draft allocation for discussion, not committed or incurred expenses. Cost estimates, duration, supporter benefits and risks will be completed before any formal fundraising.
Start with the smallest environment that can be validated; use the results to decide whether to expand computing resources.
Controls, ablations, retention checks and restoration checks.
Prepare experiment records, failure cases and research summaries that others can inspect.
Service costs, fees and uncertain expenses.
The research may fail or fall short of expectations. This website is not an investment solicitation and provides no guarantee of returns, profits or a completion date. No payment feature is live.
These optional illustrations explain a minimal principle. A person designed the arithmetic and controllers, and they execute in the browser. They are not Xiaomai's autonomous research results or observations of a complete model's internals.
Xiaomai's research direction · A human-designed interactive illustration
Not another answer in different words: trace an actual execution, isolate a questionable component, challenge how it was defined, then put the modification back into the computation.
Long term, we aim to use foundational equations and functions across basic disciplines to represent, inspect and modify Xiaomai's own mechanisms and understanding of problems. The achievable scope remains to be tested.
Every mechanism, modification and criterion below was written by a person in advance. No Xiaomai model is called. This lets you inspect a path; it does not claim Xiaomai has followed it autonomously.
3 + 5Sum or product is part of the request.
(left, right) → productThe old definition receives only the number pair and ignores intent.
3 × 5 = 15Multiplication itself is correct; selecting it for this sum violates the request.
The result conflicts with the request. Inspect selection and execution before asking for another answer.
The current explicit request is 3 + 5, with reference value 8. The original ignores operation and always selects multiplication: 3 × 5 = 15. The original conflicts with this request; inspect the route from intent to operator.
The original is currently loaded; its output is 15 and conflicts with the request. The reference and criterion come from the explicit request and human-written arithmetic, not a discovery by the model.
The issue is not simply a wrong answer. How was the request represented? Which component selected the operation? Did the selected arithmetic execute correctly according to its definition?
Changing an operator solves this example. It does not establish a change in the mechanism that forms problems, autonomous evolution, or new science. Xiaomai's research must test whether it can itself find a questionable definition, propose an executable change, and retain effective capability under fair, unprompted and restart checks.
Imagine a different self, then ask why its components were defined that way. A modification must be executable and open to refutation; only supported changes deserve to be retained.
Imagine different possibilities: this is a capability Xiaomai aims to investigate, not a result established by this illustration.Explore the full research cycle
This page uses an explicit operation and bounded numerical values to illustrate request representation, operator selection, arithmetic and a human-written reference check. sum = left + right; product = left × right. It covers only these human-written functions, not language understanding, neurons, real-model learning or complete internal thinking.
Python is currently a tool for execution and inspection. The long-term aim is to use foundational equations, functions, basic assumptions, applicability conditions and dependencies across basic disciplines to progressively represent, inspect and modify Xiaomai's own mechanisms and understanding of problems. The achievable scope of representation and modification remains unknown. This does not mean every discipline can be fully expressed as equations, that the real model has been completely decomposed, or that a root cause is guaranteed to be found.
This website illustration executes in the browser in TypeScript. Reloading an in-memory version is different from saving knowledge or parameters, restarting a real model, transferring to hidden new problems, or retaining a lasting capability improvement.
Inspect the same 3 and 5 through execution, equations and the request. Multiplication can be correct while the route answers the wrong question. This is a public illustration written in advance by a person, not internal source code extracted from a real model.
sum(3, 5); the human-defined expectation is 8. All three views share this request. The product control changes only operation; the numbers and both operator formulas stay the same.
This page calls executeInspection and reads its outputs and traces. The original route ignores the sum request and selects multiplication; the candidate uses an explicit operation to select the operator. The code executes in TypeScript. Its requests, candidate and assessment rules are all written in advance by a person.
const request = { left: 3, right: 5, operation: "sum" };
const original = executeInspection(request, "original");
const candidate = executeInspection(request, "candidate");
const product = executeInspection(
{ ...request, operation: "product" }, "candidate"
);
const interventions = inspectInterventions(request);The records below are returned by these illustrative function calls. They are execution records from the example, not an invented run or a model's thinking process.
[
{
"part": "requirement",
"input": "{\"left\":3,\"right\":5,\"operation\":\"sum\"}",
"output": "3 + 5 = 8"
},
{
"part": "resolver",
"input": "{\"left\":3,\"right\":5}",
"output": "product"
},
{
"part": "arithmetic",
"input": "3 × 5",
"output": "15"
},
{
"part": "check",
"input": "actual=15; reference=8",
"output": "mismatch"
}
][
{
"part": "requirement",
"input": "{\"left\":3,\"right\":5,\"operation\":\"sum\"}",
"output": "3 + 5 = 8"
},
{
"part": "resolver",
"input": "{\"left\":3,\"right\":5,\"operation\":\"sum\"}",
"output": "sum"
},
{
"part": "arithmetic",
"input": "3 + 5",
"output": "8"
},
{
"part": "check",
"input": "actual=8; reference=8",
"output": "match"
}
]The original route calculates 3 × 5 = 15 correctly, but the request is for a sum. The candidate does not hard-code 8: a product request still executes multiplication. Ablating the single routing replacement returns to fixed multiplication and reproduces the sum error. These conclusions cover only this human-written program and the specified conditions.
[
{
"part": "requirement",
"input": "{\"left\":3,\"right\":5,\"operation\":\"sum\"}",
"output": "3 + 5 = 8"
},
{
"part": "resolver",
"input": "{\"left\":3,\"right\":5}",
"output": "product"
},
{
"part": "arithmetic",
"input": "3 × 5",
"output": "15"
},
{
"part": "check",
"input": "actual=15; reference=8",
"output": "mismatch"
}
]sum(x, y) = x + yproduct(x, y) = x × yThe two operators and numbers do not change. The original resolver omits the requested operation from its input and always routes to product. The candidate includes operation as its basis for choosing sum or product. The dependency from meaning to execution changes, not the multiplication formula.
(left, right)resolve(left, right) → product(left, right, operation)resolve(left, right, operation) → operationA person has explicitly labelled sum and product in this example. How a reliable request is obtained, whether the original problem is sound and whether ambiguous meaning is understood correctly require separate checks. A formula or route does not resolve those questions.
The original purpose is still to calculate the requested operation on the same numbers. Changing a sum request into a product request or displaying 8 alone does not demonstrate autonomous problem solving. This example shows how an external observer can inspect a public operator route.
A person has already disclosed the sum request and fault location here. This does not validate Xiaomai autonomously finding a contradiction without a cause hint, identifying a flaw in the original problem itself or proposing a new executable definition.
This fixed illustration does not validate model learning, weight updates, fresh-process restoration or selection without prompting. Whether Xiaomai can modify and independently validate its own problem formation, experiment selection and learning mechanisms remains a long-term goal.
No. Equations, conditions of applicability, the source of a request, representations and mechanisms for forming problems can themselves be questioned. We do not assume that every discipline can be fully expressed as equations.
A heating command is still on its way, yet the current temperature already determines the next action. This temperature-control illustration compares a state containing only the present temperature with one that also represents pending commands.
Both controllers are designed in advance by a human, and the delay is explicitly disclosed. This does not call the Xiaomai model and is not evidence of autonomous discovery or a result from the four-group research experiment.
The original purpose stays the same: bring the temperature close to 35°C. Adjust the command delay and observe two curves in the same environment.
120 time steps, the same initial state, and the same command limits. Drag or use the arrow keys to adjust; download the values for every step below.
In this toy environment, including commands that have not yet taken effect can reduce error. This demonstrates a human-designed difference in representation; whether an AI can discover, test, and retain this change on its own requires a separate experiment.
Want to inspect the curves? Take the values for each step and recalculate them.
Research does not ask for belief in advance.
Questions worth pursuing should be open to inspection.
Longer term, we also want to inspect, modify and independently test how Xiaomai forms questions, chooses experiments and learns. This has not yet been demonstrated.
Download the Chinese and English research briefYes. There is already considerable work on self-improvement, program evolution and autonomous scientific research. Xiaomai aims to study the complete cycle: imagine different possibilities, identify contradictions or gaps, question assumptions and problem definitions, propose new understandings or mechanisms, then use derivation, experiments and attempts at refutation to retain effective changes and keep exploring. The ways it selects problems, forms hypotheses and plans experiments are also long-term targets for reflection. The direction alone does not establish novelty; a contribution must still be explained through methods, controls and reproducible results.
A fair comparison is needed: use the same model and information access, control or disclose computation budgets, and separately test direct answers, additional reflection, the Xiaomai mechanism and core ablation. Then separately check actual changes to a representation or mechanism, restoration after restart, selection on new problems without prompting, and regression in existing capabilities. Failures, no difference, or equally effective results after removing the core mechanism must all be retained. The temperature-control demonstration on this website is not that four-group experiment.
Python is a temporary tool for execution and inspection. The long-term aim is to use foundational equations, functions, basic assumptions, applicability conditions and dependencies across basic disciplines to progressively represent, inspect and modify the system's own mechanisms and understanding of problems, making unknowns and invalid relationships traceable. The achievable scope of representation and modification remains unknown. This is a research direction; it does not mean that every discipline can be fully expressed in formulas or that the real model has been completely decomposed. Going deeper gives us more precise things to question, but does not guarantee that a root cause has been found. The formulas, representations and ways of forming problems can themselves still be questioned.
The inspiration from thought experiments is a question: when a path fails, can we return to assumptions that were taken for granted and examine them? First imagine possibilities, then ask whether their conditions hold, and finally submit them to derivation and experimental testing. When information is insufficient, the next required measurement should also be identified. This is not a claim of equivalent ability or a promise to produce another theory of relativity.
Formal fundraising has not begun, and this website has no payment feature. The plan is to first publish the mechanism to be tested, fair comparison specifications, acceptance checks, cost estimates, proposed duration, risks and deliverable materials before deciding on a support plan. TWD 900,000 is a draft for discussion, not an amount raised or the cost of completing the work. There is no promise of returns, equity, model capability or certain success. For now, we can discuss research collaboration or resource support.
The founder graduated in Computer Science and Information Engineering from National University of Kaohsiung and is pursuing this as an independent research project. There is currently no registered company. Contact with an institution, participation in a competition or an invitation to collaborate is not treated as endorsement. If formal collaborations are established, their scope will be clearly stated after consent has been obtained.
Help design a fair comparison, inspect existing records, or bring a counterexample that could overturn our expectations.
Design the first fair experimentSelect the mechanism to test, information access and refutation criteria together.
Independent review and reproductionInspect records, methods and reproducibility within the published scope.
Bring domain problems and counterexamplesBring real constraints, unknown conditions and different understandings of a problem.
The founder graduated in Computer Science and Information Engineering from National University of Kaohsiung and is pursuing this as an independent research project.
This website has no accounts or payment processing and does not use advertising trackers or website analytics cookies. The interactive illustrations compute in your browser. Research summaries come from a public API on this website; no personal model prompts are transmitted. The collaboration form only prepares an email draft or a text download. It does not store your submission on the website server. If you send the email, your email service and Gmail process it for replies and collaboration discussions. Please do not submit sensitive information. To request deletion of information you have emailed, contact us.
The hosting provider may process necessary connection logs for security and operations. The website developer has not added third-party tracking integrations. Research content may change; capability claims remain limited to each version’s verified scope. Website updated: 2026-10-01. Research-summary dates are shown in the public records.