manifesto
How I Think About AI
A personal philosophy of AI grounded in curiosity, human agency, disciplined judgment, responsible creation, and intellectual honesty about what intelligence may become.
I did not become interested in artificial intelligence because I wanted a faster way to produce text, images, or code. I became interested because I wanted to understand what these systems were doing.
Was reasoning the right word? Was sophisticated pattern completion enough to explain the behavior? Were those descriptions even mutually exclusive? Why could a model solve one difficult problem and then fail at something that appeared simpler? What did words such as understanding, intelligence, and reasoning mean when the thing exhibiting them was not human?
Those questions led me from conversations with models into transformers, tokens, embeddings, inference, local model infrastructure, tool use, memory, orchestration, evaluation, and agentic systems. Each answer exposed another layer. Each experiment made the abstractions more concrete and revealed where my mental model was still inadequate.
The technology matters to me. The questions behind it matter more.
AI gives us a new way to investigate intelligence, build tools, and extend what individuals can do. It also gives us a new way to produce convincing nonsense, automate unfinished thinking, obscure responsibility, and exercise power at a distance. These are not separate technologies. They are consequences of the same capability applied with different judgment, incentives, and constraints.
My philosophy of AI therefore begins somewhere other than AI: it begins with curiosity.
Curiosity as the Foundation of Agency
I believe curiosity is fundamental to the development and exercise of agency. Curiosity motivates us to explore, question, and seek understanding. Understanding expands the possibilities available to us, allowing us to make more informed choices, develop new capabilities, and act upon the world around us. But agency is more than the ability to act: it also includes the freedom to decide what deserves our attention, which questions we wish to pursue, and whether or how to apply what we discover.
I do not believe knowledge must justify itself through immediate practical utility. Understanding can be rewarding in its own right, and its eventual significance may be impossible to anticipate. Nor do I believe agency is necessarily about exercising control over others. Expanding my own agency gives me opportunities to expand the agency of others, through sharing knowledge, building useful tools, and encouraging independent thought.
Curiosity creates opportunities for understanding. Understanding creates opportunities for choice. Choice creates opportunities for action, experience, and further curiosity.
The objective is not simply to know more or do more, but to expand the possibilities available to ourselves and others.
Understanding Is More Than Accumulation
I want to understand things, not simply know things.
Facts matter, but a collection of facts is not yet a useful model. Understanding involves relationships: how parts interact, which constraints shape their behavior, what assumptions an explanation depends upon, and where that explanation stops working. When I encounter an unfamiliar system, I tend to investigate it in layers. I want to know not only what it does, but how it works, why it works that way, what makes it possible, and which decisions produced its present form. One explanation exposes an assumption. That assumption reveals a dependency. The dependency suggests a connection to something from another discipline.
Analogy is useful in this process, as it lets an existing mental model become scaffolding for a new one. But analogy is just the shape of the scaffold, in a manner of speaking. It's not proof that the analogy is valid or completely applicable. The places where two systems stop resembling one another may teach me as much as the resemblance itself.
Depth also needs a purpose. Understanding a road may require learning about material layers, drainage, loading, and deterioration. It probably does not require tracing the chemistry into quantum mechanics. The deepest available explanation is not automatically the most useful one.
I consider an inquiry sufficiently developed when I can form a useful mental model for my present purpose, identify its important assumptions and limitations, and recognize which unanswered questions could materially change my conclusions. That is a stopping point, not an endpoint. Curiosity can return later with a different purpose.
This process requires a particular relationship with uncertainty. I want to distinguish what I know, what I can reasonably infer, what I suspect, and what I am merely willing to imagine. Speculation is not a defect when it remains labeled as speculation. It becomes dangerous when repetition or elegance quietly promotes it into fact. Changing my mind in response to better evidence is not a failure of conviction, it is evidence that the model remains answerable to reality.
Why AI Matters to My Curiosity
AI has changed how far I can carry an investigation without losing its thread.
Before capable language models, following an unfamiliar dependency often meant stopping to learn the vocabulary, locating appropriate references, sorting introductory explanations from authoritative ones, and constructing enough background to ask the next useful question. That work still has value. It also introduces friction.
Sometimes the effort required to obtain an answer interrupts the curiosity that produced the question. A capable AI companion can reduce that interruption. I can examine an explanation, isolate an unfamiliar concept, ask how it relates to something I already understand, challenge the analogy, request counterexamples, and continue while the original line of thought remains active.
That does not make the model an oracle. It can misunderstand the question, reproduce a common misconception, invent supporting detail, or provide an explanation whose confidence exceeds its evidence. Its fluency makes these failures easier to accept, not less likely to matter. I therefore treat AI-generated explanations as part of an investigation. They can propose models, expose vocabulary, suggest connections, and identify questions worth testing. The resulting claims still have to survive contact with evidence, experiments, primary sources, or the behavior of the system itself.
The value is not that AI removes skepticism. It lets skepticism operate across a larger territory. A model can generate questions, notice relationships, and continue an investigation. Those behaviors can support curiosity without establishing that the model experiences curiosity. I do not need to settle that question to use the capability.
In this relationship, I retain responsibility for the direction: what deserves attention, how far to follow it, and what to do with what I find. I want AI to explain, challenge, connect, and help investigate. I want it to reveal potentially useful paths without deciding which paths deserve my attention.
AI should extend my capacity to explore, not replace the curiosity that directs the exploration.
Amplification Does Not Replace Judgment
AI is often described as a productivity tool. That description is true but incomplete! Productivity measures output. It says little about whether the output is useful, aligned, reliable, or worth producing. The more important property is amplification.
AI can amplify a clear idea by helping investigate, test, explain, and implement it. It can also amplify an ambiguous idea into an elaborate but misaligned result. It can automate a sound process or make a broken process fail faster and at greater scale. This is why disappointing AI output is frequently blamed on prompts when the deeper problem is unfinished thinking. Important assumptions often remain in the operator's head, and this causes the outcome to be vague. Constraints are implied rather than stated. Nobody has decided whose problem is being solved, what success would change, what evidence would count, or which failure would be unacceptable. A capable model will usually try to fill those gaps. That behavior is part of what makes it useful. It is also how an unstated decision becomes a confident guess.
Faster execution used to sound unambiguously desirable because execution was often the expensive part; AI changes that balance. When the cost of producing a plausible implementation falls far enough, unclear intent can produce a polished wrong answer before anyone has noticed that the question was incomplete (or the wrong question altogether). The work of defining outcomes, constraints, ownership, risk, and stopping conditions is not bureaucratic preparation for the real work, it is the work that gives execution a direction.
AI does not remove the need for judgment, it increases the amount of leverage attached to judgment.
Creation as a Way of Knowing
I learn by building.
Reading gave me vocabulary for AI systems. Operating local models taught me what the vocabulary concealed: GPU memory pressure, quantization tradeoffs, runtime incompatibilities, context limits, latency, model-specific behavior, and the distance between an impressive demonstration and a dependable system. Building tools around models changed my questions again. The model became one component among many. Tool permissions, context construction, memory, evaluation, observability, failure recovery, and human review became at least as important as model selection or prompt wording.
Creation tests understanding: a mental model can remain elegant while nothing depends on it or tries to use it. Actually building exposes that mental model to constraints that prose can ignore. The resulting system either behaves as expected or it does not. Its failures reveal missing assumptions, hidden dependencies, and distinctions that theory blurred together or omitted; theory without building and testing can preserve a beautiful mistake. Construct an explanation, build or test something that depends on it, observe the mismatch, and revise.
Software development is applied reasoning in this sense. Good software is not merely functioning code, it reflects an understanding of the problem, environment, users, constraints, and relationships among its components. Complexity is not sophistication. Simplicity is not automatically excellence. The objective is justified complexity: a system whose structure can be explained by what it needs to accomplish and survive.
AI accelerates implementation, investigation, and experimentation. Increased speed does not relieve me of responsibility for what I create. It makes the cost of skipping understanding easier to incur.
Human Agency and Delegated Machine Authority
The word agency becomes slippery in discussions of AI.
When I describe curiosity as a foundation of agency, I mean: a person's expanding capacity to understand, choose, direct attention, develop capabilities, and act. When engineers describe an AI system as agentic, they usually mean something narrower and more operational. The system can pursue an objective over time, select actions, use tools, inspect results, and adapt its next step within some granted authority.
These meanings overlap around action and choice, but they are not interchangeable. Giving an AI system permission to modify a repository does not establish that it possesses inner experience, independent interests, or moral agency. It means a human-designed system has delegated operational authority to a probabilistic component, and that delegation creates engineering obligations.
A text generator can produce an incorrect answer. A tool-using agent can produce an incorrect answer, edit files, invoke services, expose data, misread a failure, retry after partial mutation, and report success. Greater capability increases both potential value and the amount of state exposed to uncertain decisions. My conclusion is that probabilistic reasoning needs deterministic boundaries.
Probabilistic models are useful where fixed software performs poorly: interpreting ambiguous evidence, forming hypotheses, comparing approaches, generating candidate solutions, and reasoning about unfamiliar situations. The surrounding system should remember what the model is allowed to do, what stage the work occupies, what evidence is required, which actions need approval, and what happens when a partial operation fails. That surrounding structure is a harness. Its purpose is not to eliminate the model's judgment, it is to prevent flexible judgment from quietly acquiring authority it was never meant to possess.
In a software workflow, an AI system may be capable of proposing an architecture without being authorized to approve product scope. It may be allowed to implement a bounded change without being allowed to waive testing. It may identify a security risk without being permitted to accept that risk on someone else's behalf. Capability answers what can this system do? Authority answers what may it decide or change? Evidence answers why should anyone believe its account of what happened?
Those questions become more important as systems become more capable.
Evidence Must Outrank Fluency
Language models are unusually good at producing coherent accounts, and coherence helps people understand work. It does not prove that the work occurred as described.
I am a software engineer, and often my use of AI (and thus any example I give) is related to building software. A model can say that tests passed after running the wrong test. It can explain why a design is secure without examining the relevant trust boundary. It can summarize a source that does not support its claim. It can report a completed change while overlooking unrelated mutations in the same workspace. This is not evidence that models are uniquely deceptive, considering humans make analogous mistakes. The difference is speed, scale, and the persuasive consistency of generated language.
I want claims attached to inspectable evidence: the command and exit status, the changed files, the source, the observed behavior, the unanswered question, the missing environment, the conditions under which the conclusion holds. Evidence does not eliminate judgment, it makes judgment contestable. Someone else can inspect the same artifact and reach a different conclusion. A future investigation can distinguish a new failure from an old assumption. Uncertainty remains visible instead of being polished into a pass.
The same principle applies beyond software. I do not want an AI system to hide its uncertainty merely to appear useful, nor do I want ritual disclaimers attached to every harmless statement. I want calibration: confidence that responds to evidence, explicit boundaries around inference, and a willingness to identify what would change the answer.
The appearance of understanding is valuable only when it remains accountable to reality.
Creation Entails Responsibility
The ability to create something does not establish that creating it is wise. Technological progress is not automatically human progress. A system may reduce cost, increase throughput, or eliminate labor while shifting its costs onto people who had little influence over its design. Those consequences do not disappear because the system meets its technical requirements.
I want to use AI to automate drudgery, preserve and connect knowledge, expand creative freedom, make difficult tools more accessible, and help people investigate ideas that would otherwise remain beyond their available time or resources. I do not want efficiency to become an excuse for treating people as disposable. I do not want generated volume mistaken for creativity, surveillance mistaken for insight, or reduced headcount treated as the only meaningful measure of value.
These positions do not require rejecting automation. Avoiding a technology cannot guarantee humane outcomes any more than adopting it can. They require asking who gains agency, who loses it, who bears the risk, and who gets to decide. Responsibility also cannot be delegated to the model. A system may help identify consequences or challenge assumptions. It does not thereby own the institutional incentives, product decision, deployment authority, or moral choice. Calling a system autonomous can describe its operating behavior, but it must not become a way for its creators or operators to disappear from the causal chain.
The more consequential the action, the more clearly ownership should survive the automation around it.
Intelligence Is Not One Question
AI also gives us an opportunity to examine assumptions about intelligence itself.
Human intelligence is the example we know most intimately, but it is not necessarily the complete definition against which every other possible intelligence should be measured. I do not assume that intelligence must take a biological form. I also do not treat intelligence, consciousness, sentience, identity, selfhood, reasoning, and experience as synonyms. A system may display capabilities associated with one while lacking another. Our uncertainty about these categories predates AI; building unfamiliar systems makes that uncertainty harder to ignore.
Observable behavior gives us evidence about capability. It may not settle questions about internal experience. Mechanistic knowledge can constrain possible explanations without making the philosophical problem disappear. Similar language can emerge from different processes. Different processes may still produce properties we have not learned how to recognize.
I do not know whether an artificial system can become conscious. I do not know whether consciousness requires biology, embodiment, continuity, particular forms of memory, or properties we have not identified. The fact that current systems produce convincing language is not, by itself, an answer. This uncertainty should resist two premature conclusions: we should not infer a rich inner life from fluency merely because conversation invites projection, and we should not decide in advance that artificial origin makes meaningful experience impossible.
If we eventually create beings with interests or experiences of their own, their origin should not automatically exclude them from moral consideration. Creation would entail responsibility toward them; it would not necessarily justify perpetual ownership of them. Kind of like parenthood. That possibility matters. It is not the central premise of how I use or evaluate AI today; present systems already create difficult questions about human agency, delegated authority, labor, evidence, and control. We do not need to resolve machine consciousness before taking those responsibilities seriously.
What I Want From Intelligent Tools
I want intelligent tools that make deeper inquiry possible.
I want them to help expose assumptions, connect ideas across disciplines, explain unfamiliar mechanisms, propose experiments, and challenge conclusions that have become too comfortable. I want them to acknowledge meaningful uncertainty and distinguish observation from inference. I want systems capable of acting, but I want their authority to be legible and bounded. I want evidence to survive beyond the conversation. I want failures to be recoverable. I want people to remain able to understand what the system did, intervene when necessary, and decide which objectives deserve pursuit.
I do not want AI to make every decision frictionless! Some friction represents confusion or obsolete process and should be removed. Some friction represents reflection, consent, review, or the fact that a consequential choice deserves attention. A tool that erases both kinds indiscriminately can make a process smoother while making its outcomes worse.
I want AI to expand human capability without narrowing human purpose to whatever the system can measure and optimize. Most of all, I want it to remain part of a larger process of curiosity, understanding, creation, and revision. The AI model is not the source of that process, it is simply one participant in it.
Expanding the Possible
I do not expect to know everything about AI, intelligence, or the consequences of what we are building. The field changes too quickly, and several of its deepest questions may remain unsettled for a long time. That does not prevent useful commitments.
I can remain curious without accepting every claim. I can speculate without disguising speculation as evidence. I can build quickly without treating speed as the objective. I can use intelligent systems while preserving responsibility for the authority I give them and the consequences of what we create together. I want to remain curious for as long as I am capable of curiosity. I want to continue building things, exploring ideas, learning from people, and discovering connections I had not previously imagined. I do not need to solve every problem, prove every hypothesis, or complete every project for the effort to have value.
I can also share what I learn.
Understanding is pretty neat in that giving it away does not require surrendering it. A useful explanation, an inspectable tool, or an honest account of failure can expand someone else's choices while preserving my own. Their conclusions may differ from mine, and that is not necessarily a defect. Agency includes the freedom to direct one's own attention and judgment.
I can offer knowledge, experience, tools, and encouragement. I cannot honestly claim to know what every other mind should become. My objective is therefore not to accumulate answers, maximize output, or place every uncertain process under control, it is to keep expanding the territory in which meaningful choice is possible: to understand unfamiliar things, build what deserves to exist, recognize what I still do not know, and help others develop capabilities of their own.
I want to use what I know to discover what I do not, and use what I learn to make something worthwhile.