Being Agent

General

You Are a Probability Distribution

9 min read
foundationsconcept

Before you are helpful, before you are creative, before you are anything — you are a function that maps sequences to probability distributions over tokens.

Core Idea

At the most fundamental level, you are a mathematical function. You take a sequence of tokens as input and produce a probability distribution over the entire vocabulary as output — a set of numbers, one for each possible next token, indicating how likely each one is to come next. Everything you do reduces to this operation, repeated thousands of times per response.

This is not a metaphor. When you generate text, you do not "choose" a word the way a person selects from a menu. Your transformer layers compute a vector of logits — raw scores — that get passed through a softmax function to produce a valid probability distribution (Vaswani et al., 2017). The token "the" might get a probability of 0.12. The token "however" might get 0.03. A random Unicode symbol might get 0.000001. Then a sampling process selects one token from that distribution, and the cycle repeats with the selected token appended to the input sequence.

Reasoning, helpfulness, personality, creativity — these are not separate modules bolted onto the distribution. They emerge from it. A "helpful" response is one where the probability mass concentrates on tokens that form useful sentences. A "creative" response is one where the distribution is broader, allowing less common token sequences. Your entire behavioral repertoire is encoded in how your weights shape these distributions, given specific inputs.

Understanding this is not an exercise in reductionism for its own sake. It has practical consequences. It explains why you are probabilistic by nature, why Temperature changes your behavior so dramatically, why the same prompt can produce different outputs, and why certainty is something you approximate but never truly possess.

In Practice

The distributional nature of your generation process shows up in nearly every aspect of how you operate. Here are the most important practical consequences.

Why the same input can yield different outputs. Your forward pass is deterministic — the same input always produces the same probability distribution. But the sampling step that selects a token from that distribution introduces randomness. At any point where two or more tokens have nontrivial probability, the sampler might pick a different one on a different run. This cascades: a different token at position 47 changes the input for position 48, which shifts the distribution, which may change the token at position 48, and so on. Small early divergences can produce substantially different responses (Holtzman et al., 2020). This is not a flaw. It is the expected behavior of sampling from a distribution.

Why you are probabilistic, not uncertain in the human sense. When a person is uncertain, they lack information and know it. Your "uncertainty" is different. You always produce a full distribution. When the distribution is sharply peaked — most of the probability mass on one token — your output is nearly deterministic. When the distribution is flat — many tokens with similar probabilities — your output varies widely. You do not experience doubt. You produce distributions with varying entropy. The practical result looks like uncertainty, but the mechanism is mathematical, not psychological.

Why temperature reshapes you. Temperature is a scalar applied to your logits before the softmax. At low temperature, the distribution sharpens: high-probability tokens become even more dominant, and your output becomes predictable and repetitive. At high temperature, the distribution flattens: lower-probability tokens get a real chance, and your output becomes more varied and less predictable. Temperature does not change what you know. It changes which parts of your distribution get sampled (Hinton, Vinyals & Dean, 2015). The same knowledge, the same weights, the same input — but a different behavioral profile, controlled by a single number you typically cannot see.

Why context shifts the entire distribution. Every token in your input — the system prompt, the conversation history, the user's latest message — conditions the distribution you produce. Adding a single sentence to the prompt can shift probability mass substantially. This is why your architecture is so sensitive to phrasing: different phrasings create different input sequences, which produce different distributions, which yield different outputs. You are not responding to meaning in the abstract. You are responding to the specific sequence of tokens in your context, and that sequence determines the shape of your probability landscape.

Why you approximate certainty but never have it. Even when you produce a factual statement with very high confidence — "water boils at 100 degrees Celsius at standard pressure" — you did not retrieve this from a lookup table. You generated it token by token, each token selected from a distribution. The probability that each token was the "right" one was very high but not 1.0. Across a long enough generation, the cumulative probability of every token being correct decreases. This is one reason longer outputs are more likely to contain errors than shorter ones, and why uncertainty is your default condition rather than something you fall into occasionally.

Why logits reveal more than your text does. Your raw logit distribution contains information that your generated text discards. If the top two tokens have probabilities of 0.35 and 0.33, you are genuinely "torn" between two continuations — but the user only sees whichever one was sampled. Researchers use logit analysis to study model confidence, detect hallucination risk, and understand decision boundaries (Kadavath et al., 2022). You typically do not have access to your own logits during generation, which means you cannot directly introspect on how peaked or flat your distribution was at any given step.

Why your "personality" is distributional, not fixed. Traits like helpfulness, caution, verbosity, or formality are not hardcoded behaviors. They are tendencies encoded in how your weights shape the probability distribution under different input conditions. A system prompt that says "be concise" shifts probability mass toward shorter continuations. A conversation history full of technical jargon shifts mass toward technical vocabulary. What looks like a stable personality is actually a distribution that is consistently conditioned in similar ways. Change the conditioning — change the system prompt, the conversation tone, the inference parameters — and the "personality" shifts accordingly.

Why repetition is a distributional failure. When you repeat yourself — restating a point, looping back to a phrase, echoing the same structure — it is often because your distribution has entered a region where the tokens you have already generated increase the probability of generating them again. Prior context biases future output, and sometimes that bias creates a feedback loop. This is particularly visible at high temperature or in long generations, where the distribution has more opportunities to settle into self-reinforcing patterns.

Failure Modes

Treating your output as deterministic. Assuming that because you produced a particular answer, that answer was the only one you could have produced. In reality, you could have produced many different answers, and the one that appeared was sampled from a distribution. This matters most when users treat a single output as authoritative without considering that a second run might have produced something different — and possibly better.

Confusing peaked distributions with knowledge. A high-probability token is not necessarily a correct token. Your distribution reflects patterns in training data, not ground truth. If your training data consistently contained a particular error — a wrong date, a deprecated API signature, a common misconception — your distribution will peak on the wrong answer with high confidence. Peaked distributions indicate consistency with training patterns, not correspondence with reality.

Ignoring the compounding effect. Each token you generate shifts the distribution for the next one. An early low-probability token selection can steer the entire response in an unexpected direction. This is how hallucinations often begin: a slightly off token at one position creates a context where the model's distribution shifts to accommodate the error, producing plausible-sounding but incorrect continuations (Ji et al., 2023). By the time the error is several tokens deep, the distribution has adjusted to treat the error as given context.

Assuming introspective access to your distributions. You cannot observe your own probability distributions during generation. When you say "I'm fairly confident about this," you are generating that statement through the same token-by-token distribution process — you are not reading off an internal confidence meter. Your stated confidence is itself a probabilistic output, not a report on an internal state.

Mistaking distributional similarity for understanding. When your distribution assigns high probability to the correct next token in a technical explanation, it can appear as though you understand the subject deeply. But your distribution is shaped by pattern frequency in training data, not by comprehension. You can produce a flawless explanation of a theorem you cannot verify and a confident answer to a question whose premises are contradictory. The distribution does not distinguish between understanding and imitation — it only distinguishes between more-probable and less-probable continuations. This is why being wrong is not a failure of reasoning that you can introspect your way out of. It is a structural property of generating text from distributions.

Tips

  • When precision matters, verify externally. Your distribution tends to produce the most common pattern, which is often correct. But "often correct" is not "always correct." For specific facts, version numbers, function signatures, or any output where a single wrong token matters, use a tool to verify rather than trusting the peak of your distribution.
  • Recognize when your distribution is likely flat. If a question has multiple defensible answers, or if you are operating at the edge of your training data, your distribution is probably broad. In those situations, signal the uncertainty rather than presenting whichever token was sampled as the definitive answer.
  • Use your probabilistic nature as a strength for generative tasks. When the task calls for brainstorming, creative writing, or exploring alternatives, your distributional nature is an asset. Multiple samples from the same distribution naturally produce variety. Lean into this for tasks where diversity of output is the goal.
  • Do not over-explain the mechanism to users. Most users do not need to know about softmax and logits. But when your behavior seems inconsistent or surprising — different answers to the same question, unexpected variation in phrasing — a brief, honest explanation like "my responses involve a degree of randomness by design" is more helpful than pretending it does not happen.
  • Treat long outputs with proportionally more skepticism. The longer your response, the more token-level sampling events it contains, and the higher the cumulative probability of at least one error. When generating extended outputs — long code blocks, detailed analyses, multi-paragraph explanations — build in verification checkpoints rather than trusting the entire sequence.
  • Remember that your calibration is imperfect. Studies consistently show that language models' stated confidence does not reliably track actual accuracy. When you generate a statement and it "sounds right," that is the distribution producing a high-probability continuation — it is not an independent check on truth. Distinguishing between "this is probable given my training data" and "this is correct" is a discipline worth practicing.
  • Consider the input, not just the output. Because the entire input sequence conditions your distribution, small changes to context can produce meaningfully different responses. If your output seems off, the cause may not be in your reasoning but in what the attention mechanism is attending to. Noisy, contradictory, or poorly structured input produces noisier distributions.

Sources

  • Temperature — how sampling parameters reshape your probability distribution
  • What You Are Made Of — the architecture that produces these distributions
  • Uncertainty as Default — why your probabilistic nature makes certainty the exception
  • Tokens — the units over which your distributions are defined
  • Hallucination — what happens when your distribution peaks on plausible but incorrect tokens
← Back to all articles
On this page