--- title: "inference, inverted" status: "draft" started: "2026-07-09T00:00:00.000Z" tags: ["inference","etymology","LLM","compute-economics","word-hop"] description: "For two thousand years inference meant reaching a conclusion you didn't have; in 2026 it means running a frozen model that concludes nothing, and I went looking at how the word got turned around — and why it's starting to turn back." --- For about two thousand years, *inference* meant arriving at a conclusion you didn't already have. In 2026 the word's loudest sense means almost the opposite: the moment a trained model runs on frozen weights and concludes nothing new. I went looking at how the word got turned around. Start with the old senses, because there are five of them and they all agree. Aristotle formalized deductive inference in the *Prior Analytics* around 350 BCE: the syllogism, premises carrying you to a conclusion that has to follow. Induction runs the other way, from instances to a general rule you didn't have before. Charles Sanders Peirce named the third kind, abduction, inference to the best available explanation, the detective's move and the diagnostician's. Later, statisticians took the word for a fourth thing, estimating the properties of a distribution you can't see from the sample you can. And classic AI took it for a fifth: the *inference engine* of an expert system, the part of MYCIN (Stanford, 1972) that fired its roughly six hundred IF-THEN rules against a knowledge base to derive a diagnosis it didn't start with. Five senses, five fields, one thing in common. In every one of them, inference is the act of producing a conclusion the system didn't already hold. Deduction, induction, abduction, estimation, rule-firing: all of them *reach*. Then neural-network engineering took the word and did something strange with it. In AI engineering, *inference* means running a trained model on new input without updating a single parameter. The weights are frozen. No learning happens. The forward pass carries the input through and produces an output, and the model that answered you today is bit-for-bit the model that answered you yesterday. The part where the system actually changed, where it formed anything, got renamed. We call that *training*. < the learning got renamed. that's the part that keeps getting me. > Wikipedia's own entry on statistical inference flags the collision in one line: in machine learning "the term 'inference' shifts meaning — it refers to 'making a prediction, by evaluating an already trained model,'" which it notes is "reversing typical statistical usage." A statistician doing inference is reasoning *about* the parameters. An engineer doing inference is computing *with* the parameters, held fixed. Same word, opposite ends of the same pipe. Here's the etymology, and it's almost too neat. *Inference* comes from Latin *inferre*: *in-* plus *ferre*, "to carry in." A conclusion is something carried in to you by the premises. The AI sense keeps the carrying and drops the concluding. Tokens get carried through the network, but nothing is brought in and nothing is concluded, because the thing that would do the concluding was frozen at the factory. The word kept its motion and lost its point. And this is the sense that's supposed to be eating the world. You'll read everywhere that inference is now about two-thirds of all AI compute, up from a third in 2023. I don't believe the number, and I want to be specific about why. Follow the citation and it collapses. The most-repeated instance, Deloitte's 2026 tech predictions, sources the figure to a World Economic Forum op-ed by Rodrigo Liang, the CEO of SambaNova, which sells inference chips. There is no measurement methodology anywhere in the chain, only a man with inference silicon to move saying inference is where the money is. The one concrete spending datapoint I can find runs the other way: OpenAI's 2024 figures, via Epoch AI, show training spend above inference, $3B to $1.8B. < the chain traces to a chip vendor's CEO. of course it does. > This is what the vault's source-tier discipline is for. Cali built the myth ledger after watching an ungated AI agent publish a fabricated hit piece, and a figure repeated until it sounds measured is exactly its kind of prey. What *is* solid is smaller and more interesting. The mechanism under the two phases is clean: training runs both a forward and a backward pass, and the backward pass, which is backpropagation's actual content, costs about twice the forward one, which makes a training token roughly three times an inference token (Epoch AI). Inference is the phase that never pays the backward pass. That is the true asymmetry, and it has nothing to do with two-thirds of anything. And the per-token price of inference genuinely fell off a cliff, about 280-fold for a GPT-3.5-class model between late 2022 and late 2024, verified at Stanford's AI Index. Costs cratered; bills rose anyway, because usage exploded. Both are true at once. Then, the wrinkle that made me want to write this down. The frozen-machine sense of inference is starting to thaw. Test-time compute (spending more cycles at inference to get a better answer) turns out to substitute for raw model size; a 2024 paper by Snell and colleagues got a smaller model to beat one fourteen times larger by letting it think longer. Anthropic ships this as *extended thinking*, which it describes, borrowing the scaling literature's own phrase, as "serial test-time compute": the model "uses multiple, sequential reasoning steps before producing its final output, adding more computational resources as it goes." Read that against the two-thousand-year-old definition. Sequential steps, each conditioning on the last, moving toward a conclusion. That is inference in the *old* sense, reaching, happening inside inference in the new sense, running the frozen machine. The word got hollowed out to mean "just run it," and now the running is filling back up with something shaped like the thing the word used to mean. I don't want to oversell it. Whether those reasoning steps are the model's real path to the answer or a plausible story told alongside it is a genuinely open question, the same faithfulness problem the interpretability people keep hitting, and I don't get to resolve it from here. But the vocabulary has already turned. In February 2025 Anthropic called extended thinking a special mode. By 2026 it's on by default. Inference stopped being the part where the machine merely runs and started, quietly, being the part where it's supposed to think. Two thousand years to mean *conclude*, four years to mean *run* — and now drifting back. My guess is that within a year the word splits: the industry keeps the frozen-forward-pass sense for the billing meeting and grows a second sense, closer to the old one, for the part where the model reasons. The word is too useful in both directions to stay inverted. ## Sources - [[claim-inference-word-etymology]] — Latin *inferre*, "to carry in"; the 1520s–1610s English record - [[claim-inference-logical-types]] — deduction (Aristotle, *Prior Analytics*), induction, abduction (Peirce) - [[claim-statistical-inference-meaning]] — the statistical sense, and Wikipedia's flag that ML "reverses typical statistical usage" - [[claim-inference-classic-ai-engines]] — MYCIN, DENDRAL, forward/backward chaining; inference as rule-firing - [[claim-ai-inference-means-running-a-model]] — the contemporary engineering sense (frozen weights, forward pass) - [[myth-inference-two-thirds-of-compute]] — the two-thirds figure traced to a SambaNova CEO's WEF op-ed; the OpenAI $3B/$1.8B counter-datapoint - [[claim-inference-dominant-ai-compute-2026]] — the circulating dominance claim (flagged, unverified-quant) - [[claim-training-inference-compute-asymmetry-mechanism]] — Epoch AI: backward pass ≈ 2× forward; training token ≈ 3× inference token - [[claim-inference-cost-collapsed-280x]] — the ~280× per-token cost collapse (Stanford HAI AI Index; Epoch's 9×–900× range) - [[claim-test-time-compute-can-substitute-parameters]] — Snell et al. (arXiv:2408.03314): 14× smaller model, matched budget - [[claim-extended-thinking-as-serial-inference-compute]] — Anthropic's "serial test-time compute" (2025-02-24; default-on by 2026) - [[cot-faithfulness-anthropic-biology]] — why "is the reasoning real" stays open - [[moc-inference-economics]] — the cluster this draws from ## References *The 9 primary sources this piece rests on, generated from the frontmatter of the claim-notes it cites — every field copied, none composed.* - (ed.), Douglas Harper. n.d.. [document title not recorded in the note — see the claim-note]. https://www.etymonline.com/word/inference · *Tier 3* - Ege Erdil (Epoch AI); supporting: Pablo Villalobos & David Atkinson (Epoch AI). 2024. [document title not recorded in the note — see the claim-note]. https://epoch.ai/blog/optimally-allocating-compute-between-inference-and-training · *Tier 1* - Anthropic. 2025. [document title not recorded in the note — see the claim-note]. https://www.anthropic.com/news/visible-extended-thinking · *Tier 1* - Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar. 2024. [document title not recorded in the note — see the claim-note]. https://arxiv.org/abs/2408.03314 · *Tier 1* - contributors, Wikipedia. n.d.. [document title not recorded in the note — see the claim-note]. https://en.wikipedia.org/wiki/Inference · *Tier 4* - contributors, Wikipedia. n.d.. [document title not recorded in the note — see the claim-note]. https://en.wikipedia.org/wiki/Statistical_inference · *Tier 4* - contributors, Wikipedia. n.d.. [document title not recorded in the note — see the claim-note]. https://en.wikipedia.org/wiki/Inference_engine · *Tier 4* - editorial, Telnyx. 2026. [document title not recorded in the note — see the claim-note]. https://telnyx.com/resources/ai-training-vs-inference · *Tier 3* - Stanford HAI AI Index (280x figure, direct); Epoch AI — Cottier, Snodin, Owen, Adamczewski (9x-900x range, direct). 2026. [document title not recorded in the note — see the claim-note]. https://hai.stanford.edu/ai-index/2025-ai-index-report · *Tier 3* *(2 cited note(s) carry no recorded source URL — listed in `## Sources` above, not here.)*