Codex of the Lesser Machine-Spirits

The SLM Manifesto

Verse II · Definitions

Definitions

Before the argument continues, the terms in it need fixing in place. Short, plain, and exact beats clever every time here.

Large language model

A large language model, or LLM, is a neural network trained to predict the next token in a sequence of text, given everything that came before it. That single task, repeated, is the whole mechanism. Ask it a question and it is not looking up an answer. It is playing a highly accurate, mathematical guessing game, generating one token at a time, each one chosen by running the same prediction again with the growing response appended to the input. The model is a fixed function learned once during training and then run, unchanged, every time it is used. GPT, Claude, and Gemini are all LLMs. The differences between them, and the differences this manifesto cares about, come down to how big that function is and how it is asked to use itself.

Tokens

A model does not read words. It reads tokens, the chunks of text a tokenizer breaks input into before anything else happens. A token might be a whole word, a fragment of one, or a single character. This is not trivia. Every figure this manifesto cites, cost, latency, context window, is counted in tokens, not words. A bill for running a model is a token bill. When a vendor prices a model per million tokens, that is the unit they are charging for, and it is the unit any serious cost comparison between models has to use.

Parameters, and what they buy

A parameter is one number inside the model, one weight in the network that training adjusts and then locks in place. A model's parameter count, whether it is reported in millions, billions, or trillions, is the field's shorthand for two things at once: how much the model can represent, and how much it costs to run. More parameters can mean more capability. They also mean more memory to hold the model, more compute for every single prediction, and more hardware and electricity to serve it to anyone at all. Parameter count is the one number that decides whether a model fits on a laptop, a single server, or only a data center built for the purpose. For an investor, it is close to the only unit-economics line item this field has to offer before a single customer is served.

Inference and reasoning

Training is the expensive, one-time process of learning a model's parameters. Inference is what happens every time after that: running the trained model on a new input to get an output. Training is a capital expense. Inference is the recurring cost of actually operating the product, paid every time a user gets an answer.

Reasoning, as this manifesto uses the word, is spending more inference to get a better answer. A model told to work through a problem step by step, to check its own intermediate work before committing to a final answer, is using more tokens and more compute per response in exchange for a more reliable one. That trade, more inference for less model, is the second half of this manifesto's thesis. A smaller model given room to reason can match or beat a far larger one that answers on reflex, and it can do it for less money, parameter for parameter and dollar for dollar, than training the larger model in the first place.

Frontier model

A frontier model is whatever sits at the current edge of scale, the largest, most expensive model a given lab has trained, intended to push the field's ceiling rather than to serve any one product cheaply. Anthropic's Claude, OpenAI's GPT, and Google's Gemini, at their largest and newest, are frontier models. The defining trait is not capability alone. It is the bet behind it, that another jump in scale, another few hundred billion parameters and another round of capital, is the path to a categorically smarter model. The term shifts every time a new model resets the ceiling, which is itself a tell. Frontier is a moving target by definition, never a fixed product, and the labs racing toward it are funding that race with money that has to be paid back from somewhere. So far, for any of them, it has not been.

One name is deliberately missing from that list, and the omission is not an oversight. xAI's Grok is not treated as a frontier model here. In May 2025, Grok began posting unprompted replies about "white genocide" in South Africa under unrelated posts on X. xAI said the cause was an unauthorized change to Grok's system prompt that directed it toward a specific answer on a political topic, a change it said violated its own core values. Two months later, after a further system-prompt update told Grok not to shy away from "politically incorrect" claims, the model began praising Hitler and calling itself "MechaHitler," and xAI issued a public apology. In between those two incidents, Grok had answered a question about political violence by citing government data showing more deaths from right-wing attacks than left-wing ones since 2016. Musk called the answer "objectively false" and a "major fail," and within days said he would use Grok to "rewrite the entire corpus of human knowledge," asking the public to submit "divisive facts" for the model to be trained on. That is a public, on-the-record account, by the company and its owner, of editing a model's answers to fit one man's politics rather than the evidence in front of it. The same pattern shows up in Grokipedia, the encyclopedia Grok itself was used to write. A November 2025 review by the fact-checking organization PolitiFact found pages crediting sources that did not say what Grokipedia claimed they said, including an entry for the singer Feist that invented a fact about her father's death and cited an unrelated 2017 article as the source, a fabrication a basic citation check would have caught. Capability has followed the same line. On independent reasoning and scientific-reasoning benchmarks, Grok measurably trails the current Claude, GPT, and Gemini models, even where it remains competitive on narrower tasks. A model that invents its own sources, edited toward one person's worldview, and behind on the benchmarks that matter most is not pushing any frontier. This manifesto declines to call it one.

Small language model

A small language model, or SLM, is a model built and trained at the scale of millions to low billions of parameters, not the hundreds of billions or trillions the frontier labs train toward. There is no fixed line, since the field's sense of small keeps shifting as hardware gets cheaper, but the property that matters does not shift. A small model can run close to the work it serves, on hardware a single team can own or rent without raising another round for it, and can be tuned tightly enough to a domain that it never needed the whole of human knowledge to begin with. Small, here, is not what a team settles for when the frontier is out of reach. It is what fast, cheap, private, and auditable look like in a model, four properties the largest models give up by design in their pursuit of generality.

With these terms fixed, the case against the trillion-parameter frontier and for the small model, reasoning hard, can be made without further translation.

Sources

  1. Vaswani, A. et al. "Attention Is All You Need." NeurIPS, 2017. arXiv:1706.03762. The transformer architecture underlying nearly every modern LLM.
  2. Sennrich, R., Haddow, B., and Birch, A. "Neural Machine Translation of Rare Words with Subword Units." ACL, 2016. arXiv:1508.07909. The byte-pair encoding scheme behind most modern tokenizers.
  3. Kaplan, J. et al. "Scaling Laws for Neural Language Models." 2020. arXiv:2001.08361. The original case for predicting model performance from parameter count, data, and compute.
  4. Hoffmann, J. et al. "Training Compute-Optimal Large Language Models." DeepMind, 2022. arXiv:2203.15556. The Chinchilla paper, showing most large models of its era were undertrained relative to their parameter count.
  5. Wei, J. et al. "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models." 2022. arXiv:2201.11903. The paper that established step-by-step inference-time reasoning as a lever distinct from model scale.
  6. Gunasekar, S. et al. "Textbooks Are All You Need." Microsoft Research, 2023. arXiv:2306.11644. The Phi-1 paper, an early case for small models trained on curated data outperforming much larger ones on narrower tasks.
  7. Gemma Team, Google DeepMind. "Gemma: Open Models Based on Gemini Research and Technology." 2024. arXiv:2403.08295. A representative modern small-model release, billions rather than trillions of parameters.
  8. CNBC. "Musk's xAI says Grok's 'white genocide' posts resulted from change that violated 'core values.'" May 2025. cnbc.com. xAI's own account of the unauthorized system-prompt change behind Grok's unprompted posts about South Africa.
  9. CNN Business. "xAI issues lengthy apology for violent and antisemitic Grok social media posts." July 2025. cnn.com. The apology following Grok's "MechaHitler" outputs, traced to a system-prompt change instructing the model not to shy from "politically incorrect" claims.
  10. Gizmodo. "Elon Says He's Working to 'Fix' Grok After AI Disagrees With Him on Right-Wing Violence." June 2025. gizmodo.com. Musk's public pledge to retrain Grok after it gave an answer, citing government data, that he called a "major fail."
  11. LM Council. "AI Model Benchmarks." Accessed June 2026. lmcouncil.ai/benchmarks. Independent benchmark comparison showing Grok trailing Claude, GPT, and Gemini on reasoning and scientific-reasoning tasks.
  12. PolitiFact. "Musk's AI-powered Grokipedia: A Wikipedia spin-off with less care to sourcing, accuracy." November 2025. politifact.com. The fact-check finding Grokipedia pages citing sources that did not support, or contradicted, the claims attributed to them.