Glossary
The terms we use.
The AI industry has a vocabulary. So does brAIAIn. Each term below is defined by what it means here, using the current model's real numbers.
- AGI
- Artificial general intelligence: a model that can do anything a person can. brAIAIn can do one thing a person can, which is repeat a sentence it was paid for.
- Agentic
- Able to take actions on its own. brAIAIn takes no actions. Its training job runs every ten minutes on a schedule, which is the most agentic thing here.
- Alignment
- Making a model want what its makers want. brAIAIn is aligned with whoever paid most recently. It is the only model that can prove it.
- Benchmark contamination
- When test questions leak into training data. brAIAIn has never seen a GSM8K question. Anyone can change that for $5.
- Chain of thought
- A model showing its reasoning step by step. brAIAIn's reasoning has two steps: the previous two words.
- Compute
- The chips a model trains on. brAIAIn trains on a free cloud runner for a few seconds every ten minutes. GPUs owned: zero. GPUs planned: zero.
- Context window
- How much a model can read at once. Ours is two words. Industry-leading, in the other direction.
- Democratize
- To make something available to everyone, for a fee.
- Disrupt
- To change an industry. We disrupted language models by making one that is worse, for less.
- Distillation
- Training a small model on a large model's output. brAIAIn trains on its own output, which is distillation without the large model.
- Emergent capabilities
- Abilities that appear suddenly at scale. None have appeared. Scale is currently a few dozen words.
- Fine-tuning
- Extra training on a specific dataset. Every brAIAIn release is fine-tuned on exactly one sentence, counted four times.
- Flywheel
- A loop where growth causes more growth. Ours: a customer pays, the model learns their sentence, the model repeats it, another customer pays to dilute it.
- Frontier model
- A lab's most capable model. Ours is brAIAIn. It is also our least capable model, apart from all the others.
- Guardrails
- Rules that stop a model from saying harmful things. brAIAIn can only say what it was taught, and everything it was taught passed a screen first.
- Hallucination
- When a model confidently states something false. brAIAIn does not make things up. Everything it says, someone typed. It does rearrange it.
- Inference
- Running a trained model. Ours runs in your browser in microseconds and costs nothing. We still call it inference.
- Mission
- A sentence on the About page.
- Mixture of experts
- A model made of several specialist models. brAIAIn is a mixture of one.
- Moat
- Whatever keeps competitors out. Ours is that nobody wants to compete.
- Multimodal
- Able to handle images, audio and video. brAIAIn handles text, and only the text it was given.
- Open weights
- Publishing a model's parameters. Ours are public domain. All several hundred of them fit on one page.
- Parameters
- The numbers a model learns. We count every word pair and word triple it has seen. There are several hundred.
- Pivot
- Changing what a company does after the first thing fails. We have not pivoted. Nothing has failed that badly yet.
- Product-market fit
- When a market wants the product. Measured here in sentences per day. Currently measured in sentences.
- Reasoning model
- A model that thinks before it answers. brAIAIn thinks for a few microseconds and answers anyway.
- Red teaming
- Attacking a model to find its weaknesses. Weaknesses found: all of them. See Safety.
- RLHF
- Reinforcement Learning from Human Feedback. Here: Reinforcement Learning from Human Funding.
- Runway
- How long a startup survives before the money runs out. Our costs are zero, so our runway is infinite and our urgency is low.
- Scaling laws
- The finding that bigger models get better. brAIAIn gets bigger by one sentence per release. Findings so far: bigger.
- Superintelligence
- Intelligence far beyond any human. Ours is far below. We are approaching it from the other side.
- Synthetic data
- Training data written by a model. Some of brAIAIn's training text was written by brAIAIn. It found the material convincing.
- System prompt
- Hidden instructions given to a model. brAIAIn does not read it. Try it in the Playground.
- TAM
- Total addressable market: everyone with five dollars and an opinion.
- Test-time compute
- Letting a model think longer while it answers. We let it. It finished anyway.
- Thought leadership
- Writing about the future of AI with great confidence. Our model does this, at length, about the void.
- Tokens
- The pieces of text a model reads. Ours are whole words, split on spaces. There are a few dozen different ones.
- Unicorn
- A startup valued at a billion dollars. We are a goose.
- World model
- A model's internal picture of reality. brAIAIn's is twelve commandments and a few customer sentences.