Archive
Glossary
The terms that show up in Cinco, without oversimplifying.
- Active parameters
- In an MoE, the parameters actually used to process each token. They determine compute cost; total parameters determine memory.
- Agent
- A system where the model decides its own path: picks tools, observes results and repeats until done. Unlike a workflow, where the path is fixed code.
- Workflows are not agentsReAct
- Benchmark
- A fixed set of tasks with verifiable answers, used to compare models. It saturates when everyone gets nearly everything right; it is contaminated when answers leak into training.
- SWE-benchHELM
- Chain of thought
- Intermediate text the model produces before the answer, “reasoning” in writing. Improves results on multi-step tasks; not necessarily faithful to the internal computation.
- Attribution graphs show the model planning rhyme
- Contamination
- When examples from a benchmark (or something very similar) were in the model's training data, inflating the score without inflating the capability.
- Context window
- The maximum number of tokens the model receives at once. The advertised window and the window actually used are not the same thing.
- A large context window does not mean the context
- Distillation
- Training a smaller model to imitate a larger one's outputs. It is how much reasoning capability reached small models.
- DeepSeek-R1
- Fine-tuning
- Continuing the training of a finished model with task- or domain-specific data. Cheap relative to pre-training; easy to break without evaluation.
- Hallucination
- Fluent, confident output that does not match fact or source. Not an occasional bug; a consequence of how the model generates text, mitigated with sources, verification and refusal.
- Inference
- Running the model to generate answers (as opposed to training it). Where recurring cost and latency live.
- PagedAttention
- KV cache
- Memory where the model stores already-computed attention keys and values so it does not recompute them per token. Dominates memory use in long-context inference.
- PagedAttention
- LGPD
- Brazil's General Data Protection Law (13,709/2018). Article 20 — the right to review automated decisions — is the most direct point of contact with AI systems today.
- ANPD's AI sandbox has three companies testing ex
- MCP
- Model Context Protocol: an open standard for exposing tools and data to models, now governed by the Agentic AI Foundation (Linux Foundation).
- MCP became the tool standard for agents — and si
- Mixture of Experts (MoE)
- An architecture where only a fraction of the parameters (“experts”) is activated per token. That is why a 770B model can cost like a 49B one to run — but still needs memory for all 770B.
- Tencent open-sourced a 770B MoE under Apache 2.0DeepSeek-V3
- Monitorability
- How much undesired behavior can be detected by reading a model's reasoning. It drops when the model produces shorter or unfaithful chains.
- GPT-6 AstraAttribution graphs show the model planning rhyme
- PBIA
- Brazil's AI Plan (2024–2028), with R$ 23 billion announced. Execution through 2025 fell far short of the announcement; the RN supercomputer tender is the first large concrete disbursement.
- The government published the R$ 959 million tend
- PL 2338/2023
- Brazil's AI regulatory framework bill, risk-based. Approved by the Senate in December 2024; awaiting the rapporteur's opinion in the Chamber.
- PL 2338
- Quantization
- Storing weights with fewer bits (8, 4, even 2) to fit in less memory and run faster, with some quality loss. It is what allows running large models outside the datacenter.
- RAG
- Retrieval-augmented generation: fetching relevant documents and placing them in the context before generating. An alternative to trusting the model's memory — and to dumping everything in the window.
- A large context window does not mean the context
- Sparse autoencoder (SAE)
- A network trained to decompose a model's activations into many sparse and, with luck, readable features. The central tool of recent interpretability.
- Millions of interpretable features extracted fro
- System card
- A document a lab publishes with a model's capability and risk evaluations. Read the limitations sections before the results.
- GPT-6 Astra
- Throughput
- Tokens processed per second by an inference server. Together with latency, it defines the real cost of serving a model.
- Token
- The unit in which the model reads and writes text — word pieces. In Portuguese, a word usually becomes more tokens than in English, which makes the same sentence more expensive.
- Verifiable reward
- A training signal that can be checked automatically (the math checks out, the test passes). It is what makes reinforcement learning of reasoning possible without human examples.
- DeepSeek-R1
- Weights license
- Terms under which a model's weights may be used. Apache 2.0 and MIT are permissive; custom licenses (like Llama's) impose restrictions. “Open weights” is not the same as open source.
- Llama 3Tencent open-sourced a 770B MoE under Apache 2.0