From 3917c48d60d981ffe6c5deeea5b620e5b0b26ae6 Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Thu, 2 Oct 2025 06:33:51 +0000 Subject: [PATCH 1/2] Add AI and LLMs 101 guide Co-authored-by: jamescashdesign --- docs.json | 3 ++- guides/ai-llms.mdx | 61 ++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 63 insertions(+), 1 deletion(-) create mode 100644 guides/ai-llms.mdx diff --git a/docs.json b/docs.json index 9bd3c8b..81b949e 100644 --- a/docs.json +++ b/docs.json @@ -18,7 +18,8 @@ "pages": [ "index", "quickstart", - "development" + "development", + "guides/ai-llms" ] }, { diff --git a/guides/ai-llms.mdx b/guides/ai-llms.mdx new file mode 100644 index 0000000..d936720 --- /dev/null +++ b/guides/ai-llms.mdx @@ -0,0 +1,61 @@ +--- +title: "AI and LLMs 101" +description: "A practical introduction to Artificial Intelligence and Large Language Models" +--- + +## What is AI? + +Artificial Intelligence (AI) refers to computer systems designed to perform tasks that typically require human intelligence—understanding language, recognizing patterns, making decisions, or generating content. + + +In this playbook, we focus on AI systems that process and generate natural language to power product experiences. + + +## What are LLMs? + +Large Language Models (LLMs) are AI models trained on vast text corpora to predict the next token in a sequence. This simple objective, scaled with data and compute, enables capabilities like drafting content, answering questions, summarizing text, and following instructions. + +- **Inputs:** prompts, system instructions, and optional context (documents, tools, or structured data) +- **Outputs:** tokens forming text, code, or structured formats like JSON + +## How LLMs work (at a glance) + +- **Tokenization:** Text is split into tokens (sub-word units). Models operate over tokens, not characters. +- **Transformers:** The architecture uses self-attention to weigh relevant context efficiently across long inputs. +- **Pretraining:** Models learn general language patterns by predicting next tokens on large datasets. +- **Finetuning & alignment:** Additional training (instruction tuning, RLHF/DPO) steers models toward helpful, honest, and safe behavior. + + +Despite impressive outputs, LLMs are probabilistic next-token predictors—not databases or oracles. Always validate critical claims. + + +## Strengths and limits + +- **Strengths:** rapid drafting, summarization, translation, pattern recognition in text, flexible interfaces +- **Limits:** hallucinations, outdated knowledge, sensitivity to prompt phrasing, difficulty with precise computation or long reasoning without scaffolding + + +Never rely solely on model outputs for high-risk domains without human review, guardrails, and clear escalation paths. + + +## Common terms + +- **Prompt:** The instructions and context provided to the model +- **Context window:** Maximum tokens of input plus output the model can handle at once +- **Temperature/top‑p:** Parameters that control randomness and diversity of outputs +- **System prompt:** High-level guidance that sets persona, goals, or constraints +- **Tools/functions:** External capabilities the model can call (search, code execution, databases) + +## Choosing a model + +Match the model to the job: + +- **General assistants:** Balanced instruction-following and safety +- **Reasoning tasks:** Models trained for extended or multi-step reasoning +- **Cost-sensitive flows:** Smaller or open models with prompt engineering and evaluation +- **Private data or latency needs:** On-prem or edge-deployed models with tight context windows + + +Evaluate models against your scenarios and rubrics. Track agreement between human reviews and automated judges before broad rollout. + + From 0c80a62526c3f6b23606f3c200cca713b2d3115f Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Thu, 2 Oct 2025 06:36:10 +0000 Subject: [PATCH 2/2] Refactor AI LLM guide for clarity and conciseness Co-authored-by: jamescashdesign --- guides/ai-llms.mdx | 42 ++++++++++++++++++------------------------ 1 file changed, 18 insertions(+), 24 deletions(-) diff --git a/guides/ai-llms.mdx b/guides/ai-llms.mdx index d936720..f7d55a1 100644 --- a/guides/ai-llms.mdx +++ b/guides/ai-llms.mdx @@ -5,57 +5,51 @@ description: "A practical introduction to Artificial Intelligence and Large Lang ## What is AI? -Artificial Intelligence (AI) refers to computer systems designed to perform tasks that typically require human intelligence—understanding language, recognizing patterns, making decisions, or generating content. +Artificial Intelligence (AI) refers to systems that perform language understanding, pattern recognition, decision-making, and content generation. -In this playbook, we focus on AI systems that process and generate natural language to power product experiences. +This playbook focuses on language AI that powers product experiences. ## What are LLMs? -Large Language Models (LLMs) are AI models trained on vast text corpora to predict the next token in a sequence. This simple objective, scaled with data and compute, enables capabilities like drafting content, answering questions, summarizing text, and following instructions. - -- **Inputs:** prompts, system instructions, and optional context (documents, tools, or structured data) -- **Outputs:** tokens forming text, code, or structured formats like JSON +Large Language Models (LLMs) predict the next token to generate text, code, and summaries from prompts and optional context. ## How LLMs work (at a glance) -- **Tokenization:** Text is split into tokens (sub-word units). Models operate over tokens, not characters. -- **Transformers:** The architecture uses self-attention to weigh relevant context efficiently across long inputs. -- **Pretraining:** Models learn general language patterns by predicting next tokens on large datasets. -- **Finetuning & alignment:** Additional training (instruction tuning, RLHF/DPO) steers models toward helpful, honest, and safe behavior. +- **Tokenization:** Text is split into subword tokens. +- **Transformers:** Self-attention weighs relevant context across long inputs. +- **Training:** Pretraining plus alignment (instruction tuning, RLHF/DPO) steers behavior. -Despite impressive outputs, LLMs are probabilistic next-token predictors—not databases or oracles. Always validate critical claims. +Treat outputs as plausible text, not facts—verify critical claims. ## Strengths and limits -- **Strengths:** rapid drafting, summarization, translation, pattern recognition in text, flexible interfaces -- **Limits:** hallucinations, outdated knowledge, sensitivity to prompt phrasing, difficulty with precise computation or long reasoning without scaffolding +- **Strengths:** drafting, summarization, Q&A, translation, flexible interfaces +- **Limits:** hallucinations, outdated knowledge, prompt sensitivity, weak math/logic without scaffolding -Never rely solely on model outputs for high-risk domains without human review, guardrails, and clear escalation paths. +Use human review and guardrails for high‑risk flows. ## Common terms -- **Prompt:** The instructions and context provided to the model -- **Context window:** Maximum tokens of input plus output the model can handle at once -- **Temperature/top‑p:** Parameters that control randomness and diversity of outputs -- **System prompt:** High-level guidance that sets persona, goals, or constraints -- **Tools/functions:** External capabilities the model can call (search, code execution, databases) +- **Prompt:** Instructions and context for the model +- **Context window:** Max tokens of input plus output +- **Temperature/top‑p:** Controls randomness/diversity +- **Tools/functions:** External calls (search, code, databases) ## Choosing a model Match the model to the job: -- **General assistants:** Balanced instruction-following and safety -- **Reasoning tasks:** Models trained for extended or multi-step reasoning -- **Cost-sensitive flows:** Smaller or open models with prompt engineering and evaluation -- **Private data or latency needs:** On-prem or edge-deployed models with tight context windows +- **General purpose:** Balanced instruction-following and safety +- **Reasoning-focused:** Trained for multi-step reasoning +- **Cost/latency constrained:** Smaller/open or on‑prem/edge deployments -Evaluate models against your scenarios and rubrics. Track agreement between human reviews and automated judges before broad rollout. +Evaluate on your scenarios and rubrics; track agreement with human reviews.