AI (Artificial Intelligence)

Concepts, Environments, and Collaboration

Brian Yandell (byandell.github.io)

2026-07-02

AI (Artificial Intelligence) Contents

Perspectives & Environmental Impact

How Might We Think About AI?

Tools Reshaping Us

“We shape our tools and thereafter they shape us.”
— Marshall McLuhan (via John Culkin)

  • Deeply engaging with AI is a process, not just pushing a button.
  • Treat it as a collaboration rather than just a passive tool.

Social Collaboration

Jaron Lanier’s There is no AI reframes AI as a medium for human cooperation, emphasizing data dignity and creator credit.

The Disrupted & Rethought Workforce

Workforce Transformation

  • Early success is seen in AI-agent-native startups.
  • Legacy companies experience shifts in workforce dynamics.
  • Focus is shifting from replacing people to rethinking how we collaborate.

Core References

Ethics & Environmental Impact

Environmental Footprint

  • Rapid growth of data centers strains local power and water resources.
  • Efficiency of algorithms is improving, but resource demand is rising.
  • Increasing focus on protecting natural resources and supporting indigenous lands.

Call for Guardrails

Scientists warn in Nature that uncritical adoption of AI in science requires urgent regulatory guardrails.

Ethics Resources:

Data Center Resources & Efficiency

Energy & Water Reports

  • UN Report: Details footprints of carbon, water, and land usage.
  • Texas HARC & UT News: Reviews data center boom and water usage in Texas.
  • LBNL: U.S. Data Center Energy Usage Report.
  • Data Center Map: Tracker for infrastructure.

Best Practices & Forecasts

  • Microsoft: Sustainable datacenter designs using zero water for cooling.
  • Google: Climate-conscious cooling commitments.
  • Salish Tribal Alliance: Zero Edge cloud data sovereignty.
  • Micro Data Centers: Shift toward Modular Edge Computing.

AI & the Scientific Community

Gleanings from the ESIIL Network Slack Channel:

AI Environments

What is an AI Environment?

Setting for AI Systems to Operate & Interact

  • Components: Hardware, software, data, and algorithms.
  • Types: Physical (robots) or virtual (simulations, cloud platforms).
  • Feedback: Allows real-time interaction and feedback with surroundings.
  • Machine Learning: Enables adaptation and improvement over time.

Applications

Autonomous vehicles, smart homes, virtual assistants, and connected IDEs.

Connected Development Environments

Model Context Protocol (MCP)

  • Open protocol introduced by Anthropic.
  • Allows LLMs to interact securely with local/cloud tools and codebase context.
  • Enables AI agents to read files, run commands, and edit code under user guidance.

Interactive Workflow

Rather than copying and pasting code, the AI agent directly reads and modifies local code blocks based on natural conversations.

Cloud Educational Environments

Enable students and teams to build and test code without complex local setup:

  • GitHub Codespaces: Cloud-hosted container integrated with GitHub and Copilot.
  • Google Colab: Interactive Python/R notebooks with Gemini integrations.
  • CyVerse Discovery Environment: Advanced biology and geospatial platform.
    • Integrates AI assistants (Roo / Cline) into its VS Code IDE workspace.
    • AI Verde: API tunnel providing access to high-performance LLMs.
  • NSF ACCESS: High-throughput computing via Jetstream2, HTCondor, and PATh.

Google Gemini & Antigravity

Antigravity 2.0 Suite

Launched May 2026 with four primary tools: - Antigravity 2.0: Multi-agentic code platform. - Antigravity CLI: Command line terminal tool. - Antigravity SDK: Building custom agent apps. - Antigravity IDE: Workspace-integrated assistant.

Transition Note

IDE version 2.0 does not carry over history automatically. See Recovering Antigravity Conversations for step-by-step help.

Gemini Resources:

Standalone AI IDEs & Tools

AI-native environments have quickly expanded beyond standard editor integrations:

  • Claude Code: Anthropic’s terminal-based coding agent.
  • OpenAI Codex: Deeply integrated code generation models.
  • Positron: Posit’s next-generation IDE built for R and Python data science.
  • Windsurf: Standalone AI editor optimizing code editing workflows.
  • Cursor: Advanced fork of VS Code featuring native MCP tool support.

Caution

Be careful when using terminal-based interactive interpreters (like radian) alongside background AI terminal execution.

Agentic AI & Multi-Agent Orchestration

Orchestrating Agent Teams

  • Modern developers orchestrate multiple agents to tackle complex tasks.
  • Each agent holds its own role, persona, and technical skills.
  • Commercial frameworks exist in Claude, ChatGPT, and Antigravity.
  • SOUL.md: Markdown-based configuration files defining agent personalities.

Team Science Platforms

Agentic AI References

Large Language Models (LLMs)

Transformer Networks

  • Structure: Massive transformer networks with nodes joined by weighted edges.
  • Weights: Nodes/edges are fixed; weights are adjusted based on training data.
  • Sequence Prediction: Learning patterns to predict the next token in a sequence.
  • Scale: Billions or trillions of nodes to generate rich responses to prompts.

LLM Explanations & Guides

What are AI Tokens?

How Models Process Language

  • Tokens: Units of data representing words, subwords, images, or audio snippets.
  • Encoding: Converting text to numerical embeddings that capture contextual meaning.
  • Contextualization: Processing embeddings through layers to understand relationships.
  • Decoding: Predicting and generating output tokens, then transforming them back to human text.

Tokenization Resources

Mixture of Experts (MoE)

Computational Efficiency

  • Conditional Routing: Activates only a subset of network layers for each query rather than the whole model.
  • Expert Sub-networks: Inputs are routed to specialized expert layers via a gating network.
  • Parameter Scaling: Allows massive scaling of parameters without a proportional increase in compute cost.

MoE Resources

Open Source LLM Specifications

Key Characteristics

  • Technical Specs: Developer lineage (Google Gemma, Meta Llama, Qwen, GLM), context windows (32k to 1M+), training tokens.
  • Operational & Licensing: Release dates, open vs. proprietary license types, quantization levels (FP16, INT8, 4-bit) defining memory/RAM needs.
  • Capabilities: Benchmark scores (MMLU, GSM8K), text-only vs. multimodal (images, audio, video), language support.

Model Sources (AI Verde)

What is an AI Harness?

The Harness Metaphor

  • Horse Metaphor: Ethan Mollick compares an LLM to a powerful horse. Without a harness (bridle, reins, saddle), it has no direction or boundaries.
  • Harness Engineering: Mitchell Hashimoto defines it as engineering solutions so an agent never repeats a mistake.
  • Normalizing Term: Mollick organizes AI into Models, Apps, and Harnesses.

Core Origins & Analogies

Harness Engineering References

Prompt & Context Engineering

Prompt Engineering Fundamentals

Documenting the Trail

  • Prompt engineering designs inputs to guide an AI’s output.
  • Always save your prompts to a file (e.g., prompts.md) to retain your workflow.
  • Ask the AI agent to update prompts.md and walkthrough.md as you code.

Core References

Prompts to Organize Workflows

Split complex workflows into human-sized, manageable chunks:

  1. Extract Functions: Move heavy operations to a helper R file (functions.R).
  2. Setup Notebook: Create a clean Quarto document (workflow.qmd) that sources functions.R.
  3. Handle Deliverables: Configure the workflow to output figures and data tables (PDF/CSV).
  4. Build Documentation: Present the workflow in clear source and rendered forms.

Note

Check out ESIIL’s Three Rivers working group AI Prompt Engineering docs for examples.

Context Engineering

Managing Attention Budgets

  • Every turn fills the agent’s context window.
  • Excessive background context or giant documents cause hallucination or crash.
  • Best Practice: Request the agent to summarize files, then save and pass the summary to subsequent chats.

Context Inflation

Avoid oversized onboarding files. Keep repository rules focused to preserve conversation token budget.

Context: AGENTS.md vs Agent Skills

AI coding assistants (Claude Code, Antigravity, Cursor) leverage two core standards to organize context:

Feature AGENTS.md Agent Skills (SKILL.md)
Purpose Static repository context (coding standards, project layout). Procedural knowledge and automation recipes (deployment, tests).
Scope Bound to a single repository. Reusable/portable across projects.
Structure Single Markdown file in root. Folder with SKILL.md + scripts.
Trigger Loaded automatically. Activated dynamically on keywords.

AGENTS.md (Repository Context)

Acts as the onboarding handbook for an AI agent entering a repository:

  • Directory Layout: Highlights where components, hooks, or styles live.
  • Design & Coding Standards: Strict rules (e.g., “Prefer existing Button components”).
  • Layered Rules: Nested AGENTS.md can layer constraints on specific sub-folders.

Real-world Examples:

Agent Skills & SKILL.md

Agent Skills represent a portable toolbox extending what an agent can perform:

  • Trigger Actions: YAML frontmatter defines triggers for specific command hooks.
  • TDD Workflows: Anti-rationalization rules prevent agents from skipping steps (e.g. bypassing test suites).
  • Automation Scripts: Scripts (Node, Python, Bash) execute complex deployment routines.
---
name: vercel-deploy
description: Deploy my web app to Vercel production.
---
# Deploy Skill
...