AI Summer

Summary repository for AI Summer 2026. Introduction to generative AI, with practical applications for inferencing and training May 5-30, 2026

Presented by Vanderbilt Data Science Institute data scientists:

• Dr. Jesse Spencer-Smith, Chief Data Scientist
• Umang Chaudhry, Senior Data Scientist
• Myranda Shirk, Senior Data Scientist

Overview

The objective of these workshops is to develop foundational skills in understanding, inferencing and training generative AI models and other transformer models.

Getting Ready for AI Summer

Getting Accounts

You’ll want to use the most advanced AI chat model that you can get access to. To get a headstart, please create an account with https://claude.ai/ . We will be spending a signficant time using Claude, Claude Cowork and Claude Code. Please sign up for a paid Claude account to gain access to these tools.

Decide on a Project (recommended)

Think about any data you might want to bring to the workshop. Also begin thinking about any projects you might want to accomplish during our month. We’ll have office hours for you to work with us to get your first project off the ground!

Workshop Schedule and Recordings

Session will run live from 11am-1pm, with an office hour from 1pm to 2pm (all times Central).

All session recordings are posted to the AI Summer 2026 YouTube playlist.

Week 1

Day 1 — Tuesday, May 5, 2026

Topics covered:

• Working with secure data: Vanderbilt’s data classification levels (1–4) and the AI tool that fits each one
• Opting out of training, ChatGPT EDU, Vanderbilt’s Amplify, and where Claude / Cowork / Code fall in the framework
• Working locally: open-source models, quantization, distillation, and PEFT
• Demo: LM Studio for fully-local model use
• Tour of Claude Chat vs. Claude Cowork vs. Claude Code (desktop and CLI)
• What separates a chat from an agent: planning, task files, sub-agents, evaluation
• How LLMs actually work: tokens, embeddings, attention, MLP, temperature
• Context as the central problem: tools, MCPs, RAG, and context rot
• Skills and the agent harness (“Agent 2.0”); progressive disclosure
• Live build: co-creating a “sensitive field detector” skill in Claude Cowork

Resources:

• Lecture notes
• Recording on YouTube

Day 2 — Thursday, May 7, 2026

Topics covered:

• Recap: chat vs. agent (replay), and what “Agent 2.0” means
• Walkthroughs of two real skills: the sensitive-field detector and the lecture-recording-to-lecture-notes skill that produced these notes
• Three surfaces for creating skills — Chat, Cowork, Claude Code — and how to match the surface to the skill
• Anatomy of a SKILL.md: YAML frontmatter, body, $ARGUMENTS, and the lesser-known flags (disable-model-invocation, user-invocable, allowed-tools)
• Progressive disclosure: how metadata / body / bundled-files load in three levels
• Where skills live: Enterprise vs. Personal/Global (~/.claude/skills/) vs. Project (/.claude/skills/)
• Writing skill descriptions that actually trigger
• Case study: the three-step composable accessible-housing research pipeline (programs → housing → audit)
• Four design patterns from the pipeline: canonical taxonomies, do/not-do tables, confidence ratings, cross-skill references
• Skill-building framework and when to reach for the Skill Creator Skill
• Model Context Protocol (MCP): what it is, how to set up a server, and how skills layer on top of MCPs
• Claude Code modes (plan / accept-edits / ask), effort levels, and routines

Resources:

• Lecture notes
• Recording on YouTube

Week 2

Day 1 — Tuesday, May 12, 2026

Topics covered:

• Recap of the agent framework with three zoom levels: the LLM, the harness, and the working space
• Where a skill actually lives on disk — global vs. project vs. session vs. Anthropic-bundled
• Anatomy of a skill folder: SKILL.md (required), scripts/, references/, evals/, LICENSE
• How to tell if Claude actually used your skill (and the slash-command trick to force it)
• One-verb rule for deciding when a task is one skill vs. many
• Two skill composition patterns: linear chain and fan-out / fan-in
• Spec-driven development: writing a precise, testable specification before you build
• Anatomy of a good spec: Goal, Inputs/Outputs, Constraints, Acceptance Criteria
• Three levels of tests: smoke, correctness, and edge cases
• Live demos: skill that didn’t trigger (vague prompt), the Customize → Skills panel, dissecting a real skill in breakout rooms
• Office hours: vetting third-party skills, prototyping security middleware, Level-3 Claude on AWS update

Resources:

• Lecture notes
• Recording on YouTube

Day 2 — Thursday, May 14, 2026

Topics covered:

• Warm-up Q&A: how prompt engineering changed once context windows got long; scoping agent security by data sensitivity
• The three surfaces revisited (chat / Cowork / Claude Code) and when to reach for each
• Live build: a ‘Claude developments tracker’ newsletter skill in Cowork with Skill Creator’s automatic eval loop
• Breakout reports: meeting-transcript-to-action-items extractor, PhD-level math tutor, Federal Reserve data downloader
• The plugin marketplace: Legal, Small Business, Bio Research — check before building from scratch
• AI Ethics guest lecture (Dr. Sarah Burriss): whiteboard activity, personal story tying PageRank to harm via Safiya Noble’s Algorithms of Oppression
• Sociotechnical ethics framework (Education / Technology / Policy) and entry points from philosophy, sociology, math, law
• Governance frameworks (NIST RMF, EU AI Act, AI Bill of Rights, UNICEF) and model cards as a practical accountability mechanism
• Office hours: AI environmental impact (Myranda’s research with Vanderbilt Law Lab), the next 6 months in agents, Obsidian-as-curated-context, AGI redefinition, whether skills are protectable IP

Resources:

• Lecture notes
• Recording on YouTube

Week 3

Day 1 — Tuesday, May 19, 2026

Topics covered:

• How you actually learn agents (use them, don’t read about them) and why models can’t reliably answer questions about themselves
• The three surfaces for building skills — Chat, Cowork, Claude Code — and how to choose; curating a subdirectory of source material as context (the 1,500-page Federal Acquisition Regulations example)
• Models and skills: why a well-authored skill lets Haiku 4.5 perform like Opus 4.7 without skills; author with Opus, run with Haiku; forcing /skill-creator; benchmarking a skill across models
• Claude Code deep dive: app (managed VM) vs. CLI (‘bare iron’); Shift-Tab plan / accept-edits / auto modes; watching the bash and git commands; converting PDFs to Markdown (OCR, context rot, data poisoning); running long, remote, and on a schedule (loop, channels, Claude RC)
• Live demo: iterating the lecture-notes skill from student feedback — standalone notes, banning the AI-narration voice, timestamps, take-home assignments; reprocessing all prior notes in minutes
• The Vanderbilt CV skill: a data-then-render architecture (parse to structured JSON, then generate .docx / .md / .pdf)
• Breakout reports: a monthly interaction gap-analysis skill (Excel + report) and a smart bus-shelter assessment rubric skill
• Efficiency: tokens are ‘thought’; push repeatable work into code inside the skill (MarkItDown for PDFs); goal loops (/goal) for clearly judgeable targets
• Agents 1.0 (designed, robust, repeatable) vs. Agents 2.0 (dynamic, skill-driven, flexible) and how they compose; putting a skill in front of existing code to absorb failure modes
• Data sensitivity: email, Level 3 data, and computer use — what’s safe to send where
• Wrap-up: the end of regular AI Summer instruction and the move to self-selected cohorts (course creation; general projects)

Resources:

• Lecture notes
• Recording on YouTube

Week 4,

Breakout Rooms

How to Breakout

Remember we are all learning and exploring

• Please share your video upon entering the room and unmute
• Share your screens–someone volunteer to share their screen upon entering, and everyone be ready to share your screen to show what you’ve found
• Make notes of what you’ve discussed in the Response Reports below
• Everyone be ready to report out (random)
• Make some friends
• Breakout Rooms Worksheets

Report Documents

Google Docs has a limit of 100 people viewing/editing a document at one time.

Special Breakout Room Groups

Please be sure your display name is set in Zoom. If you are in one of the following special groups, please pre-pend your name with one of the following qualifiers.

• Data Science for Social Good: DSSG
• If you are in a lab and would like your own breakout room: Labname (keep it short, please!)
• If you are faculty and would like to be in a breakout room with other faculty: Faculty

For example, I might be DSSG-Jesse Spencer-Smith

Workshop Video Recordings

Video recordings of these workshops are on YouTube: the AI Summer 2026 playlist, or browse all of our playlists on the VUDataScience channel.

Looking for the code resources for Summer 2025? See the archive folder.

Course Resources

• Visual overview of Generative AI from 3Blue1Brown: https://www.youtube.com/watch?v=wjZofJX0v4M
• Semester-long course on transformer models, DS 5690. Graduate students and advanced undergraduates can register by contacting Jesse Spencer-Smith. We welcome auditing by a select number of postdoctoral fellows, and drop-ins from faculty!

Other Resources

Compute Grants for Vanderbilt Faculty and Students

DGX A100 Compute Grant: https://forms.gle/2mGfEy9DB4JU2GpZ8

Python

Transformers

• Natural Language Processing with Transformers by Lewis Tunstall, Leandro von Werra and Thomas Wolf. If you are affiliated with Vanderbilt University, you can access this pre-print book (and any book by O’Reilly) free by logging into O’Reilly Media using your Vanderbilt email address. Vanderbilt licenses all content from O’Reilly. The book covers Transformers for purposes beyond text.

Getting the Most out of this Course

To get the most out of this workshop:

• Open Colab (workbook) notebooks and actively write code along with the instructor
• Actively participate in discussions
• Actively participate in breakout rooms
• Work on homework assignments before coming to class
• Relax your mind and ask questions