# Chris Re

> ~1979– · Computer Scientist, AI Researcher
>
> **Recorded contribution:** Data-centric AI; weak supervision (Snorkel); FlashAttention; Together AI; foundation model data

## How to use this dossier

Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory.

## 1. Historical orientation

Stanford computer scientist Christopher Ré works across database systems, machine learning, and AI infrastructure. His group helped develop data programming and Snorkel, where domain experts write noisy labeling functions that are statistically combined into training labels. He also co-authored Hogwild! and FlashAttention with their respective teams and co-founded several companies. The registry's previous automated source match to Chris Lattner was false and is corrected here.

## 2. The problem inherited

For many applied machine-learning tasks, labeling enough high-quality examples costs more than model construction, while expert knowledge exists as rules, heuristics, and weak external sources.

## 3. The central contribution

Ré and collaborators developed data programming and Snorkel as a framework for modeling and combining weak supervision, and contributed systems ideas for efficient learning.

## 4. Reconstruct the mechanism

1. Ask domain experts to write labeling functions that vote, abstain, or conflict on examples.
2. Apply those functions to unlabeled data and record the vote matrix.
3. Estimate function accuracy and dependence without treating any one source as ground truth.
4. Produce probabilistic labels and train a discriminative model that can generalize beyond the rules.

## 5. What changed downstream

- Weak supervision became a practical alternative to exhaustive hand labeling in specialized domains.
- The work strengthened the view that data and systems design are first-class parts of machine-learning research.

## 6. Attribution, limits, and uncertainty

- Snorkel, Hogwild!, and FlashAttention each have distinct co-author lineages; Ré is not sole creator of all registry-listed work.
- Correlated or systematically biased labeling functions can create confident error, and expert-written rules may encode institutional inequity.

## 7. Reconstruction lab

Write five labeling functions for twenty examples, including two correlated rules. Estimate a simple weighted vote, compare it with majority vote, and audit every confident disagreement. Hide the gold labels while authoring rules, then reveal a small validation set and measure coverage, conflict, and correlation separately. Add two near-duplicate labeling functions and explain why majority vote overstates their independent evidence. Train a discriminative model on probabilistic labels and audit where it exceeds the rules’ support. Snorkel’s first-principles move is to model supervision sources as noisy data rather than pretending heuristics are ground truth.

## 8. Evidence trail

- [Christopher Ré](https://www.cs.stanford.edu/people/chris-re) — Stanford Computer Science
- [Data Programming: Creating Large Training Sets, Quickly](https://arxiv.org/abs/1605.07723) — NeurIPS
- [FlashAttention](https://arxiv.org/abs/2205.14135) — NeurIPS

---

*Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*