Study Group on
ML Systems
in Production
Cerulean LabsPretraining
Learn from data.
+
Post-training
Refine behavior.
Post-training
A crowded frontier.
Frontier
Pretraining
OpenAI
Google DeepMind
Anthropic
Meta
DeepSeek
Alibaba / Qwen
Kimi
Grok
Frontier
Pretraining
OpenAIGoogle DeepMindAnthropicMetaDeepSeekAlibaba / QwenKimiGrokDeepSeek-V3 pretraining
US$5.328M
Estimated compute cost
We overthrow
pretraining.
More room to explore.
Architecture choices.
Learning efficiency.
Beyond CUDA.
More hardware.
More people doing science.
Open the whole process.
Code.
Experiments.
Results.
The study group
A place to become
a researcher.
Built around three things.
Hands-on work
Labs-centered
Automated scoring
People to think with
Deep-dives into papers
Office hours to chat with
Compute
16,000
H200 GPU-hours
Secured for the project.
People to think with.
3
Papers
First Author, Paper in AAAI 2024 Student Abstracts
First Author, Paper in IJCAI 2023 Symposium on LLMs
Authors, Paper in TAICHI 2026
TPC Member, FLLM 2023 - 2025
Dept. of CS Course Lecturer
Honorable Mention, HiPAC 2025
Participant, ISC 2026
Honorable Mention, HiPAC 2026
Creator of Superfile
Founding Product Manager, Clustron
Senior Backend Engineer, Clustron
3
HPC Competitions
First Author, Paper in AAAI 2024 Student Abstracts
First Author, Paper in IJCAI 2023 Symposium on LLMs
Authors, Paper in TAICHI 2026
TPC Member, FLLM 2023 - 2025
Dept. of CS Course Lecturer
Honorable Mention, HiPAC 2025
Participant, ISC 2026
Honorable Mention, HiPAC 2026
Creator of Superfile
Founding Product Manager, Clustron
Senior Backend Engineer, Clustron
23k+
GitHub Stars
First Author, Paper in AAAI 2024 Student Abstracts
First Author, Paper in IJCAI 2023 Symposium on LLMs
Authors, Paper in TAICHI 2026
TPC Member, FLLM 2023 - 2025
Dept. of CS Course Lecturer
Honorable Mention, HiPAC 2025
Participant, ISC 2026
Honorable Mention, HiPAC 2026
Creator of Superfile
Founding Product Manager, Clustron
Senior Backend Engineer, Clustron
The learning arc
Foundations
GPT-2 architecture and training
Optimizers and experiment design
Research sprints
Embeddings, data, and architecture
Build a deep learning framework
Your research
One week to plan
Three weeks to investigate
6–9
credits of workload
Every Tuesday
20:00 – 22:30
In person
ML and software engineering experience expected.
Join us.
Applications close
September 25, 23:59
Results by September 28
Apply through the SDC registration form.