Single-task optimization has a ceiling
Most production AI systems are built to do one thing well. That works until the task changes, the data distribution shifts, or a new requirement appears that the model was never trained to handle. Generalizing across tasks is a different design problem — and it requires different architectural thinking.
This workshop covers the patterns that researchers and engineers are currently exploring to push toward greater generality: modular architectures, memory-augmented networks, mixture-of-experts designs, and world model integration. We work through each pattern with code examples and documented failure cases.
Lab sessions
Three of the five workshop days are lab-heavy. You will implement a simplified memory-augmented agent, instrument a mixture-of-experts router to observe load distribution, and stress-test a world model on out-of-distribution inputs. All labs use Python with PT and open-source tooling — no proprietary APIs required.
Who this is for
Engineers with at least two years of hands-on deep learning experience. You should be comfortable writing custom training loops and reading architecture papers. The workshop moves fast — there is no time to cover fundamentals during sessions.
AGI Architecture Workshop: Designing Systems That Generalize
Single-task optimization has a ceiling
Most production AI systems are built to do one thing well. That works until the task changes, the data distribution shifts, or a new requirement appears that the model was never trained to handle. Generalizing across tasks is a different design problem — and it requires different architectural thinking.
This workshop covers the patterns that researchers and engineers are currently exploring to push toward greater generality: modular architectures, memory-augmented networks, mixture-of-experts designs, and world model integration. We work through each pattern with code examples and documented failure cases.
Lab sessions
Three of the five workshop days are lab-heavy. You will implement a simplified memory-augmented agent, instrument a mixture-of-experts router to observe load distribution, and stress-test a world model on out-of-distribution inputs. All labs use Python with PT and open-source tooling — no proprietary APIs required.
Who this is for
Engineers with at least two years of hands-on deep learning experience. You should be comfortable writing custom training loops and reading architecture papers. The workshop moves fast — there is no time to cover fundamentals during sessions.
Program structure
What gets covered and in what order — no filler, no repetition.
Workshop Days
- Day 1 — Modular architectures: routing, composition, and failure isolation
- Day 2 — Memory systems: differentiable memory, episodic buffers, and retrieval-augmented generation beyond RAG
- Day 3 — Lab: building and testing a memory-augmented agent on a multi-step task
- Day 4 — World models: learning environment dynamics for planning under uncertainty
- Day 5 — Lab: out-of-distribution stress testing and architectural post-mortem
Ready to start?
Seats fill up quickly — once the cohort closes, the next opening is months away.