Loading…
Loading…
Quote
“A field of study that gives computers the ability to learn without being explicitly programmed.” — coined while teaching a checkers program to improve through self-play.
Arthur Samuel coins “machine learning”
source→Paper
Krizhevsky, Sutskever & Hinton showed deep CNNs could shatter the ImageNet benchmark by a margin nobody expected. The GPU-trained network that lit the deep-learning era.
ImageNet Classification with Deep Convolutional Neural Networks
source→Paper
Vaswani et al. replaced recurrence and convolution with self-attention. The architecture under every major LLM that followed.
Attention Is All You Need
source→Paper
Kaplan, McCandlish, Brown, Amodei et al. quantified how loss scales smoothly with model size, dataset size, and compute. Turned “bigger is better” from intuition into a law.
Scaling Laws for Neural Language Models
source→1 source·4 chapters·24 sections·11 sandboxes
ML pulse · external
ML Sandbox is an interactive companion to the foundational works of machine-learning systems. Each source in the librarycontributes chapters of prose, figures, and runnable sandboxes — you can read, simulate, and tweak parameters from the page itself.
The dashboard above is not a marketing widget. Every number is sampled from this platform’s own backend: chapter counts come from the database, ops metrics come from the running Celery workers, the live simulation tile streams ticks over Server-Sent Events from the same Redis pub/sub channel every Pattern B sandbox uses.
Anonymous use is supported — no account required to read or run. Create one to save runs to a project and keep your history.
ImageNet Classification with Deep Convolutional Neural Networks
Ops
Live workers
1
Background tasks
2
0 reserved
Connected clients
54
Platform uptime
17h
Live simulation · backend
pausedThroughput
—
p99 latency
—
Utilization
—
Queue depth
—
A continuous batching simulator runs on the backend (Celery + Redis pub/sub), streaming snapshots to your browser over Server-Sent Events. Same machinery that powers every Pattern B sandbox in the library.
Concept of the day
The world changes; your training set doesn't. Concept drift is when the relationship between features and labels shifts over time — users behave differently, fraudsters adapt, fashion changes. Without monitoring, a model that was accurate last quarter is wrong today.
Featured chapter
Chapter 3 · Designing Machine Learning Systems · Chip Huyen
Where data comes from, what shape it sits in, and how it moves
Where data comes from, what shape it sits in, and how it moves. The chapter the rest of the platform leans on most heavily — every Pattern B sandbox in chapter 3 sits inside one of these sections.
Featured sandboxes
Hand-picked across the library. Run them right here — no sign-in required.
Explore latency vs throughput tradeoffs in request batching
Discrete-event simulation of a Poisson event stream under batch and stream processing; latency vs throughput trade-off