ResearchOct 20
Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
Upvotes51Authors11
Loading…
Glossary term
The process of adjusting a model's parameters using data so it performs a task better.
Training is how a model learns. It repeatedly makes predictions on examples, measures how wrong they are with a loss function, and nudges its parameters to reduce that error.
Large models are trained for weeks or months on thousands of chips, which is why training is the most expensive part of building a frontier model.
The core of a training loop (PyTorch)
for x, y in loader:
pred = model(x) # 1. predict
loss = loss_fn(pred, y) # 2. measure error
optimizer.zero_grad()
loss.backward() # 3. compute gradients
optimizer.step() # 4. update the weightsTraining cost and data quality largely determine what a model can do, and they drive the industry's demand for chips and power.
ResearchOct 20
Upvotes51Authors11
ResearchOct 19
Upvotes84Authors10