GOAT

Training compute

2.4×10²² FLOP

Parameters

3.5M

Published

Jul 27, 2021

GOAT is an AI model developed by DeepMind (United Kingdom), first published in July 2021. It works in the games domain, on tasks such as open ended play.

Training it took an estimated 2.4×10²² FLOP of compute (estimation method: hardware). The model has 3,472,816 parameters. It was trained on roughly 798.7T datapoints. Training ran on Google TPU v3. The compute alone is estimated at $85K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 228 citations. Epoch AI rates the confidence of this record as speculative.

Full record

Organization: DeepMind
Country of organization: United Kingdom
Domain: Games
Task: Open ended play
Training compute: 2.4×10²² FLOP
Compute estimation method: Hardware
Parameters: 3,472,816
Dataset size: 798.7T
Training hardware: Google TPU v3
Training cost (2023 USD): $85K
Model accessibility: Unreleased
Open weights: No
Citations: 228
Epoch confidence: Speculative