No title
Recommendation systems
Self driving
Producing insights from data
...
AI tools?
Unsupervised learning
Detects patterns in data, without labels.
Reinforcement learning
Learning by trial and error.
Used in robotics, gaming, etc.
Generative AI
Generating new content, by looking at existing data (image, text, audio, etc.).
Supervised learning
Captcha / Classification.
GPT Assistant training pipeline
Pretraining
Uses large datasets.
Uses Transformer Architecture.
High GPU usage and months of training.
Supervised finetuning
Reward Modeling
Gets ratings for answers from humans.
Reinforcement learning
Uses the data from the previous step to improve the model.
ChatGPT used Reward Modeling and Reinforcement learning to improve their model.
AI journey
2010 to 2020 was the decade of supervised learning
AdTech Targeting
Search
Personalization
Needs large upfront investment in data and compute.
Requires large datasets.
2020 to 2030 is the decade of Generative AI
Generative AI Applications can be made in 6months
Supervised learning
Get labeled data (1 month)
Train AI model (3 month)
Deploy and iterate (3 months)
Prompt-based AI
Get labeled set, if you have a very exotic use case (NQL, 1month)
Specify prompt (1 hour)
Deploy and iterate (days)
Generative AI Applications can be made in 6months
Supervised learning
Get labeled data (1 month)
Train AI model (3 month)
Deploy and iterate (3 months)
Prompt-based AI
Get labeled set, if you have a very exotic use case (NQL, 1month)
Specify prompt (1 hour)
Deploy and iterate (days)