AI & automation, page 3

Where AI sits in a screening mammogram read, and where people decide

1. Mammogram taken
2. AI scores the exam
3. Triage
4. Radiologist reads
5. Recall decision
6. Follow-up

From training data to a reviewed answer: how an LLM is made and used

1. Pretraining
2. Instruction tuning
3. Preference tuning
4. Evaluation
5. Answer with retrieval
6. Human review

From data to a discovery that holds up: where AI helps and what checks it

1. Data
2. Model
3. Search
4. Filter
5. Lab
6. Review

Where fairness is decided, from framing the problem to monitoring

1. Frame
2. Data
3. Train
4. Evaluate
5. Decide
6. Monitor

Training compute of three landmark models, and the EU AI Act's risk line

• 3.14 ×10²³ FLOP
• 5.76 ×10²³ FLOP
• 100 ×10²³ FLOP
• 380 ×10²³ FLOP