Sakana AI Evolutionary Model Merging Quiz

Answers and explanations appear as you take the quiz.

1. What does Sakana AI's evolutionary model merging actually do?
   • Uses evolutionary search to combine existing pretrained models into new ones
   • Trains every new model from scratch on raw text
   • Deletes the weights of bad models randomly
   • Only fine-tunes a single model with more data
2. Who is a notable co-founder of Sakana AI?
   • The inventor of the GPU
   • A co-author of the original Transformer paper, Llion Jones
   • A former president of a country
   • The creator of Linux
3. What does the name 'Sakana' mean and reflect?
   • 'Mountain,' reflecting size
   • 'Fire,' reflecting speed
   • 'Fish' in Japanese, reflecting collective swarm-inspired AI
   • 'Cloud,' reflecting data centers
4. Sakana evolves model merges across which two spaces?
   • Audio and video only
   • Hardware and software licenses
   • Marketing and sales
   • Parameter space and data-flow (layer arrangement) space
5. What can reusing pretrained weights avoid in an evolutionary model-merging project?
   • Training a new base model from scratch
   • All compute for search and evaluation
   • The need to check model compatibility
   • Testing on a benchmark task
6. Besides model merging, what ambitious system did Sakana introduce?
   • A social media app
   • The 'AI Scientist' aimed at automating research
   • A video game console
   • A new programming language
7. How does the evolutionary search compare candidate model-merging recipes?
   • It chooses the recipe with the longest name
   • It selects candidates without running any evaluation
   • It evaluates them on task-specific benchmarks and uses scores to guide later candidates
   • It always keeps the first combination it tries
8. What is an important limitation when interpreting a merged model’s benchmark score?
   • It proves the model has no inherited limitations
   • It guarantees equal performance on every language
   • It means no evaluation data was used
   • The score reflects the evaluated tasks and setup, not guaranteed performance everywhere