DeepSeek R1 reasoning open-weights

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-modelx-deepseek-r1-reasoning-builders

1. What makes DeepSeek R1 unusual compared to most commercial AI assistants?

   1. It is the fastest AI assistant available
   2. It can only answer questions about mathematics
   3. It is an open-weights model that achieves reasoning quality close to expensive frontier systems
   4. It was released before any other AI model

2. What does it mean that DeepSeek R1 has 'open weights'?

   1. The model is free to use on any website without registration
   2. Anyone can download the model files and run them on their own hardware
   3. The model cannot be used for commercial purposes
   4. The model's training data is publicly available

3. What is 'chain-of-thought' reasoning in AI models?

   1. A feature that lets models write their thoughts as social media posts
   2. The model shows its internal reasoning steps before giving a final answer
   3. A method for connecting multiple AI models together
   4. A technique for making models faster by skipping steps

4. How does DeepSeek R1's cost compare to OpenAI's o-series models?

   1. R1 is more expensive because it's open-source
   2. R1 is slightly cheaper but not dramatically different
   3. R1 and OpenAI o-series cost roughly the same
   4. R1 costs about 1/20th of OpenAI's pricing

5. What is 'distillation' in the context of AI models?

   1. Combining two models into one
   2. A method for removing unwanted behaviors from a model
   3. Creating a smaller, faster model by training it to mimic a larger model's outputs
   4. A technique for making models generate more text

6. Why does DeepSeek R1 often have slower response times than simpler AI models?

   1. It spends time generating chain-of-thought reasoning before answering
   2. The model is poorly optimized for speed
   3. The servers are always overloaded
   4. R1 uses older hardware

7. What is the main advantage of R1-Distill-Llama-70B over the full R1 model?

   1. It requires much less computational power to run while keeping most of the reasoning ability
   2. It thinks faster because it's smarter
   3. It can be used commercially without restrictions
   4. It produces more accurate answers

8. What hardware can run R1-Distill-Llama-70B?

   1. A smartphone
   2. A single high-end GPU like an H100
   3. A standard laptop without a GPU
   4. A network of 100 GPUs

9. In what way is DeepSeek R1's quality described relative to frontier models?

   1. It is significantly worse than frontier models on everything
   2. It has surpassed all frontier models
   3. It achieves near-frontier quality but only on selected reasoning tasks
   4. It matches frontier models on all tasks

10. When should someone still pay for high-effort GPT models instead of using R1?

    1. For translating between common languages
    2. For summarizing short documents
    3. For writing emails and social media posts
    4. For frontier competition math and novel scientific reasoning where maximum accuracy matters

11. What is the main tradeoff when choosing between DeepSeek R1 and a frontier model like OpenAI's o-series?

    1. R1 is open-source but much slower; frontier models are closed but faster
    2. R1 requires internet; frontier models can be run offline
    3. R1 offers lower cost and openness but slightly lower peak capability on the hardest problems
    4. There is no real tradeoff; they are equivalent

12. Why might a startup choose DeepSeek R1 over OpenAI's reasoning models?

    1. Because OpenAI doesn't support startups
    2. Because OpenAI's models are illegal to use
    3. Because R1 is made by a more trusted company
    4. Because R1 provides similar reasoning capability at a fraction of the cost

13. What does the lesson imply about the future of open-weights reasoning models?

    1. They will be replaced by API-only models
    2. They represent a new category that is changing the AI landscape
    3. They will always be significantly worse than closed models
    4. They are only useful for hobbyists

14. What is required to achieve the 'realistic self-host target' mentioned in the lesson?

    1. A single advanced GPU like an H100
    2. A supercomputer
    3. A connection to DeepSeek's servers
    4. A powerful server with multiple CPUs

15. What makes R1 different from a model like GPT-5.5 in terms of accessibility?

    1. R1 cannot be used by individuals
    2. GPT-5.5 is more affordable for developers
    3. Both models have identical accessibility
    4. R1 allows self-hosting while GPT-5.5 requires API access