An analysis of DeepSeek's R1-Zero and R1
arcprize.org
An analysis of DeepSeek's R1-Zero and R1
1–10 of 280 posts
Re: An analysis of DeepSeek's R1-Zero and R1
#2Well o3 scored 75% on AGI-1, R1 and o1 only 25%.... watch this space though....
Re: An analysis of DeepSeek's R1-Zero and R1
#3The claim is that this removes the human bottleneck (aka SFT or supervised fine tuning) on domains with a verifiable reward. Critically, this verifiable reward is extremely hard to pin down in nearly all domains besides mathematics and computer science.
Re: An analysis of DeepSeek's R1-Zero and R1
#4>Ultimately, R1-Zero demonstrates the prototype of a potential scaling regime with zero human bottlenecks – even in the training data acquisition itself.
I would like this to be true, but doesn't the way they're doing RL also require tons of human data?
Re: An analysis of DeepSeek's R1-Zero and R1
#5The claim is that this removes the human bottleneck (aka SFT or supervised fine tuning) on domains with a verifiable reward. Critically, this verifiable reward is extremely hard to pin down in nearly all domains besides mathematics and computer science.
By verifiable do they mean it in the complexity theory P/NP sense of the word?
Re: An analysis of DeepSeek's R1-Zero and R1
#6I predict that the future of LLM's when it comes to coding and software creation is in "custom individually tailored apps".
Imagine telling an AI agent what app you want, the requirements and all that and it just builds everything needed from backend to frontend, asks for your input on how things should work, clarifying questions etc.
It tests the software by compiling and running it reading errors and failed tests and fixing the code.
Then, it deploys the software in production for you. It compiles your app to an APK file and publishes it on the Google play store for example.
Sure an LLM now may still not be able to get everything perfect as far as it's outputs go. But surely there's already systems and workflows in place that will auto run your code, compile it, feed errors back to the LLM, some api to interact with cloud providers for hosting etc?
Re: An analysis of DeepSeek's R1-Zero and R1
#7Well o3 scored 75% on AGI-1, R1 and o1 only 25%.... watch this space though....
with 57 million(!!) tokens
Re: An analysis of DeepSeek's R1-Zero and R1
#8The idea that a lot of compute is moving towards inference has a huge consequence for the current "AI investments". This is bad news for NVDA particularly. The inference focused solutions have better economics than paying NVDA those huge margins (e.g. Grog)
Re: An analysis of DeepSeek's R1-Zero and R1
#9I predict that the future of LLM's when it comes to coding and software creation is in "custom individually tailored apps". Imagine telling an AI agent what app you want, the requirements and all that and it just builds everything needed from backend to frontend, asks for your input on how things should work, clarifying questions etc. It tests the software by compiling and running it reading errors and failed tests a…
It doesn't need to write tests: it can just use the application and figure out if it works.
Re: An analysis of DeepSeek's R1-Zero and R1
#10The claim is that this removes the human bottleneck (aka SFT or supervised fine tuning) on domains with a verifiable reward. Critically, this verifiable reward is extremely hard to pin down in nearly all domains besides mathematics and computer science.
By verifiable do they mean it in the complexity theory P/NP sense of the word?
I think it just means that you can objectively score an answer as being correct or not. (e.g. if the generated program passes some tests; a discovered proof is valid, etc).