Suggestions
Gary Basin
Cofounder @ Pineapple Mortgage
Gary Basin is a fintech entrepreneur based in Miami, Florida, with a strong focus on transforming the mortgage industry. He holds a Bachelor of Business Administration in Finance from the University of Michigan's Stephen M. Ross School of Business. Throughout his career, Gary has accumulated extensive experience in trading and finance, having worked as a trader and software engineer at Balyasny Asset Management L.P., and as an Equities Proprietary Trader at Assent LLC, among other roles.
In the fintech sector, Gary co-founded Kapital Trading, a company designed to support traders. He has also served as a technical advisor for the Reserve project and is currently an advisor at Dapi and Bitwyre. Additionally, he is a co-founder of Propshop, a startup leveraging AI to innovate within the mortgage industry.
Highlights
clever paper, may be a hint at how o1 is implemented.
essentially use CoT as an information bottleneck: use RL to train an update function (fine-tune of base model) to generate CoT which can then be used on its own to predict the answer by the base model / frozen prediction function (without seeing the question).
This pushes a CoT's tokens to casually link to the answer, rather than allowing the model to rely on activations or tokens from the question to get the answer right even when CoT is irrelevant
what if: zeno's paradox for ai
unless we crack continuous learning (which seems legit insurmountable for neural nets), i can see a world where AI keeps getting more powerful (and crushes benchmarks) but asymptotes in usefulness because it's always trailing the leading edge of knowledge.
