Suggestions
Nathan Gould
Founder @ Endeavor Labs
Nathan Gould is a professional based in Brooklyn, New York, currently serving as a Practitioner at Crosslake Technologies since October 2022. He also founded Endeavor Labs in May 2022, where he continues to lead initiatives. Prior to these roles, he was the Founder and CEO of GridReason from June 2013 to July 2014 and worked as a consultant for Advanced Microgrid Solutions for a brief period in 2014. His early career includes internships at California ISO, focusing on markets development and grid operations in 2008 and 2009.
In addition to his professional experience, Nathan is proficient in Spanish, which enhances his communication skills in diverse environments.1
Highlights
The AI headlines these days focus on frontier models, cutting edge infrastructure, and all the shiny new tech making previously impossible things possible.
But your average business, possible isn't good enough. It has to also be easy.
That's why I get jazzed when I see tech companies putting their dollars toward small improvements that make life easier for customers everywhere.
One recent example: MLFlow is now available as a fully-managed service as part of AWS SageMaker.
https://lnkd.in/gFuXe-MT
As a result, thousands of ML practitioners will be able to use a best-of-breed tool without waiting on DevOps support to self-host, or advocating for additional budget to spend outside of AWS.
Thanks Amazon Web Services (AWS)! ❤️
(This is not an ad. I just stumbled on this announcement and found myself moved to post about it.)
A common sentiment I’ve heard from developers building with LLM’s:
“It’s not worth spending much time on prompting strategy, since a better model is going to come out next month anyway.”
There are plenty of valid reasons to hold off on improving the performance of an LLM pipeline, but this is not one of them IMHO.
These days, gains in performance from methodology often significantly outweigh the gains you’ll likely get from a more powerful model.
The “More Agents is All You Need” paper offers a neat illustration of this point by showing that a simple sample-and-vote strategy can reduce the rate of bad outputs by ~50%. (See chart below.)
Of course, sample-and-vote is a very "brute force" sort of approach, and not actually one I'd actually recommend using. But I've come to sort of think of this ~50% gain as a sort of lower bound on the improvements that one can get from prompting methodology alone.
There are lots of better methods, ranging from simple few-shot learning, to multi-agent systems, automated prompt tuning, and RAG.
Using any of these requires learning something new, and experimenting a bit. But if you're looking for an edge, this is where to find it.
Don't sit on your hands and wait for GPT-5!
My 2c.
