Sign In

David Mack

Successfully exited YC Founder, Deep Product + AI expertise

David Mack is an experienced founder, full stack engineer, ML developer, product designer, and manager with a passion for creating new products.

He holds a BS in Political Science and Economics from American University.

David has held key positions in various organizations, including Director of Public Affairs Campaigns at Meta, SVP Policy and Public Affairs at Eaze, and Senior Director of Public Affairs at Lyft.

His previous roles include Principal at 463 Communications, Account Manager at Bite Communications and Chlopak, Leonard, Schechter & Associates, and Director of Government Affairs at National Small Business Association.

Currently, David Mack is focused on building a new company while also enjoying outdoor activities like rock climbing, ski touring, and dog walking in the Sierra Nevada mountains.

Highlights

Jun 20 · linkedin

I’m happy to share that I’m starting a new position as Co-Founder, CTO at Presto Design!

Apr 14 · linkedin

LLMs seem to have firmly entered a performance enhancement phase as opposed to their previous era of rapid breakthroughs. This lines up with OpenAIs long standing prediction curves of exponential training flops required for larger parameters

This is troubling as LLMs have yet to achieve the critical level of being able to replace a general human worker, the point at which they can turbo charge service growth and affordability.

I believe that we might be able to eke out higher intelligence with higher compute spend on current models:

  • fine tune long running assistants to provide memory and better alignment with their users
  • use intermediate self talk for llm to “think out the options” before speaking
  • use long running self talk to “ruminate” and “ideate” on harder questions (akin to shower thoughts). This idea might stretch llms too far as they currently quickly get distracted from their goals and go dorm rabbit holes. To this end, creating reinforcement from human feedback on much much longer tasks could be very valuable
  • fine tuning LLM‘s on using common tools like a memory store just like humans use pen and paper to think about bigger questions could unlock further performance

Moreover, it does seem like we really need an order of magnitude larger parameters to solve a lot of LLM and adequacies. It is really encouraging the large amount of optimization research coming out, often being able to shrink down the storage size of parameters and disperse with parameter layers altogether. I really hope that this unlocks being able to train much larger parameter sizes.

David Mack
Add to my network

Location

Truckee, California, United States