I agree 100% with and fully support this letter by Joseph Suarez, who I work with on Neural MMO, an open-source RL environment for massively multiagent learning.

The Letter

Read the full letter on X/Twitter

Title: Rejecting Libraries Hamstrings AI Progress: An open letter to the NeurIPS2023 D&B Organizers

Request: Add “open-source libraries and tools that enable or accelerate ML research” to the call for papers.

Key Points

“Progress in AI research is dependent upon libraries, infrastructure, and middleware—not just algorithms, data, and hardware. The magnitude of this dependence varies by subfield. In reinforcement learning, it is the single largest factor preventing researchers from working on harder tasks. As evidenced by recent NeurIPS competitions (ProcGen, MineRL, NetHack, Neural MMO), many interesting RL problems can be investigated with a single GPU.”

“Creating a clear path to publication will enable fundamentally new research. Increasing compute and data requirements have impeded a large chunk of academic research. Software is an axis where academic freedom can still be more valuable than large budgets. We have failed to leverage this advantage: developers have to fight to publish years worth of work.”

Impact on Research

The lack of recognition for infrastructure work means:

  • Researchers duplicate effort instead of building on existing tools
  • Important tools remain unmaintained
  • Academic researchers are disadvantaged vs. industry (which values infrastructure)
  • Progress in compute-intensive fields (like RL) is unnecessarily slowed