Findings
This report examines the critical factors that either facilitate or impede AI research breakthroughs within large technology organizations. Based on in-depth interviews with AI researchers and practitioners, we identify key enablers such as existing relationships, access to high-quality data, and discoverability of internal documentation. Significant barriers include organizational silos, access to computational resources, regulatory hurdles, internal politics, and the complexities of collaboration..
3 Enablers of AI Research Breakthroughs
Fostering Connections, Relationships Across Silos and Industries: Pre-existing relationships, both internal and external, significantly accelerate research. One interview highlighted the crucial role of business development teams in leveraging existing connections: "We had a team of business development who had pre-existing relationships with Stanford." Another researcher emphasized the importance of long-term collaborations, stating, "We've been collaborating with [Researcher] at [University] for many years. So this paper was a one-off. One of these areas was data exchange." This underscores the value of networking and sustained partnerships in gaining access to expertise, data, and resources.
Existing Good Data: Access to large, high-quality datasets is a fundamental enabler. As one researcher put it, "[Our organization’s leader] said we are sitting on 10 billion dollars worth of data." The interview data consistently showed the organization possessing large quantities of relevant data. For large companies like Google, this means the ability to reuse StreetView data and Waymo data for 3D modeling. Existing datasets, even those initially collected for different purposes, can be repurposed to fuel new AI research initiatives. The ability to leverage existing data infrastructure, such as pipelines for data scanning and modification, was also mentioned as a key advantage.
Discoverability of Internal Docs: A culture of internal documentation and knowledge sharing allows researchers to build upon previous work and avoid redundant efforts. One approach, as described by a researcher, involves circulating concise project proposals: "I created a provenance 1-pager and circulated it... You create 50 1-pager and hope one takes off." This suggests a system for proposing and disseminating new research ideas, although the effectiveness of such systems can vary.
5 Barriers to AI Research Breakthroughs
Significant obstacles were also reported, hindering the progress of AI research:
Organizational Alignment and Internal Factors (Org Silos): Lack of alignment between research teams and product teams, along with fragmentation within large projects, can impede progress. Researchers cited many challenges exist in coordinating research across different organizational units.
Access to Compute (Resources): Limited access to computational resources, particularly specialized hardware like GPUs, can be a major bottleneck. A recurring theme in the interviews was the challenge of securing sufficient compute. One researcher recalled their struggle financially and politically to obtain access to compute. This underscores the critical role of computational infrastructure and highlights how resource allocation can become intertwined with internal politics.
Regulations, Approvals: Regulatory hurdles, particularly concerning data privacy and legal compliance, can significantly slow down research, especially in sensitive domains like healthcare. One interview revealed the scale of the challenge, speaking to the size of the regulatory team hired. Another researcher lamented, "Too many legal blockers." A particularly telling example involved image generation research: "Text-to-Image people are really frustrated because they can't train on video data because legal and regulatory people need to go annotate it." This highlights the inherent tension between leveraging vast datasets and adhering to privacy regulations and ethical guidelines.
Internal Politics, Reorgs: Organizational restructuring and internal politics can disrupt research projects and divert resources, as several interviews revealed. One researcher described a project's demise: "One PM we worked with got laid off and that's why their project didn't launch." Another simply stated reorgs were their biggest pain point. A third shared the direction of their research changed due to reorgs, and lack of manager alignment was a significant driving force.
Collaboration Complexity: The inherent complexity of collaborative AI research, particularly in large-scale projects, presents challenges in model quality assessment, prioritizing tasks, and knowledge sharing. As one AI researcher described it: "Collaboration complexity, model quality, stack ranking options is tricky. What do you do when the model is not performing well but you have 20 ideas of what to pursue? It's tempting to get stuck in something."
Discussion and Conclusion
The interviews reveal a complex interplay of enablers and barriers within a large technology organization's AI research ecosystem. While the availability of vast datasets and internal expertise provides a strong foundation, organizational structures, resource constraints, and regulatory concerns can significantly impede progress.
Recommendations:
To foster a more conducive environment for AI research breakthroughs, large technology organizations should consider the following:
Streamline Internal Processes: Reduce bureaucratic hurdles related to approvals, data access, and resource allocation.
Promote Cross-Functional Collaboration: Implement mechanisms to facilitate communication and collaboration between research teams and product teams, as well as across different organizational units.
Invest in Computational Infrastructure: Ensure sufficient access to computational resources, particularly specialized hardware, for all research teams.
Develop Clear Data Governance Policies: Establish clear guidelines for data usage and privacy that balance innovation with ethical and legal considerations.
Foster a Culture of Knowledge Sharing: Encourage internal documentation, knowledge sharing, and mentorship to maximize the impact of research efforts.
Long-Term Strategic Vision: Support long-term, fundamental research, recognizing that not all projects will have immediate product applications. The importance of this was emphasized by one researcher, who highlighted the need to prioritize "'expected knowledge gained' not just 'improve products.'"
By addressing these challenges and building upon the existing enablers, large technology organizations can create a more fertile ground for groundbreaking AI research that benefits both the company and society as a whole. Further research could explore quantitative measures of these enablers and barriers, and compare the experiences across different organizations.