In an attempt to gain a better understanding of the different types of intelligence models that developers are looking to build, we recently sent a brief survey to all of the registrants within our Private Beta Program. Closing in on 550 responses from the initial round of surveys we could not be more appreciative of the engagement and feedback we have received from our growing community. To those that participated, this information will be critical in helping shape the direction of our ongoing product investments. Thank you.
Given the current industry discussion trying to gauge the status and direction of AI adoption, we thought it would be useful to publish the survey results to help further inform the dialogue. While the questions asked were in some cases specific to the use of Bonsai, due to the horizontal applicability of the platform we think the the responses can also serve as a decent proxy for how people are looking to apply AI more broadly. However, as you read through the results below please keep in mind that the data was derived from a non-random sample as all survey respondents had already self-selected into the Bonsai Beta Program. With that in mind, let's take a look at the responses...
With six different use cases each flagging as relevant to 30% of the respondents, there is clearly a wide range of applications looking to benefit from increased intelligence. Notable in these results is what appears to be an emerging interest in use cases that fall outside of classic machine learning use cases of prediction, recommendation and natural language processing. As AI expands its reach, and new tools and technologies become available, we would expect to see continued expansion into these non-traditional use case including adaptive control and system/process optimization.
For personal projects and/or hobbyists, prediction, NLP, perception, and chatbots ranked as the top four use cases. The composition of these responses is not unexpected given these applications are built on data sets (e.g. text, images) that are more easily accessible to individuals. These applications have also received considerable attention across the AI community, as such there are a number of readily available APIs that can be leveraged to help expedite development.
Selected by 42% of respondents, TensorFlow has clearly developed an early mindshare leadership position (at least within our community) for AI toolkits. In contrast, across the IaaS players offering machine learning services we have yet to see the emergence of a clear leader with AWS, Azure and Google all selected by 10-12% range. While still very early days in the uptake of these services it will be notable to monitor the adoption trajectory of cloud-based services versus the likes of TensorFlow.
Lacking time-series data as this is our first survey, what we can’t decipher from this data is where on the adoption trajectory any of these respondents fall with any one of these tools. Are they happy with the tool they are using, and looking to build on top of them? Or are they displeased and looking for alternatives? Beyond TensorFlow, it is interesting to note the second most selected option was “None of the Above” at 37%. It is encouraging to see people that are new to AI, and likely lacking machine learning expertise, exploring the different tools and technologies that are accessible to them in the market.
However, with 74% of respondents selecting, “I have never used a simulator“, it is clear that we are very early days in the use of simulations as an AI training source. This lack of use, and likely overall familiarity with simulations leads us to believe that most users haven’t yet considered how a simulator could be combined with AI to solve their problem.
Dialing into only those respondents looking to address control or optimization use cases, and those that have never used a simulator approaches 100%. This is an extremely telling response as it relates to both the opportunity and challenge for developers building AI for optimization and control use cases. Control and optimization problems have very large and complex problems spaces that are extremely inefficient and expensive to model in the real world. In these situations simulations are the most viable and cost effective alternative, yet adoption to date as evidenced by this survey is extremely low. We as an industry clearly have our work cut out for us to better convey the benefits of simulation-based training, while also improving the developer experience by proactively integrating and perhaps even hosting select simulators within our platforms.
If interested, you can still complete the AI Adoption Survey, or you can sign up for the Bonsai Private Beta program and we will onboard you as fast as we can. If you do apply for the program, please note that we are on-boarding new program entrants as fast as we can. Your patience is greatly appreciated as we work through our backlog of beta registrants. We look forward to working with you to shape the future of AI.