Interview with UbiOps

1) I think it is currently too difficult for many to get their AI systems from the lab to the field, actually running them in production. Mainly because there is a gap in tech skills and understanding between software teams and data science teams. In order to deliver value, this should improve.

2) Data quality and availability often remains a limiting factor and killer of many AI projects. A good data collection and processing infrastructure should become even more of a boardroom priority. Especially with a long-term vision behind it.

3) Sharing building blocks for AI systems and democratizing the development of algorithms. This will greatly improve the efficiency of teams within companies, but also for the broader ecosystem.

1) Trust. It is an essential aspect that people understand how their data is used and that AI is not magic.

2) The energy footprint of AI should not be underestimated and should get more awareness. AI models become larger and larger, and cloud services make it too easy to just add more compute to a problem instead of making the analytics more efficient. Energy labels for AI systems should become a thing.

3) Access to high quality data will become an even more distinguishing factor in who can develop good performing AI systems and who can't, resulting in unfair advantage of data-rich entities.

In the next 5 years, businesses will improve greatly in how to run AI at scale so more applications can have advanced analytics in them. Both in the cloud as on the edge. The ability to run AI in production is an important step towards broader adoption, as it becomes easier to run "AI behind anything". Practices and developments like MLOps will help us to create standards and reduce risks around running AI in production.

10 years from now AI will have a more decentralized character. I expect the current paradigm of migrating and processing all our data in central hubs (e.g. cloud data centers) will slowly transform in a more decentralized approach, both for AI training and inference practices. This distribution and democratization will result in more efficiency, but also better information security by design.

20 years from now is a long time. There will definitely be a form of analytics or AI in almost all electronic devices available. Assuming the developments in AI research have led us to models and algorithms that are capable of associative tasks and reasoning beyond the context of their training data, it would make them able to perform a different realm of tasks beyond narrow AI. But AI capabilities will remain limited if the data distribution they are trained on remains a sparse representation of the real world. Therefore, to get to a point beyond narrow AI we first need to improve data capturing technology and data quality if we really want to enable to train models on a digital data equivalent of what’s going on in the real world.

Better understanding climate dynamics and weather patterns would be an amazing area for AI to flourish. The complex nature of these systems lends itself well for a data driven approach, combined with the fact that there is a lot of data available. This could help us in analyzing and understanding local weather patterns and climate effects in vulnerable areas, but for instance also anticipating energy demand.

That the lack of transparency in where AI is used and what happens with people their personal data will result in distrust instead of adoption. We should strive for a system where consumers, and everyone actually, has more ownership on their data and insight in how it is used.

Related to this is the trend that it becomes harder and harder to distinguish real information from fake information. AI technologies fuel this as they are capable of synthesizing fake photos, speech and text. When they become more advanced, this will lead to a society where we don't know what information we can trust. That would be a scenario we all should be very afraid of and work on together to avoid.

We have some cool developments to share with the community and want to involve people in what's going on in the MLOps field.