Happy Sunday and welcome to InsideAI!  I'm Rob May, a Partner at PJC, investing in AI and Robotics.  If you are starting an early stage company in the AI space, please reach out.

Here are the most popular articles of the week from our daily newsletter:

European Parliament members adopted a new AI report yesterday that calls for a ban on lethal autonomous weapons, as well as “highly intrusive" AI-based social scoring applications. AI programs can't replace human contact or decision-making, according to the report, which provides guidelines for both military and non-military AI uses.

Microsoft was granted a patent for creating chatbots using people's social media profiles, prompting internet users to compare it to an episode of “Black Mirror." As UberGizmo notes, one application could be conversing with chatbots of loved ones after they've died.  In "Black Mirror's" first episode of its second season, a young woman's boyfriend is killed in a car accident and she uses AI technology to resurrect him as a chatbot and later, a robot.

Google opened an investigation into one of its AI researchers, Margaret Mitchell, under claims that she shared thousands of internal company files with outsiders. A source told Axios that Mitchell, a co-leader of Google's ethical AI team, used automated scripts to search through messages for instances of discriminatory treatment toward AI researcher Timnit Gebru.  Gebru, Mitchell's former colleague, who was also a co-lead of Google's ethical AI team, was ousted last month after Google asked her to retract an AI research paper she co-authored. Gebru later voiced frustration at managers' requests to retract the paper, as well as its treatment of women and minorities, particularly in hiring, which generated support from colleagues and people online.

One of the things I've been asked a lot lately is about MLOps as an area of investment, and how it differs from DevOps, and why the Devops tools won't just own the space.  On the latter question - will devops own the space - I don't have an opinion at the moment.  Long term, maybe they will.  But what I want to do today is briefly explain for those of you who don't work in the space how the workflows are different.

The use of data in most systems, up until recently, is either for analysis of some type, or retrieval for a user.  When you build a system and store data in the database, you pull it out as information.  The data doesn't usually affect how the underlying application works.  That is handled by the code.   Since changes to code might impact how the application functions in the future, those changes are tested with a series of what programmers call unit tests, and then functional tests, then deployed on a staging server (a server that simulates a real world production environment but isn't), and archived so you can roll back to the previous version if necessary.

Data, however, is usually backed up or archived for disaster recovery purposes or legal purposes, but you really don't need a copy of every change of the data to make the application work or possibly roll back if the application doesn't work as planned after some changes. 

In machine learning models, the data is used to train a model, and the model performs a task.  The code is used to define the structure of the neural network, and doesn't change as often as the data.  So the functionality of your application is more impacted by changes in data than changes in code.   That isn't how most software works today. 

As a result, MLOps is emerging as a set of tools that manage, analyze, and archive these models and their related performance characteristics so that you can understand how data changes might impact performance.  The workflows here are different.  For example, most web applications that don't use ML don't need a data engineering pipeline, but most ML applications do. 

The workflows are still changing, and there aren't set engineering standards and best practices to the same extent there are in traditional software just yet.  Until that happens it will be difficult to predict where MLOps will go.  My guess though, is that there will be 1 or 2 MLOps companies that get big enough fast enough that they become standalone or very large acquisitions, and the rest that show some traction will get swallowed up by companies like Snowflake, Github, Splunk, and the like. 

What will determine the final outcome?  One key thing to watch is how similar the workflow and related UX/UI is between MLOps and DevOps.  The other is how fast MLOps grows.  If it's slow and steady, expect DevOps tools to add it over time.  If it's fast, then MLOps startups will benefit from being focused just on that.

There is a lot more depth I could cover here, but, I hope this was a good high level explanation of what is going on in the market at the moment.

Thanks for reading.

@robmay