Blunders and Big Wins: A Guide to Building a Successful AI Fellowship

Original Source Here

Blunders and Big Wins: A Guide to Building a Successful AI Fellowship

How to attract, develop and retain scarce Machine Learning talent

Finding the successful formula. Image by author. AI generated using Dall-E

The world of AI is moving at an incredible pace. There are tons of AI startups emerging each year and the amount of research published is increasing year-on-year. Although this is great news for those building AI companies, there are also some problems that go with it.

First of all, AI talent is scarce. You will need to be an attractive employer, which can either be solved through compensation or secondary benefits. Secondly, your AI talent can fall behind quite quickly. What was state-of-the-art two years ago can already be outdated now. Keeping up with the latest trends is crucial if you want to stay competitive and interesting as a business.

Investing in AI R&D within your company can solve both problems. From our experience at Slimmer AI, people working as a Machine Learning (ML) engineer, ML researcher, or data scientist, usually have an academic background and prefer to have a research component in their work. They are motivated to learn and should be given the opportunity to do so. When talent development is taken seriously, it will boost job satisfaction and your in-house knowledge. A win-win!

So, what should this AI R&D program look like?

Within Slimmer AI, we’ve attempted to tackle this need for more AI R&D many times and learned what did and didn’t work (the hard way). After several iterations we now have a successful formula, which we call our AI Fellowship, and our engineers love it! In this blog post, I will share the lessons we learned, introduce you to our AI Fellowship and how to set it up in your own organization.

If you’re interested in knowing more about our AI Fellowship program, please have a look at this post, which shares more details on the exact implementation.

Our Failed Attempts (So You Can Avoid Them)

As an applied AI B2B venture studio we are continuously scouting for new opportunities. Knowing what is and what isn’t possible with AI enables us to focus on opportunities in exciting new areas and stay away from others.

However, just following the high-level trends is not enough. By knowing what is currently state-of-the-art and knowing how to actually implement it, what the risks are, etc., we’re gathering the necessary experience to deliver the best quality AI products.

The following sections outline our attempts at tackling this need over the past years and show what didn’t work and why.

Image by author. AI generated using Dall-E

Attempt #1: Divided We Fall

ML engineers are most likely working on at most one or two projects at the same time. This focus can be good short-term but limits their field of view in the long-term. We had several teams that worked on diverse projects and were using different data and AI techniques. To reduce the isolation, we introduced weekly AI stand-ups where we discussed the progress and issues we were facing. That sounded like a wonderful way of sharing knowledge and troubleshooting together.

However, after several months, we noticed that there was little to no interaction. Whenever someone was hitting a roadblock and asked for help, no one was able to. As we were all focused on our own projects, we simply didn’t have the knowledge or experience to help in a slightly different area. It was clear we needed something else.

Attempt #2: Out of Focus

To tackle this problem, we thought it would be helpful to share more in-depth knowledge on the AI that we were applying in our projects. This would enable everyone to better understand the issues others were facing when needed. ML engineers prepared half hour long technical deep dive presentations with some room for discussion at the end. We shared exciting new papers and experiences we gained when applying new models or techniques.

Although these sessions were great and helpful, it was difficult to keep them aligned with current work. This was because these technical deep dives weren’t focused on our core business. Additionally, people found it hard to find time to prepare these sessions as there were project deadlines to be met.

Even though these sessions didn’t meet our intended goal, we repurposed them to let everyone within our company share knowledge and educate others on diverse topics. The frequency and depth of machine learning content has gone down, but the communication within our company has increased a lot. It’s a nice way of keeping in touch with all the different departments, especially when a lot of us are working remotely.

Attempt #3: Who’s Driving this Car?

At this point, we were still lacking a way to increase knowledge and, maybe more importantly, hands-on experience. The next initiative was meant to tackle both.

The idea was centered around selecting a few topics of interest and forming groups to do some research and development. The development component, however, wasn’t that strong and we lacked someone driving and overseeing the program. Again, within a few months time people were prioritizing other work and the program stopped.

Attempt #4: Forgetting the Real World

We attempted to patch up the previous initiative by assigning clear leaders per topic and telling everyone that they had dedicated time to work on the R&D work. This helped noticeably. With clear leaders the work stayed on track and people were pulled in to be more engaged.

However, not every appointed leader had equal affinity with the role. Teams came together to do R&D work on fixed days which increased everyone’s motivation. However, it didn’t take long for everyday work and the pressure of deadlines to creep up on us and slowly disintegrate the teams.

That the R&D topics mostly focused around the engineers’ interests was fun at first, but without a clear link to the company’s near term strategic priorities, the work soon started to feel rather insignificant.

Based on these past attempts, we set out to create something new and lasting. Something that people get excited about, feel they can spend time on, and something that would really contribute to increasing our knowledge and hands-on experience.

Enter: The AI Fellowship

The Slimmer AI AI Fellowship program is part of the fabric of our culture; a program which our ML engineers consistently rate as very or extremely valuable. A program that not only the engineers are excited about, but everyone in the company. Our engineers are motivated and become more knowledgeable, which means we deliver higher quality products. Better products make happy customers.

The AI Fellowship shares many similarities with Chapters from the Spotify model, which represents a group of people with a similar competency area. Apart from sharing knowledge and ideas within this area, we added a dedicated R&D program to the mix.

This program centers around four main objectives:

1 Accelerate Innovation We work on topics that are aligned with Slimmer AI’s near-future needs, such that we stay ahead of the game.

2 Attract, Develop, and Retain Talent We work on topics that are aligned with both gaps and interests in our employees’ expertise.

3 Thought Leadership We strengthen our thought leadership position by sharing and publishing our findings.

4 Strengthen Venture Ties As a venture studio, we care not only about the Slimmer AI team but the ML capabilities of our portfolio companies as well. We include venture employees in our Fellowship and help them set up their own programs.

The R&D program

So, what is the success behind our Fellowship? Based on the identified patterns of our failures there are a few key aspects that make a great AI Fellowship:

  • Having a clear structure
  • Mandatory participation and management commitment
  • Dedicated time per week and good planning
  • Topics that are aligned with business goals and employees’ interests
  • Having someone drive innovation

For details on the exact implementation of our R&D program, please refer to this post. For now, I’ll just briefly highlight what you need to know and share some helpful tips & tricks.

Clear Structure

At Slimmer AI, we have four R&D cycles per year. One R&D cycle constitutes 3 months and consists of a 7-week research and 2-week development phase. The remaining time is used as a buffer or to publish interesting findings in, for example, a blog post, research paper, or open-source software. Check out our Medium page and GitHub if you’re interested in our past work.

Mandatory Participation & Dedicated Time

One of the most crucial aspects of the program is dedicated time. This is only possible through support at the most senior levels of your organization. The continued success of the AI Fellowship is part of every ML engineer’s role card and they are evaluated on their participation in advancing our R&D during reviews. This commitment from the leadership, combined with half a day per week of dedicated time during the research phase and 6 full days during the development phase, ensure our engineers feel confident they can spend their time on this important investment in our future.

Topics that Matter

During an R&D cycle, two or three AI topics are tackled in parallel. These topics are selected based on our short-term business needs, as well as the gaps and interests in our employees’ expertise.

Driving Innovation

Finally, we made it someone’s job to drive the AI Fellowship and its R&D program. Ideally, everyone is self-organized, proactive, and plans their own development path. But unfortunately, this is not reality. Especially in a time where many people work remotely, even getting a discussion going in a Zoom call can be quite challenging.

Therefore, having someone, or a group of people, drive the program is crucial. Responsibilities include: deciding which topics to tackle in cooperation with product managers, leading weekly stand ups, facilitating fruitful discussions, and making sure that the goals we set are being met. Don’t leave this to chance.

Tips & Tricks

After launching it in early 2021, the AI Fellowship has seen several iterations of improvement given the feedback by our ML engineers. I will share what we learned and what we changed over time, so you won’t have to reinvent the wheel.

Here are our tips and preferences regarding planning, topics, and general setup.

Image by author. AI generated using Dall-E


  • Sharing updates with the group once a week is preferred.
  • A one week development phase isn’t long enough. Two weeks is the minimum.
  • The development phase can feel short, therefore preparing for this phase during the last weeks of the research phase is crucial (e.g., prepare a dataset and the necessary code for loading it).
  • Don’t force the 7 and 2 week R&D timeline, but provide people the freedom to alternate between phases. Sometimes, it’s more convenient to have a development week half way through the cycle.
  • Let people plan their own R&D time. Some prefer to work on it half a day per week, others one day every two weeks. Both options are absolutely fine.
  • Consider organizing focus days where people come together to work on R&D work. This is especially useful for those that struggle with context switching or generally are distracted easily.


  • The topics should be accompanied with clear use cases or example research questions. Alternatively, time can be allocated to brainstorm on a research question later in the cycle to provide some focus.
  • Giving the engineers the freedom to choose what topic to work on feels liberating. Additionally, we provide people the opportunity to submit their own proposal.
  • Ideally, each topic has about the same number of people assigned to it. This is to keep the discussions balanced and not create isolation.
  • Add the possibility of continuing a topic during a follow-up R&D cycle. This allows for more ambitious goals.

General setup

  • Working in groups is more fun and motivating than working alone. Don’t make the groups too big to keep a strong individual sense of ownership. Two or three people per group is enough.
  • Having everyone write down their research allows others to read up on it later. Our engineers prefer to have this in a chronological order, rather than organized per subject. This is so they can easily catch up if they missed a stand-up.
  • For big collaborative R&D projects having someone act as a project lead is crucial. This also provides a great opportunity for someone to learn project management.
  • Mixing up the R&D cycles now and then is great for morale. For example, check out this blog post on an internal AI competition we hosted.
  • Interns are not actively involved in the R&D program, but are invited to our weekly AI stand-ups where they can learn and be part of the discussions.
  • This AI Fellowship format works perfectly for small to mid-sized teams of around 5–20 people. If your team is larger than that, consider creating a Fellowship per specific focus area — such as NLP, computer vision, etc.

In conclusion

AI is a fast moving field and as an applied AI company you need to stay up to date with the latest developments. Not only to stay ahead of the game, but also to attract and retain scarce talent.

In this blog post I introduced the AI Fellowship: an R&D program for a group of ML engineers. It was created based on the learnings of several failed attempts and can be implemented in any company that develops and applies AI. Our engineers love it and it has become one of our unique selling points.

The key characteristics of success for the Fellowship are: clear structure, mandatory participation, dedicated time, relevant topics, and having someone drive innovation that is tied to strategic company goals.

If you are planning on implementing an R&D program in your organization I’m interested to hear from you and share ideas. On the other hand, if you already have experience on setting up such a program in your company, please share your thoughts and wisdom in the comments section.

And if you want to know more about our work at Slimmer AI, feel free to reach out or head to our website to learn more.


Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot

%d bloggers like this: