Three Years in the Making: From AI Research to Finalizing the Concept | SpeakBUDDY Development Team Roundtable Vol. 1
（こちらは以前に公開された記事の英訳版です。日本語版はこちら）
In 2023, a wave of AI-generated services emerged, and within this trend, SpeakBUDDY launched the "BuddyChat" feature in September. This feature allows users to enjoy free conversations with AI characters (referred to as "AI buddies")
-How was this feature, utilizing the latest AI technology, created?
-What kind of discussions took place within the development team at that time?
This note reports on a roundtable discussion with the "BuddyChat" development team. It shares the team's thoughts on service development and their approach to feature creation, directly in their own words. The discussion was a mix of Japanese and English, but in this article, we will present the content as accurately as possible in English.
Technical research from 2020, starting with GPT-2
Interviewer: The development team began technical research on generative AI well before it became a popular topic of discussion. Could you share more about that time?
AI Engineer・Nick (hereafter N): We started tracking OpenAI's GPT technology from about 2020, beginning with the release of GPT-2 around 2020. At that time, other companies like Google and Facebook were also introducing conversational AI models, so we followed these developments as part of our regular tech research.
I think the project really kicked off with the release of GPT-3. Depending on the API, integrating it into the service became very easy, so we started creating prototypes with Michael and our CEO Tsuyoshi Tateishi.
Interviewer: What was the prototype like?
N: It was a prototype that allowed users to have conversations with AI Buddy (Character that serves as the conversation partner in “SpeakBuddy” App) about news topics. However, it didn't seem promising at the time as we couldn't see how it would lead to anything... Maybe it was because the quality of the conversation wasn't good. We had other higher priorities, so we set it aside temporarily (he laughs).
However, shortly after that, the model was updated, and we realized it had significantly improved—the conversation quality became much better. That's when we understood that it needed to be integrated into the service as quickly as possible.
The turning point: Slackbot Prototype
Interviewer: From the perspective of learning experience and effectiveness, how did you view things during this period, Sayaka?
Education Lead・Sayaka (hereafter S): I remember that around this time, Michael showed me Playground (*), and it was my first experience with GPT.
(*Playground: A web-based interface provided by OpenAI that allows for easy experimentation with GPT).
At first, I felt it might be a bit too tech-oriented. But as I continued to explore it with some guidance, I began to understand what GPT could do and started thinking that it could be very useful.
This was the first time we really leveraged Slackbot in our development. That turned out to be a true game-changer for us.
By using Slackbot, we truly began to understand what GPT could bring to the company. It led us to start thinking about how we could best leverage this technology to maximize its potential for our services and users.
Design Lead・Michael (hereafter M): We first created the news prototype in 2021, and after the GPT model was updated, we experimented with Playground and Slackbot in 2022. For about 3-4 months after that, we went back and forth, thinking about the concept and concrete features.
We debated what the focus should be-- allowing users to freely talk with Protty (Speakbuddy's official character), having Protty offer advice, letting users interact with AI buddy, or even incorporating it into a level check test... Each idea was completely different, so we kept going back and forth, considering what features to implement, what impact they would have, and what they would lead to.
Employee demo of the prototype: What was discovered?
Interviewer: The prototype that allowed us to talk with Protty was shared company-wide, and we tried it out on Slack. I remember it well.
M: That's right, we conducted a demo where participants could "choose one of four conversation themes and talk with Protty". We shared this with everyone in the company, and the results were extremely interesting.
M: First, we realized there was a fundamental issue: "There aren't any real topics to discuss with Protty."
S: Yes, there was a lack of information about Protty, so we didn't really know what to talk about.
Another point that we noticed: The participants had varying levels of English proficiency, which led us to the discovery that the conversations would be "more enjoyable if we were more fluent."
Enjoying a conversation isn't easy, and this made us certain that some support is necessary for free talk.
Interviewer: So, you discovered that just "talking freely" isn't enough to learn effectively?
S: We already understood that "just talking freely" isn't enough to learn effectively, so it felt more like a "rediscovery." The biggest discovery during this time was about having a "conversation partner that keeps the discussion going."
Even if you’re talking about travel with Protty, it’s hard to know what topics to bring up if you don’t understand whether a robot like him can actually go anywhere. On the other hand, with AI buddies like Zac or Ashley, who are frequently used in lessons, there are more natural topics of conversation.
To have a conversation that flows smoothly, the conversation partner needs to have a distinct personality and concrete background. This was the biggest realization we had at that time.
Also, now that I think about it, the prototype included a “GPT summary feature”, which was very impressive. The reactions from the team members who tried it were positive, and it was very well-received.
Narrowing down the final concept with SpeakBUDDY's unique perspective
Interviewer: How was the final feature to be developed decided?
S: Around winter of 2022, we started narrowing down the list of ideas. Since there were many different factors to consider, we assigned scores to each idea to evaluate them more effectively.
Interviewer: The final two proposals were "Protty Chat," where users would talk with Protty, and "Real Life Success," which eventually led to the current BuddyChat. What was the deciding factor between these two options?
S: The deciding factor was… “Fun” —whether it would be enjoyable and engaging for users. We focused on creating something that was not only fun but also provided a valuable learning experience.
iOS Engineer・Hiroki (hereafter H): “Protty Chat” stayed in the running due to its ease of "DEV" (development). However, it ultimately lacked impact in terms of "Learning" effectiveness, which made it less appealing as a final choice. Plus, the idea of greeting Protty every day would likely become repetitive and boring over time (he laughs)…
S: Low engagement was the deciding factor for dropping this idea. Learning is not a festival; the new feature needs to be something that users can continuously use and enjoy. From this perspective, I thought this proposal was too challenging.
In contrast, Proposal C, which led to the current “BuddyChat”, was built on the learning experience that already existed at that time. By utilizing GPT, I thought it could evolve and complete the learning cycle that we originally wanted to achieve.
Interviewer: In the list of considerations, there is also one more aspect: “AI”. What is this about?
Product Manager・Goki (hereafter G): This is about whether the feature has an “AI-like quality” or “AI characteristics.” I thought this feature's development was an opportunity to broadly promote the AI aspects of our product "SpeakBuddy" to the public, so I considered this perspective to be important.
In that regard, I initially advocated for proposals with high “AI” characteristics. Proposal C was a straightforward approach, but it’s difficult to explain the function in a single sentence. That was a drawback. However, everyone quickly realized the issue of low engagement and the possibility that some other proposals might be hard to continue using and end up being a one-time feature.
From around the time the “Real Life Success“ concept emerged, I think the team began to see GPT not just as a function but as a tool. Instead of focusing on "features" like free talk and conversation generation, GPT is seen as a tool to bridge the gap between a "learning environment" and a "real-life opportunity*" for acquiring English conversation skills.
*The term "real-life opportunity" here refers to the chance to use English in real-world scenarios.
With that in mind, I came to the conclusion that even a straightforward approach would be acceptable.
Interviewer: How did everyone feel when Proposal C, which led to the current functionality, was decided upon?
S: I had been supporting Proposal C all along, so I felt, 'Yes, awesome!' (She laughs)
H: I was advocating for Proposal B, ProttyChat, due to its AI characteristics, flashiness, and clarity. I had concerns that Proposal C might end up being a bit too plain.
But when he joined the design team and I saw that the design turned out to be quite futuristic and innovative, I realized it wasn't plain at all. Instead, this feature was something that elevated “SpeakBuddy” to another level, and my impression of it changed significantly.
G: I feel the same way. Proposal C of Real Life Success is structurally appropriate. However, if we were to choose this option, I thought it would be crucial to have a design that especially 'looks new' and 'looks fun.' Rico's design fits perfectly with that.
UI/UX, designed by our new Team Member
Interviewer: I’d also like to hear Rico’s thoughts on this. Rico, do you recall your impressions of this feature proposal?
UI/UX Designer・Rico (hereafter R): I joined the company after the proposal was decided, but when I heard all the proposals, I thought we should have pushed forward with Proposal B, Protty Chat. It could have been implemented quickly, and competitors would likely release similar features soon. I felt that “SpeakBuddy” needed to release something early, even if it wasn't perfect.
But since the proposal had already been narrowed down, I decided to put my best effort into making the chosen proposal work.
G: I’d also like to hear from Michael, our Design lead, about this assignment. After all, it was a significant task given to Rico who just joined the company.
M: A lot of timing factors overlapped. I was planning to assign Rico to another project right after he joined, but there was a bit of a delay. At the same time, I had plans to take parental leave. That’s when the BuddyChat design project came up.
From the portfolio we reviewed during the hiring process, we understood that Rico had worked on projects quite different from our previous style. Additionally, our Product Manager, Goki specifically sought “freshness” in the UI/UX for this feature.
I thought that having Rico, rather than someone who has been with “SpeakBuddy” for a long time, handle the design would result in a fresher approach. Additionally, working with our Education Lead, Sayaka would provide Rico with valuable experience in learning experience design. While it seemed somewhat challenging, I believe it was a good decision for everyone involved.
G: I also remember discussing this. Even though it was his first and an extremely important project, we decided to assign it to Rico because we believed he could handle it.
After this, the development proceeded to designing the learning experience.
The development team, which prioritized “learning effectiveness" over all other considerations in planning and narrowing down the options, decided to utilize generative AI/GPT technology not merely as a feature but as a tool to serve the role of bridging "real-life opportunity" and "learning environments." What will happen next on the road to the release of “BuddyChat”? Stay tuned for the follow-up article!