My History of Failures and Learnings in AI Product Development

About two years have passed since I was shocked by ChatGPT, and I have created several products as an individual or in small teams. However, all of them failed in terms of monetization, and as a result, I withdrew from them. Among them, I will introduce the main ones along with the situation at that time and look back on why I withdrew.

The reason for publishing the history of failures is that I hope it will be helpful for those who are also taking on challenges in the AI field. Also, I want to convey the necessity of why the current product "Evame," which I finally arrived at through these experiences, was born.

Past Projects

AI Agent Autonomix (2023/02~04)

Situation at the time

In February 2023, I became an LLM unemployed.

The first time I realized the potential of LLMs was with GitHub Copilot, released in June 2021. Initially, it was only a very localized support, but by the beginning of 2022, it had evolved to the point where it reduced manual code input by 20-30%, and even I, a web engineer not familiar with machine learning, clearly felt the future.

I was fascinated by LLMs, but at the time, LLM-related job openings were mainly for researchers, and it was difficult for web engineers to enter the field.

Then, like many others, I was convinced that "the world had changed" with the advent of ChatGPT in November 2022. I proposed an LLM business to my employer and discussed it with the president for about a month, but I was unable to move the company, partly because my presentation was immature.

I decided to create a product myself and resigned from the company in February 2023. At that time, the LLM boom was not as big as it is now, but through deepening exchanges with some people who were enthusiastic on X (Twitter), I felt that AI agents would become mainstream in the future.

When I released the LP with the concept of "replacing work," I immediately had more than 500 people register on the waitlist, and I felt a clear need, so I developed Autonomix.

Product Overview

An AI agent that automatically executes chat instructions → email replies, SNS posts, etc. via APIs such as Zapier.

Reasons for Withdrawal

It's a common issue with AI products, but I fell into the swamp of spending an infinite amount of time and money on improving accuracy after creating an attractive prototype.

Specific technical issues:

1. The success rate remains at about 80%, which is unacceptable for practical use.

2. GPT-4's external tool calling is unstable. As a result of repeated retries and failsafes, the logic became bloated, and costs soared due to multiple LLM calls. It costs 3 yen just to send an email.

However, there was a more fundamental problem. To realize a truly easy-to-use UX, it should operate directly within platforms such as Gmail and Twitter. And there is no reason why Google and Twitter would not provide this feature.

Faced with both technical and strategic challenges, I decided that "I can't beat Big Tech" and decided to withdraw.

LINE Bot Talk-chan (2023/04~08)

Situation at the time

After the failure of Autonomix, I realized the limitations of tool-based software. In the words often used by startups, there is no Moat (competitive advantage). After all, if Google or Microsoft get serious, they will provide equivalent or better functions for free.

So I changed my strategy and looked for areas where Big Tech would be less likely to enter. That was the field of entertainment, especially AI characters. In this area, it's not simply a matter of throwing in money to create something good, but rather human and creative elements such as the charm of the character and building relationships with users become important. I thought there was still a chance to win here.

There are two main approaches to AI characters. The platform type, like Character.ai, where users create their own characters, and the content type, where the operator prepares attractive characters. I chose the latter. I wanted to create an AI like Doraemon that encourages users and supports their growth.

At that time, there was no concept of RAG, and there was no AI with long-term memory. I thought that if I applied the technology I had acquired so far, I could achieve long-term memory, and the relationship built by it would become a Moat. And that's how Talk-chan was developed.

Product Overview

An AI friend with long-term memory. Hybrid search of conversation data stored in vector DB and graph DB, and conversations while referring to what is now called RAG.

Reasons for Withdrawal

There were a few users who paid, but it did not reach a level where the business could be continued. There were two fundamental issues.

First, there is the cost problem. In order to realize long-term memory, it is necessary to constantly refer to past conversation data, and the amount of token consumption becomes enormous. In order to make a profit, it was necessary to set a monthly fee of about 3000 yen.

However, we were not able to provide enough value to justify that price. After all, even though it has memory, it is just an AI that "can only chat." We were not able to create an attractive experience that would make people want to use it even if they paid 3000 yen per month. This was fatal.

In other words, even though it was technically possible, we could not solve the structural problem of "high cost and low added value," and we decided to withdraw.

We should have assumed the AVC (Accuracy Value Curve) of how much added value the product we are making can actually provide.

Reference article

https://note.com/takahiroanno/n/ncb7d77bfd9f1

Stuffed Animal AI (2023/08~10)

当時の状況

After the failure of Talk-chan, I thought about what an "AI character worth paying for" is. As an answer to that, I focused on the importance of physicality.

The amount of information is much richer with voice and body than with simple text chat. I thought that this would make people feel that it was worth the price. With that in mind, I decided to develop a stuffed animal AI that can converse and express movements.

Product Overview

A Wi-Fi connected stuffed animal equipped with a microphone, speaker, and motor to heal loneliness with voice interaction and comical movements.

Reasons for Withdrawal

As I progressed with development, I saw the reality of the hardware business. Manufacturing, quality control, logistics, customer support... This is an area that requires 10 years of commitment.

So I asked myself, "Do I want to spend 10 years on this business?"

The answer was "NO." The reason was simple: I didn't think this product was really needed in the world. Even if a person suffering from loneliness is temporarily healed by talking to a stuffed animal, I thought it was only a symptomatic treatment.

I myself have suffered from mental problems such as bipolar disorder. What brought about a fundamental solution at that time was the wisdom of early Buddhism, centered on a meditation called Vipassana. Vipassana meditation is a practical method taught by the Buddha so that we can surely experience enlightenment. I thought that what people suffering from loneliness really need is the right wisdom to support inner growth, such as meditation, and not a toy to distract them.

In other words, there was no product-founder fit at all. I can't spend 10 years on something I can't wholeheartedly believe in. I decided that and decided to withdraw.

Looking back

Looking back now, it's not that each product was bad. In fact, there are products that are attracting a lot of attention, such as Devin and Manus, in the AI agent space, and there are several success stories such as Oz Chat in the content-type AI character space. Some people may be saved by pet robots like Lovot. If I had continued to work on it seriously, I might have been able to achieve some results.

It's not that these products were bad, but that they weren't what I really thought I should be doing.

Bonus Track - Other Experimental Products

After that, while working on commissioned projects, I created some experimental products without thinking about monetization while searching for what I really wanted to create.

Subsidy Selection AI

I realized after I resigned that I didn't know anything about the various benefit programs, including unemployment insurance. I thought that there should be many people like me, and it would be helpful to have an AI that guides them to the appropriate programs.

However, as I progressed with development, I judged that each system was too complex with many exceptions, and that it was beyond the limits of current LLMs, so I gave up.

Reference article

https://note.com/tkgshn/n/n4fd785cf1381

Translation Extension

I developed it because I wanted to use LLM translation when browsing overseas sites because LLM can do more natural translation than Google Translate. However, translating the entire page resulted in high API charges, and it was very slow compared to Google Translate, so the UX was poor for practical use.

Buddhist AI

A chatbot with the Buddhist scriptures as a data source. If you ask AI normally, the teachings of early Buddhism and the teachings of Mahayana Buddhism are mixed, and it gives contradictory answers, so by limiting the data source, I made it answer based only on the teachings of early Buddhism.

What I'm making now - Evame

Current situation

In addition to these, there are a lot of things that I worked on for about a week. What I realized through these experiences was that I can only be serious about things that I sincerely believe are valuable. And what I really think is valuable is something that contributes to people's fundamental happiness.

For example, a mechanism to significantly reduce living costs, an app that makes learning fun and gamified, and something that conveys the wisdom of Buddhism, including Vipassana meditation, to the modern age.

Evame's concept

While continuing to explore, I thought about creating a product related to Vipassana meditation.

At first, I thought about creating an AI that teaches Vipassana meditation, but I changed my mind because I thought it would be very difficult. This is because meditation guidance requires delicate responses tailored to individual situations, and a deep relationship of trust between the learner and the instructor is essential.

So, I decided to publish the modern Japanese translation of the Buddha's teachings in full for free.

The Buddhist scriptures (Tripitaka) are valuable documents containing the fundamental teachings of Buddhism, but the full Japanese translations currently available are from before and during the war, and there are no complete translations in modern language. Also, what is available for free is limited.

I thought that if I built a translation system that utilizes LLM and created a mechanism for users to evaluate translations and make improvement suggestions, I could produce a better translation than before.

Current vision

While focusing on the translation of Buddhist scriptures, I also thought about publishing public domain texts from around the world because I wanted to widely share knowledge and stories from around the world. However, the cost of the database became enormous, and it was not realistic to publish it for free.

So, I reorganized my policy and decided to continue publishing public domain texts, including Buddhist scriptures, for free, while adding a blog platform function where users can post articles in order to achieve sustainable management.

That's this site, Evame, where this article is now.

Users can simply write articles here, and multilingual translations will be generated automatically, allowing them to deliver their words to the world.

You can also use a mechanism that automatically improves translations. For example, if you set the display language of this article to a language other than Japanese and click on the displayed translation, you should see voting and additional UI.

Currently, free members can write articles in up to 2 languages, but we plan to monetize by enhancing services for paid members, such as supporting up to 4 languages for paid members. Currently, the paid membership function has not been implemented yet, but we plan to operate with this model in the future.

I want to make Evame a global version of note, the world's best blog community site that surpasses Medium.

In other words, Evame is a platform that provides both "free publication of public domain content such as Buddhist scriptures" and a "multilingual blog platform." We will provide social value with the former and achieve sustainable management with the latter. We believe that this combination will allow us to continue delivering valuable information to the world in the long term.

Even if monetization doesn't go well, I will continue to develop as long as I can. This is because I am convinced that the world needs it.

Summary

The most important thing I learned through my failures so far is that I can't work on something in the long term unless I sincerely believe in its value.

Technical feasibility and market needs are important, but more than that, I feel that whether or not there is a "product-founder fit" is crucial for developers.

I will write about Evame in more detail in another article, but please try writing an article once.

Thank you!!!