Dennis had spent about a week fighting a chatbot. He was building an AI "language exchange companion," a conversation partner from another culture for people learning a language. He said it had been hard to make an AI such as the GPT-5 API behave like a real person from a different culture. The model kept falling back into what he called a "very stubborn assistant mode": How can I help you? What can I do for you? In the end, he wrote the same instruction into the system prompt five times: do not ask a question at every turn. A system prompt is the set of hidden instructions a developer gives a model before any conversation begins.
Dennis described the problem during the second audience Q&A ("Ask Me Anything") round of Moonshots with Peter Diamandis. The episode was recorded on September 18, 2026, and published on September 22. Earlier in the same episode, another founder based in Japan asked about a different side of the subject. The two questions cover two problems with AI language partners. One is what learners actually want from the product. The other is how to make a model behave like a person rather than a service desk.
"They want to be fluent and funny and interesting"
The first caller introduced himself as Alex. He said he is originally from Bulgaria and moved to Japan about 30 years ago. His company provides English-language training for universities and corporations and has operated for 17 years. Over that time, it has built its own learning platform, curriculum and assessment systems, along with a large store of teaching content and data. He said the company is adding AI to the product, "probably not fast enough." Its goal is to turn a successful education-services business into something more scalable and technology-driven. He asked Dave Blundin, founder and general partner of the investment firm Link Ventures, to answer as a venture capitalist. What would the company need to build or prove over four to five years to interest an investor? And what would show that it was still fundamentally a services business?
Blundin did not start with the technology. He described English-language learning in Japan as a $10 billion business, "probably 20 billion by now." Then he argued that the product was being defined too narrowly. "When you talk to the students, they don't want to learn English," he said. "They want to learn to be fluent and funny and interesting in English."
In his view, that calls for more than an AI avatar that teaches correct speech. The AI should be able to tell a student whether a remark was genuinely funny or entertaining to someone in America or England. It should also be able to say whether a word was pronounced the way people say it in one part of Ireland rather than in Liverpool. Blundin then took the argument further. People learn English because they want to change their lives, he said. After that, they want to be entertaining and smart and to have a life plan. Once learners are "hooked on learning a language," he said, becoming their life-coach platform is "a very natural segue." He estimated that the business model "scales to many hundreds of billions of dollars when it moves from language to life plan." Alex replied that it was something he had never thought of.
A companion, not a help desk
Dennis approached the subject from the side of building the product. He told the panel he has a background in East Asian studies and has spent a lot of time learning Chinese and some Japanese. He is currently in Japan. From that experience, he said, people first seek out someone from another culture for the language. Over time, the relationship becomes more about friendship and companionship. He described it as projecting "unfulfilled social needs" onto someone from another culture who seems like a blank slate. His app tries to recreate that feeling. It is nominally about language learning but mainly about companionship.
He called the resulting challenge "my own version of the alignment problem": getting the model to behave like an actual person from a different culture. His repeated system-prompt instruction stopped the constant questions. He found it much harder to teach the model to take initiative, show open curiosity and gradually build a picture of the user. He tried adding milestones: after a certain number of turns, the AI should know a certain fact about the user. What he could not achieve was making this feel natural and organic, like a relationship. He asked the panel which approaches and tools he should try.
The panel's answer: fine-tuning
Alexander Wissner-Gross, a computer scientist and founder of Reified, called the case "a poster child for fine tuning." Fine-tuning means taking an existing model and training it further on material chosen for one job. The desired behavior is then built into the model instead of being requested through instructions. Wissner-Gross said OpenAI had decided it was no longer interested in fine-tuning because no one was using its fine-tuning API, then decided it was interested again. He pointed Dennis to both open-source and closed-source tools and described two routes.
The first route is supervised fine-tuning to achieve what he called "style transfer," which he thought was what Dennis wanted. OpenAI's supervised fine-tuning guide describes training a model on example prompts paired with the responses you want, so its style and behavior become more consistent. One of the uses it lists is correcting cases where a model fails to follow instructions. The guide recommends setting up evaluations first, meaning tests that show whether the model is improving. It suggests starting with about 50 well-crafted examples. The tuned model should then be compared with the original on examples held back from training. For a companion app, those examples would be conversations written in the voice the developer wants. The model would learn the pattern from them instead of being told again and again what not to do.
The second route is reinforcement fine-tuning. Wissner-Gross said it applies if you can measure how well a model performs "in some quantitative way." OpenAI's reinforcement fine-tuning documentation describes a different training signal, one that does not rely on copying examples. For each prompt, the system produces candidate answers and scores them with a grader that the developer defines. It then shifts the model toward the behavior that scored higher. Graders can combine simple checks on an answer's format with judgments from another model about its quality. The documentation warns that a model can improve against one grader while neglecting another requirement. It suggests comparing training and validation scores to catch this kind of overfitting. For a companion like Dennis's, the difficult part would be writing a grader that measures curiosity and initiative rather than something easier to count.
The supervised fine-tuning page also carries an undated notice. It says OpenAI's fine-tuning platform is being wound down and is closed to new users. Because the notice has no date, the documentation does not show how it relates in time to the renewed interest Wissner-Gross described. He also mentioned open-source tools as an option.
Blundin's suggestion: an open model
Blundin called the advice great and suggested where to start among open-weight models. These are models whose trained parameters, called weights, can be downloaded and modified. He suggested "the new GLM model" without naming a specific version. He described it as having fewer parameters but still a very long context window, meaning it can take in a lot of text at once. He said that gives more flexibility. He added that he finds it "less locked into its training" than closed-source models. His advice was to find open-source fine-tuning code, layer it on top of GLM, and see whether Dennis "might be able to manipulate it a lot more."
Blundin also recalled a time when fine-tuning was almost unavoidable. With early models such as GPT-2 and GPT-3, "you almost had to fine tune it to make it do anything," he said. That practice later faded.
The fine-tuning the panel described would shape a product's manner through a training process that the developer designs, using chosen examples or a scoring rule. That differs from a model that keeps learning from each conversation with each user. The advice did not directly address the part Dennis had tried to force with milestones: building up a picture of the user over time. Blundin ended the exchange with a request: "If you crack the code, definitely check in with us and tell us how you did it."