• We are a profitable company. The application has been available on the App Store since January. In March, TechCrunch included us in the ranking of the top 10 most profitable AI startups. Currently, we have almost 5 million downloads on the App Store, and the subscription cost varies by country. For example, in the USA, the weekly subscription is $9.99, the monthly is $24.99, and the annual is $99.99. We often conduct AB tests and adjust prices, aiming to understand what keeps users in the app the longest.

Pricing is a crucial aspect in the App Store. What subscription to show to a person first: weekly or monthly? There are many such questions, and everything needs testing. Fortunately, we have an audience on which we can test hypotheses.

• Which subscription is most commonly purchased, and how many users stay in the app?

• I think the weekly subscription is the most popular. The active audience is around 1 million people per month. Regular sales started for us at the end of February. Unexpectedly, we became the first app from Kazakhstan to enter the top 10 App Store rankings in France, Germany, South Korea, Poland, and other countries. This requires a large number of daily downloads, not just a one-time spike in downloads.

Erbol Kopzhasar

Another important metric is usage frequency. The problem with many apps is that they are downloaded from social media, but users forget to open them immediately.

• What is the current valuation of your company?

• I'm afraid to make a mistake with the figure, but I can say for sure that in the last couple of years, we have raised $1.2 million in investments. This includes funds from 500 Startups, MOST VC, and Murat Abdrahmanov. We are currently preparing for a new funding round and aim to raise $30 million.

"The Valley is structured in a way that only fundamental companies can compete there."

• In your view, what is the greatest benefit of the application for users?

• For some, it's the opportunity to quickly get the information they need. For example, how to wash a cat or to hear that everything will be okay. In September, we observed increased activity among students. The chat is also actively used for emails. When we ran ads on TikTok, the most conversion video for a male audience showed how GoatChat helps a guy win back his girlfriend with just one message.

Erbol Kopzhasar

Imagine a service where you can do anything you want. Continuous model development will create new use cases. But an even greater contribution can be made by advancing research direction.

• Why did research become important to you? The application is already profitable.

• In reality, there are enough analogs to GoatChat, and many of them initially had their own language models. Silicon Valley is structured in a way that only fundamental companies can compete there. For the same reason, we invest a lot of money in hardware and specialists. In Kazakhstan, we managed to find only three AI researchers.

Currently, in addition to developers and the product team, the company has five researchers. The field of AI research is developing at a rapid pace. 7-8 scientific papers are published daily, each of which theoretically can change an industry. We have already prepared one scientific publication and are working on several more.

• What topics are you researching?

• Our first paper addresses the area of censoring responses on forbidden topics such as violence. With the example of GOAT 7, we proved that restrictions lead to a deterioration in the overall quality of any responses. All language models, except GPT 4, face the same problem. No one knows exactly how it is actually structured and how it was trained.

Erbol Kopzhasar

Language models have two more unresolved issues. The first is "hallucinations" or the issuance of false information. The second is context length. When chatting, there is a word limit in messages or one session. Currently, the limit for language models is 32 thousand tokens or 16 thousand words. Developers still need to create a model capable of writing a high-quality "War and Peace 2."

We are writing books even now. Our new model generates stories on 128 pages in 5 minutes. In the future, we want to create movie scripts using it. The test version has already been downloaded by 10,000 people. Just specify the plot and main character, and the model will write everything itself.

"Where collecting data in a simple and structured form is not possible, changes will take a long time."

• Will neural networks eventually leave screenwriters and other specialists without work?

• This is a significant danger for the entire AI industry. OpenAI allocated about $10 million for the security of its developments. But this amount is nothing compared to how much the company spends on hardware.

In areas where it is not possible to collect data in a simple and structured form, changes will take a long time. AI optimists believe that neural networks will become more of an auxiliary tool rather than a complete replacement. However, I disagree with them. As soon as McDonald's installed self-checkout machines, the company reduced the number of cashiers. Capital will always strive for increased labor efficiency. Accordingly, you need to look for areas where your competencies cannot be replaced by any language or other model.

For example, in California, Cruise already offers driverless taxi services. There is no time for jokes there now. People are going on strike. And this is happening in the country with the most developed economy in the world. All of this does not inspire optimism, but I think many new professions will emerge. Conditional screenwriters will spend more time editing scripts and coming up with ideas for them than writing from scratch.

• Share your company's near-term development plans?

• We would like our research and models to be usable by others. OpenAI, for example, is primarily a research company that managed to monetize its research. We will continue to release new B2C products or update existing ones with new technologies.

We also plan to fully relocate the team to San Francisco soon. Bakytzhan is already there, and some team members have also moved. We registered in the USA three years ago because we initially wanted to develop in a market where thousandfold growth is possible. By the way, we were offered to buy the Goat.ai domain for a hundred thousand dollars. We never thought that it would be the most profitable investment in our lives.

"In AI, all basic consulting will disappear, especially surface-level questions."

• There are many different LLMs available on the market. How do you understand which language model is the best?

• I approached our ML engineer with this question. The open-source community evaluates models using tests that include thousands of different questions. Their results show how well the model constructs a logical chain of information. For Chat GPT, this score is around 70 points, and for Lama 2, it's 68. The second one, by the way, is open-source, and enthusiasts can fine-tune it, engaging in competitions among themselves. We also played this game. Our Goat-7B received good ratings, and it was downloaded about 10,000 times.

• Share your predictions for the development of the AI industry?

• First and foremost, it's frame-by-frame video generation. Once models can create videos without distortions, people will simply stop trusting information. Imagine that in three years, scammers will send you a video message from your mother, generated by a neural network. Basic deepfake technologies appeared about 5 years ago, but they continue to improve, reaching a 95% similarity of images. On the plus side, using the same logic, you'll be able to create animation.

Also, all basic consulting in AI will disappear, especially some surface-level questions. If the issue of context limitation is resolved, programmers will have to code less. Already now, we don't hire junior developers because there are plugins that write code for developers. But in any case, it needs to be checked, debugged, etc.

Erbol Kopzhasar

Everything will change precisely with the advent of so-called Self-play. Currently, we train neural networks ourselves, but separate models may emerge for this task. In simple terms, one program will train and evaluate another program. New articles on this topic appear every week, but so far, all attempts have failed.

• What about the regulation of the AI industry? Developers have already been urged to slow down the development of artificial intelligence.

• Regulation of AI in the United States will work as capital desires. Let's say Warner Bros. decides to protect its copyrighters, pays $10 million to some senator, and he goes to defend the necessary law. The law is extortion, and if the Democrats come to power, supported by AI companies, there will be corresponding legislation.

• Recently, there has been talk in Kazakhstan about the importance of developing artificial intelligence technologies. In your opinion, how realistic is this?

Erbol Kopzhasar

• To do this, you need to go to Taiwan and ask them to supply at least NVIDIA DGX H100 100 graphics cards to universities. One such card costs $300,000. To conduct research, you need 64 of them. But even if you decide to purchase the hardware, the provider will ask you to stand in line or pay a lot of money. Owning such resources is currently a political and elitist issue.

https://digitalbusiness.kz/2023-11-24/kazahstantsi-zapustili-chat-bot-v-ssha-i-voshli-v-top-10-samih-pribilnih-ai-prilozheniy-vot-kak-im-eto-udalos/