Section Insights
Introduction to AI Query Limitations
Why do AI responses vary in quality based on the user's understanding of the topic?
AI often provides average answers for nuanced questions due to its training on summarization rather than specific data.
- AI excels at generating general responses but struggles with nuanced queries.
- The quality of AI responses can be misleadingly confident despite lacking depth.
- Understanding the limitations of AI can help users frame better questions.
Challenges in Achieving Expert-Level AI Outputs
What are the challenges in improving AI query responses?
The main challenge is the economic cost of enhancing AI models to produce expert-level outputs.
- Improving AI responses requires more extensive data searching during training and runtime.
- Current economic constraints limit the ability to achieve high-quality AI outputs.
- AI agents can ask more specific questions compared to human users.
Innovating AI Information Access
How is Kinabalu addressing the limitations of current AI search methods?
Kinabalu aims to provide frictionless access to information by innovating index structures and creating a self-learning loop for AI agents.
- Kinabalu is designed to optimize AI information retrieval processes.
- The platform focuses on understanding how agents seek information to improve search efficiency.
- Innovative index structures can enhance the performance of AI models.
Introduction to Web Query Language
What is the web query language and how does it improve search capabilities?
Web query language allows users to interact with the internet as if it were a database, enabling more nuanced searches.
- Web query language simplifies complex queries into actionable searches.
- It enhances the ability to find specific information, such as product availability and pricing.
- The technology is designed to address common search challenges faced by users.
Practical Applications of Web Query Language
How does web query language facilitate efficient information retrieval?
It can scan relevant documents, extract prices, and provide evidence of product availability with a single query.
- Web query language streamlines the process of gathering detailed information.
- Users can obtain comprehensive data with minimal input, enhancing productivity.
- The technology is positioned to improve user experience in searching for specific products.
Transcript
0:02 Hello everyone. My name is Andrey Stiskin. We are building web search and web query language for AI. you all have this experience when you ask AI any question, any random question, the answer is perfect. But if you ask the question that you really, really understand, the answer is like and it is not wrong, it is like very, very average. Let's dive deeper into this phenomenon. Let me give you example of my favorite question.
0:33 So, the question is, what is the best search API? And you look, amazing table, use cases, clear winner, justification. It looks like this AI just went to business school because what is missing? Benchmarks. No data. It is absolutely ungrounded and absolutely confident recommendation. So, why it happens? Because LLMs trained on such content. They are trained, they are amazing summarization machines that train on variety of the data and for this specific, for each specific nuanced topic, the result is the result is average.
1:16 What should we do with this? What is the problem? How to achieve expert level of outputs, of answers? Clear, models should just try harder. They should search much more during runtime and during training to get the best, not average type of content, reason over it, and produce answers. Sounds easy. What is the problem? What is the bottleneck? Right now, it is economically extremely hard to do this loop because of cost.
1:55 ChatGPT makes around one query per three messages, which is around one such query per 5,000 token generated tokens, which is 20 cents per $1. Huge. It is slow. And also agents, they search differently. In comparison to humans who are lazy, agents, they have a goal. They have a purpose. They are not lazy. They can generate very, very specific questions. They use quotes, sites, date filtering in order to achieve the goal. And all the indexes that exist in the world were optimized for human query traffic.
2:38 so, we are building Kinabalu to address this bottleneck and give all the models and give all the agent builders frictionless access to information. We are innovating index structures and we are creating self-learning loop that continuously improve and understand how agents ask to seek information and seek what they need. On top of this, we provide web search API search and fetch endpoints. And also, we provide web query language, which is literally the concept that you operate with internet as like database.
3:13 >> >> who we are? This is our third query. So, so this is our third search engine. I was CEO of Yandex Search and my co-founder, Matias, was main scientist behind Amazon web search. Let me give you an example of web query language that I touched. This is a question. Sounds pretty simple. What is the cheapest available H100? Many of you ask this on daily basis. And it is nuanced because cheapest and available, because if you if you like try to find the answer on this question, maybe availability would be the problem because it is always contact sales.
3:56 But, if you put it into web query language, it will do scan the relevant papers, extract prices, extract feedback, and find the evidence that somebody used this and was able to get some GPUs from this provider with date, which is important signal if it it will work. So, imagine it is possible with just one prompt and one query, and it is possible with web QL. So, behind this QR code, you can get free access to web search API and put into waitlist for web QL till end of the August. So, give, please, your AI access to the best knowledge, and we are keen to hear feedback from you.
4:47 and to connect back to the my favorite question, what is the best web search API? Here are benchmarks. >> >> So, thank you.
Summary
- Current AI models often provide average answers to nuanced questions due to their training on summarization rather than specific data.
- The economic cost and inefficiency of querying during AI training and runtime limit the quality of responses.
- Kinabalu aims to address these issues by providing frictionless access to information and innovating index structures for better data retrieval.
- The platform includes a web search API and a web query language that allows users to interact with the internet as if it were a database.
- An example of the web query language shows its ability to extract specific information, such as the cheapest available H100 GPUs, by scanning relevant data and providing evidence.
- Stiskin invites feedback and offers free access to the web search API and a waitlist for the web query language.
Questions Answered
Why do AI responses vary in quality based on the user's understanding of the topic?
AI often provides average answers for nuanced questions due to its training on summarization rather than specific data.
What are the challenges in improving AI query responses?
The main challenge is the economic cost of enhancing AI models to produce expert-level outputs.
How is Kinabalu addressing the limitations of current AI search methods?
Kinabalu aims to provide frictionless access to information by innovating index structures and creating a self-learning loop for AI agents.
What is the web query language and how does it improve search capabilities?
Web query language allows users to interact with the internet as if it were a database, enabling more nuanced searches.
How does web query language facilitate efficient information retrieval?
It can scan relevant documents, extract prices, and provide evidence of product availability with a single query.