The current AI models, such as large language models (LLM), are essentially advanced 'guessing games' that can be nonsensical, costly, and not stable enough; even with more computing power, progress is limited.

Don't just focus on upgrading the models; the key is to build good 'surrounding equipment': smart routing (which problems to solve with APIs and which with LLMs), security guards (to prevent errors), smart caching (cost-effective and fast), and model selectors (to choose cost-effective models).

This way, AI can help ordinary people achieve great things, like Canva, rather than just being a flashy tool.

For example, if you ask Klok 'What is the price of Bitcoin today?', it won't let GPT-4 guess randomly but will check the API directly, making it fast and accurate.

Additionally, their 'blind testing' for model selection identifies affordable and effective alternatives like GPT-4o mini, reducing costs by 90% while improving quality!

Klok has positioned itself as a 'living example' of this approach, encouraging everyone not to fixate on flashy demos but to focus on creating production-level AI applications.