The near future of AI is action-driven, according to the ReAct model developed by Yao et al. (2022, arxiv). This model takes three iterative steps: Thought, Act, and Observation. It emphasizes the importance of incorporating cognitive assets like search into the actions taken by the AI system. This action-driven approach is what will truly lead to exciting and advanced applications of AI, resembling something close to AGI (Artificial General Intelligence).

The near future of AI is action-driven, according to the ReAct model developed by Yao et al. (2022, arxiv). This model takes three iterative steps: Thought, Act, and Observation. It emphasizes the importance of incorporating cognitive assets like search into the actions taken by the AI system. This action-driven approach is what will truly lead to exciting and advanced applications of AI, resembling something close to AGI (Artificial General Intelligence).

LLMs (Large Language Models) are known for their ability to perform well in question-answering tasks, especially when prompted to "think step by step" (Kojima et al. 2022, arxiv). However, they can achieve even better results when given access to external cognitive assets. By fetching data from external sources, LLMs can bridge the resource gap and enhance their performance. This suggests that external resources play a crucial role in the effectiveness of AI models.

OpenAI's 002-text-davinci model has demonstrated remarkable capabilities, and its success can be attributed to a combination of instruction tuning and Reinforcement Learning from Human Feedback (RLHF). With RLHF, humans rate the success of a given prompt, allowing the system to learn and improve over time. However, the true potential lies in actual reinforcement learning, where AI systems can be trained to produce better results based on specific metrics of interest.

In the AI landscape, some startups will thrive by creating powerful feedback loops. They will identify a customer pain point, start with a simple solution, collect data on how to improve, train their models, and iterate. This iterative process will lead to more consistent and effective AI models, creating a competitive advantage in the industry. This can be seen as the current version of a moat in the AI field.

As AI agents become more domain-general, the possibilities for automation and offerings will expand. The spaces that can be automated will increase, leading to a wider range of applications and services powered by AI. This opens up new opportunities for businesses and entrepreneurs to tap into the potential of AI and create innovative solutions.

The pay-to-surf model, which gained popularity in the late 1990s, offered a different approach to online business. Companies would share advertising revenue with their users as a form of reward for watching promotional content. However, this model faced challenges, including attempts by individuals to defraud the companies and spamming issues. As a result, many companies had to terminate user accounts. The surviving companies in this space now operate on a rewards-based structure, where users earn points by surfing the web or completing tasks, which can be exchanged for gifts.

Brave, a browser known for its privacy features, has introduced an alternate compensation model for browsing. Users are given tokens that are promised to be exchangeable for dollars in the future. This operates in a similar manner to cryptocurrency and provides users with a different way to benefit from their online activities.

Combining these two seemingly unrelated topics, we can see a potential synergy. The pay-to-surf model, with its focus on sharing rewards with users, aligns with the action-driven approach in AI. By incorporating the concept of external cognitive assets, users could be rewarded for allowing AI models to access their browsing data and perform actions on their behalf. This would create a win-win situation where users benefit from the AI's capabilities while being compensated for their participation.

To leverage this potential synergy, here are three actionable pieces of advice:

1. For startups entering the AI space, consider building feedback loops into your product development process. Collect data from users and use it to train your models, continuously improving their performance. This iterative approach will help you create more effective AI solutions and establish a competitive advantage.

2. Explore partnerships with companies operating on a rewards-based structure, such as those in the pay-to-surf space. By incorporating their user base and reward systems, you can create a mutually beneficial ecosystem where users are compensated for their participation and AI models can access valuable data to enhance their actions.

3. Embrace the power of external cognitive assets. Give your AI models access to external resources and data sources to bridge the resource gap and improve their performance. This can be done through partnerships with relevant platforms or by developing your own data acquisition strategies.

In conclusion, the near future of AI lies in an action-driven approach, where AI models act as agents making choices and performing actions. The incorporation of external cognitive assets and the utilization of feedback loops will be crucial for the advancement of AI capabilities. By connecting the concepts of action-driven AI and the pay-to-surf model, new opportunities for collaboration and innovation can emerge. As AI becomes more domain-general, the possibilities for automation and offerings will expand, creating a dynamic and exciting future for AI-driven applications.