"The Near Future of AI: Action-Driven Schedules and the Rise of AGI"

"The Near Future of AI: Action-Driven Schedules and the Rise of AGI"

In recent years, the field of artificial intelligence (AI) has witnessed significant advancements, paving the way for exciting possibilities in the near future. One particular model that has garnered attention is the ReAct model, introduced by Yao et al. in 2022. This model takes a three-step approach, consisting of Thought, Act, and Observation, to enable action-driven AI.

The concept of action-driven AI is revolutionary, as it allows AI models to act as agents that make choices and take actions. This approach brings us closer to achieving Artificial General Intelligence (AGI), which has long been a topic of debate among academics. LLMs (Language Models) have shown promising results in various tasks, especially in question-answering when prompted to "think step by step" (Kojima et al., 2022). However, by incorporating external cognitive assets, such as fetching data from external sources, the performance of LLMs can be further enhanced.

OpenAI's 002-text-davinci model has been particularly successful in harnessing the power of external resources. This model combines instruction tuning and Reinforcement Learning from Human Feedback (RLHF), where humans rate the success of a given prompt. This approach allows for iterative improvements and fine-tuning to produce better results. However, the true potential lies in actual reinforcement learning, where the system can be trained to achieve better outcomes based on a specific metric of interest.

In the realm of AI startups, those that can create powerful feedback loops are likely to thrive. By identifying and solving customer pain points, even starting with simple solutions, these startups can collect valuable data on how to improve their offerings. They can then train their models to be more consistent and iterate on their products or services. This iterative process becomes the foundation for building a competitive advantage or a "moat" in the AI landscape.

Interestingly, the concept of scheduling also plays a crucial role in optimizing productivity in the world of AI. Paul Graham, in his essay "Maker's Schedule, Manager's Schedule," highlights the different approaches to time management between makers (such as programmers and writers) and managers. For makers, having a meeting is disruptive and akin to throwing an exception in their workflow. They prefer to work in larger time units, like half a day, as it allows them to delve deep into their creative process.

On the other hand, managers operate on a different schedule, one that revolves around commanding and coordinating tasks. Meetings are an integral part of their routine, but they often fail to realize the toll it takes on the productivity of makers. This misalignment can hinder ambitious projects that require intense focus and sustained morale.