The future of search

Google is planning a significant redesign of its search engine, the New York Times reports. The project, dubbed "Magi," aims to achieve two primary goals. First, Google intends to incorporate a conversational agent, similar to ChatGPT, into its existing search interface. The practical implementation details remain uncertain. Secondly, for the long term, Google is working on an entirely new search engine that leverages the latest advancements in large language models to their full potential.

In both cases, Google will continue to monetize these large language models through ads. The recent launch of Google Bard, which received mixed reactions due to its limited capabilities compared to ChatGPT, demonstrates Google's cautious approach in adopting AI-driven technology. This caution stems from concerns over potential reputational damage resulting from the hasty integration of LLMs without proper calibration and alignment.

The New York Times article highlights Google's unwavering commitment to its core business model—generating revenue through ads. Despite developing new search interfaces and reimagining web experiences, Google remains focused on incorporating ads in some form. I wonder if this business model will stay relevant in the foreseeable future.

For the past few decades, the dominant internet revenue model has been advertising, as evidenced by Google and Facebook's success. However, as the landscape evolves, it's worth considering whether this model will remain viable. ChatGPT, for example, uses a subscription-based model, offering a free version with limitations alongside a paid subscription. As users become more aware of the implications of "free" products (“if it’s free, you are the product” as the saying goes), subscription-based models may become the future of internet services.

While startups can quickly adapt and change their business models, established companies like Google face greater challenges in pivoting their approach. This raises questions about how internet giants will adapt to a changing landscape and if they can successfully transition to alternative business models if necessary.

Enhancing AI Agents: Pursuing Efficiency and Reasoning Abilities

Last week, I worked on creating an autonomous AI agent using GPT 3.5. I developed a version of BabyAGI without Pinecone and using numpy only. I also added an evaluation agent to set success criteria, which helps determine when to stop the autonomous agent process. To improve the AI's efficiency, I set a condition that requires meeting a certain percentage of the success criteria before stopping the process.

One challenge in autonomous agents development is achieving convergence, which is making sure the agent completes necessary tasks quickly and efficiently. The risk if you do not achieve convergence is that the agent never stops creating tasks and may theoretically run indefinitely. To address this, I incorporated an evaluation agent and set specific conditions for success.

Developing AI agents with reasoning abilities is another area of interest. Although GPT-based agents can generate text and appear to think autonomously, their factuality and reasoning abilities are limited. I'm considering working on a reasoning agent that can verify the AI's thought processes.

There are different approaches to AI reasoning, such as symbolic AI, which involve teaching logical rules to AI systems. While this is not a scalable approach, it could be interesting to explore. Combining a large language model with a reasoning agent that can autonomously control logic, factuality, and bias would be an exciting development. I look forward to working on these prototypes and will keep you updated on my progress.