Symbolic Artificial Intelligence: The Future of Chatbots

Augmented Intelligence claims that its artificial intelligence can enhance the usefulness of chatbots.

There exists an alternative to the neural network architectures that underpin many AI models, including OpenAI's o1. This alternative is known as symbolic AI, which employs specific rules related to tasks like text rewriting to solve more complex problems.

Symbolic AI can effectively tackle tasks that pose challenges for neural networks. Recent research has confirmed its scalability potential, although historically, symbolic architectures have not been computationally efficient.

Breakthroughs in scalability have laid the groundwork for numerous startups utilizing symbolic AI across various domains. Examples of such companies include Orby and TekTonic, which develop automation tools for businesses, as well as Symbolica and Likely AI, founded by one of Alexa's co-founders, William Tanstall-Pedoe.

One of the new projects garnering attention is Augmented Intelligence, which has attracted $44 million from investors, including former IBM president Jim Whitehurst.

Augmented Intelligence has developed a conversational AI model called Apollo. The company claims that this model "combines two seemingly opposing technologies—neural networks and symbolic AI—into a coherent, actionable, and learnable model." This makes the outcomes more predictable and "agentic"—a new buzzword in the AI realm—compared to typical neural network-based systems. For instance, instead of merely answering questions about flights and providing booking instructions, a corporate client can leverage Augmented Intelligence's AI to obtain a list of fares and book a ticket for the client, as indicated in the company's newsletter.

But wait—can’t ChatGPT already do this? Yes, acknowledges CEO Ohad Elhelo. However, he emphasizes that this requires much more customization and manual integration than Augmented Intelligence's technology.

"There is a significant difference between chatbots like ChatGPT, which are primarily designed for user interaction, and conversational agents that perform actions or operate on behalf of companies," Elhelo told TechCrunch. "Once you connect AI to tools for information retrieval or action execution, the model no longer relies on its training data, and the quality of intelligence drops sharply."

Augmented Intelligence’s AI can manage chatbots that respond to inquiries on a wide range of topics, such as "Do you set the price for this product?" integrating with existing company APIs and workflows. Elhelo asserts that the AI has been trained on data from tens of thousands of customer service agents' conversations.

Why do companies choose Augmented Intelligence over other AI providers? For instance, Elhelo claims their AI is trained to utilize tools for retrieving information from external sources to perform tasks. AI from OpenAI, Anthropic, and others can also use tools, but Elhelo claims that Augmented Intelligence's AI performs better than neural network-based solutions.

Additionally, the AI is more transparent, providing a log of responses to inquiries and explaining why it made certain decisions, allowing companies to adapt and improve its performance. It doesn't train on company data, only utilizing resources it has permission to access in specific contexts, Elhelo reports.

"Apollo does not require training on corporate information," the company's newsletter states, "and adheres to the instructions of the deploying company."

This aspect of not needing to train on client data will undoubtedly attract companies concerned about confidential information leaks from engaging with external AI. For instance, Apple reportedly prohibited its employees from using OpenAI tools last year due to data leak concerns.

Nevertheless, Elhelo makes some controversial claims, such as that augmented reality AI can "eliminate hallucinations." However, despite this, the company, with 40 employees, continues to attract business, recently forming a strategic partnership with Google Cloud for deploying its models on the platform.

Elhelo did not provide specific revenue information. However, he informed TechCrunch that Augmented Intelligence's latest funding round of $10 million valued the company at $350 million—a rather high valuation for an AI provider just starting to bring its products to market and not founded by a large player in the field.

"Traditional language models are based on transformer architectures, which are successful in pattern recognition and language generation," Elhelo noted. "However, these architectures struggle in situations where models need to perform actions, make decisions, or interact with tools. The neuro-symbolic architecture of Apollo and the capabilities it offers to companies address these issues."