AI World. Generative Artificial Intelligence News - Week 2/52 2024

Improving LLMs for Complex Logical Reasoning
LLMs are great, but even though they are incredible at natural language, they struggle with logical reasoning, especially when they need real-world knowledge.
Knowledge Graphs to the Rescue: These graphs represent factual knowledge through entities, relationships, and logical rules. They are great for data, but not so good at generating fluent textual explanations.
Tree of Thought (ToT): This approach breaks down problems into coherent thought steps and explores different reasoning paths. It’s like mapping a problem into a tree where each node is a partial solution. However, it relies heavily on language generation at each step.
Putting It All Together: Combining knowledge graphs with ToT boosts the reasoning power of LLMs. For instance, in a deductive task, each application of a deductive rule becomes a step in the process.
Thought Generator and State Evaluator: The LLM creates candidate thoughts for each reasoning step and evaluates each state (the current sequence of thoughts). This applies to various types of reasoning, such as deductive, inductive, abductive, and analogical.
Breaking Problems Down into Steps: Instead of tackling a problem in a single step, ToT breaks the process down into smaller, more manageable steps, each with its own logic and evaluation.
Imagine it’s like having a super-smart virtual assistant that not only knows a lot of things, but can also think and reason step-by-step to solve complex problems. It’s like giving an LLM a knowledge treasure map and a compass to navigate it. Pretty mind-blowing!
Replacing Siri and Google
John decided he was fed up with assistants like Siri and Google, so he built his own smart home voice assistant. But not just any assistant—a sarcastic one, running entirely locally without relying on the internet. He used some powerful graphics cards and HomeAssistant. His assistant can control the whole house and replies with funny, sarcastic comments. Pretty impressive, right? If you want to see how he did it, check out his website.
Source: https://johnthenerd.com/blog/local-llm-assistant/
Automatically translated from the Spanish original.