Key takeaways for recent AI Development

Join us as we delve into the exciting frontier of AI

Discover how Perplexity's groundbreaking AI tool, Kyutai's pioneering open-source conversational AI, and Microsoft's sophisticated data structuring tool are revolutionizing AI capabilities. Curious to learn more? Let's dive into the following article.

Perplexity’s ‘Pro Search’ AI upgrade makes it better at math and research.

The AI search startup facing questions about its ethics says its Pro Search is capable enough to ‘pinpoint case laws for attorneys.’

Perplexity has launched a major upgrade to its Pro Search AI tool, which it says “understands when a question requires planning, works through goals step-by-step, and synthesizes in-depth answers with greater efficiency.”

Examples on Perplexity’s website of what Pro Search can do include a query asking the best time to see the northern lights in Iceland or Finland. It breaks down its research process into three searches: the best times to see the northern lights in Iceland and Finland; the top viewing locations in Iceland; and the top viewing locations in Finland. It then provides a detailed answer addressing all aspects of the question, including when to view the northern lights in either country and where.

Perplexity’s AI search tool can also generate a detailed report based on a prompt with a feature called Pages. But recent reports from Wired and Forbes have accused Perplexity of committing plagiarism, with a report from Wired calling the self-proclaimed “answer engine” a “Bullshit Machine,” with animations that misrepresent what it’s doing and data scrapers that bypass rules in robots.txt files.

Key Takeaways:

Moshi: Open-Source Conversational AI Assistant from Non-Profit AI Lab Kyutai

French AI lab Kyutai has introduced Moshi, an AI assistant capable of natural conversations, with an open-source release planned soon. Developed by a team of eight in six months, Moshi stands out due to its real-time speech capabilities and low latency, ranging from 200 to 240 milliseconds. Unlike typical models that convert speech to text, Moshi uses an "Audio Language Model," compressing audio data and processing it as pseudo-words, making it inherently multimodal. The training process included diverse data sources, and voice actress Alice provided recordings for consistent voice synthesis.

Kyutai envisions Moshi revolutionizing machine communication, particularly aiding accessibility for people with disabilities. The demo is available online, with a US-specific link for better latency. Kyutai plans to release the technology as open source, allowing developers to explore and enhance it. The company, founded in 2023, has attracted significant investment and renowned AI researchers due to its commitment to open science and transparency.

Key takeaways:

Discover Complex Data with GraphRAG

Microsoft is excited to introduce GraphRAG, a revolutionary tool for question-answering over private or unseen datasets. Now available on GitHub, GraphRAG offers a more structured and comprehensive approach to information retrieval than traditional RAG methods. With the solution accelerator, users can deploy GraphRAG on Azure in just a few clicks, without writing any code.

What Makes GraphRAG Stand Out?

GraphRAG leverages a powerful large language model (LLM) to create an intricate knowledge graph from text documents. This graph highlights the semantic structure of data, even before any questions are asked. By detecting "communities" of closely connected nodes, GraphRAG partitions the graph into high-level themes down to specific topics, providing a clear hierarchical summary of the data. This means users get an overview without knowing what questions to ask first.

Answering the Big Questions

One of the standout features of GraphRAG is its ability to tackle global questions—those that address the entire dataset rather than just parts of it. Traditional RAG systems struggle here because they only pull from chunks of text similar to the query, often leading to misleading answers. GraphRAG, however, uses community summaries, considering all input texts to deliver accurate, comprehensive answers.

Here’s how it works:

  1. Group community reports within the LLM context window.

  2. Map the question across these groups to generate community answers.

  3. Reduce these answers into a final, global response.

Proven PerformanceMicrosoft tested GraphRAG against traditional RAG and hierarchical text summarization using GPT-4. Evaluating responses based on comprehensiveness, diversity, and empowerment, GraphRAG consistently outperformed, offering detailed and diverse perspectives while using fewer tokens. It was especially efficient in delivering high-level community summaries at a fraction of the token cost.

Innovative Insights and Future Developments

Research shows that LLMs can create rich knowledge graphs from unstructured text, supporting a new class of global queries that traditional RAG methods and costly hierarchical summarization struggle with. While setting up the graph index involves upfront costs, the benefits of structured knowledge and community summaries make it worthwhile for many use cases.

Key Takeaways: GraphRAG on GitHub