Welcome, to this week’s AI To the World Newsletter. 

This newsletter is a summary of our weekly AI meetups in Bali. You are welcome to join us online or in person – 7pm Bali at B-Work – where we discuss the latest news and tools.

This week was certainly a big week in AI with new upgrades to the consumer product of Chat GPT and a fascinating paper doing the rounds making some bold predictions about how we can tackle one of the biggest problems in AI: energy efficiency. 

But does it actually hold water? The group was skeptical. 

“As ever, some expert opinions from our esteemed team of engineers, academics, and business people as we explore practical solutions for integrating AI into business and cutting edge research.”

– Harry 

Attendees

1. Harry Verity (host)
2. Will (Intelligent Marketer and Product Manager, joining from Australia)
3. Bengi (Joining from Turkey)
4. Alan (ML Engineer)
5. Ela (AI Automation and AI Training, Joining from Poland)
6. Nuno (Design at Burger King)
7. Nuno (Nuno’s boss at Burger King)
8. Dr. Bron (AI Training for Academics)
9. Justin (Y Combinator Graduate)
10. Eric (CEO, Empathy School, Bali) 

Key Takeaways

• OpenAI released Canvas, a new interface for ChatGPT allowing real-time collaboration and editing
• A research paper claims to reduce AI model energy consumption by 95% through using addition instead of multiplication
• Anthropic introduced “contextual chunking” to improve retrieval-augmented generation (RAG) accuracy by up to 67%
• OpenAI’s new voice model shows promise but currently lacks RAG capabilities, requiring workarounds
• The group expressed skepticism about the energy efficiency claims in the “Addition is All You Need” paper
• Participants shared various use cases and experiences with AI tools in their respective fields

Topics

1. OpenAI Canvas Release

Opinion:

• Will expressed interest in the coding aspects but Noted that Cursor already offers similar functionality
• Harry Verity highlighted the potential impact on content creation tools and workflows

2. Energy Efficiency in AI Models

• Paper titled “Addition is All You Need” claims 95% energy reduction for element-wise operations
• Proposes replacing complex floating-point multiplications with integer additions
• Aims to address high energy consumption of large language models (e.g., ChatGPT using 564 MWh daily)

Opinion:

• Alan (ML engineer) expressed skepticism about the bold claims and requested more time to review the mathematical details
• The group collectively raised concerns about the lack of information on the authors and their company
• Harry Verity noted the importance of lifecycle analysis and environmental product declarations in addressing AI energy consumption

3. Improved RAG Accuracy

• Anthropic introduced “contextual chunking“ for better retrieval-augmented generation
• Adds context to document chunks, reducing information retrieval error rates by up to 49%
• Combined with reranking, improvements of up to 67% possible
• Implementation guide with code samples available on GitHub

Opinion:

• Will compared it to graph RAG and vector embeddings, expressing interest in experimenting with the new technique
• The group discussed similarities to traditional qualitative research methods in contextual analysis

4. OpenAI Voice Model

• New advanced voice model released, but not yet available via API
• Lacks ability to add RAG or extensive information
• Workaround: Using text-to-speech (e.g., ElevenLabs) to read long prompts for context

Opinion:

• Will shared his experience using the voice model for practicing presentations and brainstorming
• The group discussed potential applications in sales calls and content creation
• Concerns were raised about the model’s tendency to cache information between sessions