You are currently viewing Llama 3.2 Mobile Upgrade, OpenAI’s Shift, and RAG Innovations

Llama 3.2 Mobile Upgrade, OpenAI’s Shift, and RAG Innovations

AI Founders Meetup 1st October 

Hi Friends,

This week was a big week for model upgrades with Meta bringing out a significant upgrade to Llama 3 which now allows you to run the language model on mobile.

Our CTO friend Frank only ever seems to swing by when there’s a model upgrade to Llama; the last time he came was the week Llama 3 came out and he was pretty chuffed that he’d managed to get it working on his laptop. So it was fitting that he swooped in to give us his opinions on 3.2 this week. 

We also discussed non-technical RAG and the challenges involved with generating Wikipedia articles using AI, something researchers at Stanford have been digging their teeth into. 

Let’s dive in

Attending This Week…

  • Harry Verity (Host)
  • Eric (Empathy School founder)
  • Lucas (AI service developer)
  • Frank (Tech professional, CT)
  • Alan (ML Engineer)
  • Douglas (Social media marketer from Malaysia)
  • Reda (AI advisor and investor)
  • Dr Bron (AI Training for Academics) 

Key Takeaways

  • OpenAI shifting to more profit-driven model; key founders leaving
  • Llama 2 released with vision capabilities and mobile versions
  • New tools emerging for secure, non-technical RAG implementations
  • Wikipedia article generation being automated by Stanford researchers
  • Group members expressed concerns about AI data privacy and security
  • Discussions highlighted the rapid pace of AI tool development and potential impacts on various industries

Topics

1. OpenAI Business Model Shift

  • OpenAI moving to for-profit model; key founders like Mira Murati leaving
  • Rumors of ChatGPT price increasing to $44/month by 2028
  • Shift from academic/research focus to commercialization
  • Concerns raised about ethical implications and original mission

Opinion:

  • Frank expressed skepticism about the true motives behind the shift
  • Alan noted the contrast between OpenAI’s original mission and current direction
  • Some members speculated on potential government influence in AI development

2. Llama 3.2 Release

  • Meta released Llama 3.2 with vision capabilities and mobile versions
  • Two small models (1B, 3B parameters) for smartphones
  • Two large models (11B, 70B parameters) with image understanding
  • New architecture using adapter weights to integrate image and language models
  • Outperforming Claude and GPT-4 on some benchmarks
  • Available on Hugging Face for download/experimentation

Opinion:

  • Frank shared his experience running Llama 3 on his laptop locally
  • Alan provided technical insights on the model architecture and distillation process
  • Group discussed the significance of running AI models on mobile devices

3. RAG as a Service

  • New tools emerging for easy, non-technical RAG implementations
  • Allows businesses to create AI agents with private knowledge bases
  • Addresses data privacy/security concerns with GPTs
  • Example shown: $29/mo for 5000 messages, though may be limiting for larger orgs
  • Debate on true security/privacy guarantees of such services

Opinion:

  • Eric expressed interest in using such tools for his educational institution
  • Some members were skeptical about the true level of data security provided
  • Alan suggested that open-source alternatives might be more secure and customizable

4. AI Tools Directory

  • Future Tools directory by Matt Wolfe showcases new AI tools
  • Searchable by category (e.g. 19 copywriting tools)
  • Good starting point, but personal recommendations still valuable
  • Site itself built using AI for tool discovery/descriptions

Opinion:

  • Group generally agreed on the usefulness of such directories
  • Some members emphasized the importance of personal recommendations alongside directories

5. Automated Wikipedia Article Generation

  • Stanford researchers developed STORM system to automate Wikipedia article prep
  • Uses multi-perspective questioning and expert simulation
  • Gathers sources, creates outlines automatically
  • Similar to Perplexity AI’s approach
  • Raises questions about future of human-curated Wikipedia

Opinion:

  • Discussions on potential impacts on content creation industries
  • Some members expressed concerns about potential for misinformation or bias in automated content
  • Group debated the future role of human writers and editors in light of such technologies

Leave a Reply