AI: Weekly Summary. RTZ #318

Welcome to the Saturday weekly AI summary here on ‘AI: Reset to Zero’. 

1. Google calls out OpenAI for training Sora on YouTube: Google YouTube CEO Neil Mohan underlined that OpenAI training its next gen text-to-video Sora service on YouTube videos, would be breaking YouTube’s terms of service if trained on YouTube videos. Sora of course is OpenAI’s latest LLM AI innovation, making ripples in Hollywood and beyond. Google of course can train Gemini on YouTube data since it’s in their Terms of Service already. This is another skirmish around Video, the next major source of Data to train next generation LLM AI models, as outlined some months ago. Raises questions around TikTok/Bytedance video content, and of course Meta’s Reels video content in Instagram and Facebook. The ‘Fair Use’ debate around copyrights now extended to broader data rights to train LLMs.It also highlights the broader industry push around larger sources of data for next gen LLM AI models. Broader take on the AI Data Quest here.

2. Amazon retrenching on AI-driven Self Checkout Stores: Amazon is dropping the large store implementation of its “Just Walk Out Tech”, after years of investment around AI-augmented checkout technologies. Axios provided broader context of humans needed to scale these technologies. This issue remains a broader reality for AI companies like Scale AI and others, especially in the critical Data labeling, prep and fine-tuning part of the industry. Broader issues around ‘humans in the loop’ for AI data discussed here.

3. Anthropic highlights ‘Guardrail challenges’ on next-gen models: Anthropic’s latest AI Research highlights ‘Many-shot jailbreaking’ technique that “can be used to evade the safety guardrails put in place by LLMs. Techcrunch puts the technique in broader context for the industry, highlighting how difficult it continues to be to corral next generation LLMs from possible abuse. Deeper dive here.

4. OpenAI offers ChatGPT access without sign-up: OpenAI’s latest move to offer ChatGPT via GPT 3.5 for now without sign-up highlights its efforts to widen the mainstream access to its services. Search Engine Journal highlights the caveats with the service, but the move is notable for its directional effort to simplify access for OpenAI’s other models. Echoes moves to simplify mainstream access by Google, Amazon and others decades ago.

5. Google DeepMind’s Demis Hassabis warns on AI Hype: Deepmind founder and Google AI head Sir Demis Hassabis highlighted the underlying scientific progress in AI relative to the broader investment rush into AI at large. The comments underlines Google’s efforts to both commercialize and leverage AI for research via efforts like AlphaFold, and other efforts. And provides a pacing contrast with the industry at large.

Other AI Readings for weekend:

1. New Machine Learning, AI & Data (MAD) landscape 2024 report by VC FirstMark highlights the rapid changes in the AI ecosystem, with hundreds of new entrants, going from 1416 last year to 2011 thus far in 2024.

2. AI Training Software company Replit highlighted new initiatives like Code Repair and Replit Teams at its Developer Day. Continued differentiation with Microsoft’s CoPilot for Github and others. Company is focused on over 30 million developers worldwide, increasingly focused on AI products and services.

Thanks for joining this Saturday with your beverage of choice. 

Up next, the Sunday ‘The Bigger Picture’ tomorrow. Stay tuned. 

(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here)