Beyond Tomorrow: Companies Train Models On YouTube Videos, OpenAI's New "Mini" Model — Weekly AI Roundup
Companies are training their AI on YouTube videos illegally, a U.S. senator speaks again and OpenAI's new mini model, here's a roundup of everything AI from the week gone by.
It's been a tumultuous week in tech with a major incident involving Crowdstrike causing widespread IT outages. A bug in a Crowdstrike update disrupted Microsoft operating systems, resulting in significant disruptions to global IT infrastructure throughout Friday.
Whether you're returning to the office or continuing to work from home, spare a thought for your IT team—they've earned a well-deserved break after navigating through the chaos. For a detailed analysis of Friday's events, you can find our comprehensive breakdown here.
From big tech's questionable practices in AI training resurface and a US politician finds her voice with AI assistance to OpenAI introduces its latest, more affordable model, here are the stories that made the news on artificial intelligence space this week.
AI Companies Used YouTube Videos To Train AI
Several leading AI companies, including Nvidia, Anthropic, and Apple, have reportedly leveraged subtitles from over 173,000 YouTube videos across 48,000 channels to train their models, according to a report by Proof News. This practice, which contravenes YouTube's Terms of Service, has sparked outcry from affected content creators, including major news outlets and popular YouTube channels.
Predictably, channel owners are upset, and they have every right to be.
Many creators invest considerable time and effort into crafting their content, only to find their work repurposed without permission for AI development. The dataset in question was compiled by EleutherAI, a non-profit AI lab focusing on model interpretability and alignment.
For further insights, read Proof News' full report here, available free of charge.
US Politician "Speaks" Once More
In a heartening development, Representative Jennifer Wexton of Virginia's 10th Congressional District has regained her voice through AI technology.
Diagnosed with Progressive Supranuclear Palsy, which severely impairs speech and mobility, Wexton now uses an AI-generated voice model based on her previous speeches.
This advancement represents a significant breakthrough in aiding individuals facing similar challenges, moving beyond conventional text-to-speech solutions.
OpenAI Releases GPT-4o mini
OpenAI has unveiled the GPT-4o mini, a more compact and cost-effective iteration of its GPT-4o model. Designed to enhance accessibility for developers and businesses alike, the mini model offers improved affordability and efficiency. Effective July 18, it has replaced GPT-3.5 for all Free, Plus, and Team users on ChatGPT, with enterprise access set to follow next week.
Priced at ¢15 per 1 million input tokens and ¢60 per 1 million output tokens, the GPT-4o mini represents a 60% cost reduction compared to its predecessor, GPT-3.5 Turbo. Featuring a context window of 128K tokens and supporting up to 16K output tokens per request, the model achieves an impressive 82% score on the MMLU benchmark, surpassing several competitors in textual intelligence and reasoning.
As AI continues to evolve, offering new possibilities and advancements, stay tuned for more updates and breakthroughs in the field.
Beyond Tomorrow is a weekly newsletter published every Saturday to give you a roundup of everything AI in the last week.
