AI Roundup 024: The Llama King

July 21, 2023

Artwork created with Midjourney

The Llama King

Perhaps the biggest news this week was Meta's release of Llama 2, its latest and greatest large language model.

Why it matters:

• OpenAI, Anthropic, and Google Deepmind own the best AI models. Many businesses want to use those models in their products but worry about long-term ownership and data privacy.

• Llama V1 was a great open-source model, but it wasn't licensed for commercial use, so using it was a legal gray area.

• Llama V2 is a better model, trained for ChatGPT-style use cases, AND is commercially available. So it's a huge win for startups that want to build products on top of it.

For a deeper look at Llama 2 and its potential impact, take a look at Wednesday's post:

Elsewhere in the FAANG free-for-all:

• Apple is working on a framework for LLMs and a chatbot that's been internally dubbed "Apple GPT."

• Microsoft's upcoming Copilot for Microsoft 365 will cost $30 per user per month. The company also showed off Bing Chat Enterprise, a privacy-focused version of the chatbot.

• And Google published details on its AI Red Team, an internal "team of hackers that simulate a variety of adversaries."

Artwork created with Midjourney.

No biz like showbiz

Following in the footsteps of the Writers Guild and Screen Actors Guild, the Authors Guild released a letter to AI leaders. In it, they ask for consent and compensation when using copyrighted materials to train AI models.

Where we're headed:

• The AI pushback from white-collar workers is going to continue. Fields with strong labor organizations will likely ask for restrictions or recompense regarding AI training data and output.

• With good reason - Google is already pitching an AI news article writing tool to major publications, marketed as an assistant rather than a replacement.

• And G/O Media, the owner of Gizmodo and Jezebel, is moving forward with AI-generated stories despite employee pushback and a botched launch.

• Meanwhile, OpenAI has partnered with the American Journalism Project to use its content as training data.

Elsewhere in AI anxiety:

• The White House announced voluntary pledges from leading AI companies, including cybersecurity and watermarking systems.

• Researchers have identified over 200,000 OpenAI credentials for sale on the dark web. Hackers are reportedly showing an increased interest in generative AI tools.

• And the SEC is worried that chatbots could fuel a market panic.

End of an era

This year, one major story has been the realization that AI companies are profiting from the Internet's publicly available content. It's led to lawsuits from artists and programmers and barriers from user-generated sites like Reddit and StackOverflow. It feels like the end of the era, as the incentives for publishing content for free are changing.

Why it matters:

• Every creator and community that resides online will need to grapple with AI content - from indie artists to admins of Wikipedia.

• And photos, articles, and voiceovers are just the beginning - creators of more complex media will have to deal with this problem too. This week, we saw tools that can generate entire websites and episodes of South Park.

• We’re also facing a flood of new AI-generated content. Microsoft, OpenAI, and others are testing “synthetic data,” as developers find generic internet data is no longer good enough for training advanced LLMs.

Source: Fable Studio

Things happen

ChatGPT gets customizable instructions. There’s no magic formula to reliably distinguish AI-written text. “85% of AI startups will be out of business in 3 years.” Despite generative AI mania, other startup funding remains bleak. Hands-on with Baidu’s Ernie, a leading Chinese ChatGPT competitor. NPR’s internal guidelines on using generative AI. The fierce debate over whether GPT-4 has gotten worse. How to use AI to do stuff. The Artificial Inventor project.

Last week’s roundup