The week in AI: OpenAI's DevDay & xAI's first language model - Grok

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools or services, and exciting projects in open source. Even if you aren’t an engineer, we’ll keep you in touch with what’s going on in AI.

• ‘GPTs’ were announced - custom chatbots within ChatGPT that fulfill certain roles better than the default ChatGPT (for example, a math tutor chatbot might be more thorough/adept at explaining math problems than the default model). Users can create their own just by prompting the GPT builder with natural language, and the custom instructions will be set up automatically. It’ll name your bot, create a profile picture for it and even suggest default questions to display for the user. Custom GPTs can utilize browsing with Bing, DALL-E 3, code interpreter (re-renamed from Advanced Data Analysis), etc. - which should make for some very creative use cases.

• These custom bots will be part of the GPT Store, which will be launching in just a few weeks. CEO Sam Altman says OpenAI is going to “pay people who build the most useful and the most used GPTs” a portion of the company’s revenue. Considering how easy it was from what we saw to create an effective GPT bot in just a few minutes, that could be a pretty lucrative side-hustle for some forward thinking individuals.

• The GPT-4 Turbo API is coming - and not only is it 2-3x faster and cheaper than the current GPT-4 API, it has a 128k token context window (upgraded from a max of 32k). This allows you to add up to about 300 pages of information into your prompt - and the model will keep context much better than if you had simply uploaded the same information in a file. It’s trained up to April 2023.

• OpenAI API’s are going multimodal: DALL-E 3, GPT-4 with Vision, Text-to-Speech HD, and Whisper for text transcription (coming later this month).

The updates are being rolled out over the next weeks; some users are gaining access to select features already.

Elon Musk’s 4-month old company xAI just announced their first large language model, Grok, which will soon be available in beta - for verified X/Twitter users only. Based on self-reported results for benchmarking, the model seems very capable for its parameter size, but it’s definitely not state of the art by any possible metric. According to the post, “A unique and fundamental advantage of Grok is that it has real-time knowledge of the world via the 𝕏 platform. It will also answer spicy questions that are rejected by most other AI systems.” Futurism has published a highly critical piece on Grok/Musk.

If you’re a verified X/Twitter user, you can join the waitlist for Grok.

Cruise has since officially recalled 950 vehicles from San Francisco streets, with more recalls likely to come. They’ve halted several operations and Cruise fleet vehicle production, and are in search of a new Chief Safety Officer. Shockingly, the article notes that remote human operators have been needed to step in and assist the company’s ‘autonomous’ vehicles with issues once every 2.5 to 5 miles. Cruise lost more than $700m in the third quarter of 2023.

AI-guru Ethan Mollick already has a great blog post up detailing his early experimentation with the aforementioned ‘GPTs’ - or easily customized multi-modal chatbots for ChatGPT. He goes through the process of using GPT Builder and troubleshooting a GPT, including some multi-modal use cases and some words of warning/wisdom about current limitations (including hallucinations). These GPTs will be useful tools for just about anyone - it’s easy to imagine schools, companies, government agencies, etc. building libraries of GPTs for specialized task assistance.

Hundreds of documents obtained by 404 Media show how Fusus, an AI-powered surveillance platform, is becoming widespread in American towns and cities. Fusus integrates footage from public and private cameras, including doorbell cameras, drones, and police body cams, into a single network accessible to law enforcement. The adoption of Fusus has raised privacy and surveillance concerns, especially due to the lack of transparency and community engagement in its implementation.

• Fusus can add AI to existing cameras, allowing them to scan for specific clothing or objects, effectively creating a blanket of surveillance without direct public oversight.

• There is ambiguity regarding Fusus' facial recognition capabilities, with conflicting information about whether or not it includes this feature, despite claims by Fusus to the contrary.

• The system turns ordinary cameras into automatic license plate readers, collecting data on vehicles' locations and times, potentially tracking citizens' movements.

• Nearly 150 jurisdictions were reported to be Fusus customers, with the system connected to over 33,000 cameras across 2,400 U.S. locations.

More in AI this week:

• AI companies have all kinds of arguments against paying for copyrighted content

• Biden’s elusive AI whisperer finally goes on the record - here’s his warning

• Google researchers deal a major blow to the theory AI is about to outsmart humans

• Five takeaways from UK’s AI safety summit at Bletchley Park

• (Microsoft) After OpenAI DevDay, what’s next for the Microsoft/OpenAI partnership

• (Anthropic) Thoughts on the US Executive Order, G7 Code of Conduct, and Bletchley Park Summit 

• Xbox partners with Inworld AI to build generative AI tools for game development

• Google’s AI chief says there’s a 50% chance we reach AGI in 5 years

• Generative AI Bible: The ultimate guide to genAI disruption

We’ll see you next Thursday - thanks for reading!