Welcome back to The Prompt.
On Monday, OpenAI said it fixed an error in ChatGPT where it appeared that the chatbot was messaging users without being prompted to do so. ChatGPT proactively asked a user “How was your first week at high school?” The person posted a bizarre exchange on Reddit, where they asked the AI chatbot “Did you message just me first?” The company claims that the error occurred because a previous message hadn’t been delivered correctly.
Also: Want to know about the most cutting-edge technologies before they become mainstream? Sign up for our new weekly newsletter The Prototype. The first issue drops this Friday!
Now let’s get into the headlines.
AI DEAL OF THE WEEK
Salesforce Ventures, the investment arm of the software giant, is earmarking another $500 million to generative AI companies, bringing the total funding that it has devoted to AI startups in the last 18 months to $1 billion, Forbes reported. Through the firm’s generative AI fund, it has backed leading startups in the space like Anthropic, Cohere, Mistral and Hugging Face.
Also notable: World Labs, a startup cofounded by Stanford scientist and “godmother of AI” Fei Fei Li, has raised $230 million from Andreessen Horowitz, NEA and Radical Ventures to build spatial intelligence through AI models that can understand and interact with objects and places in the physical world.
BIG PLAYS
OpenAI released a new family of large language models called OpenAI o1. They are designed to take more time than GPT models before they respond to queries so that they can hash out different ways to answer a question and provide better, more accurate responses. They are built to “reason” about complex problems, accepting and rejecting options as it works through a task.
Nikunj Handa, who works on OpenAI’s API team, tells me that the model’s intelligence is on par with an “extremely smart PhD” and can perform advanced tasks across legal, coding and scientific fields. It does that by using what’s called “chain of thought” reasoning, a step-by-step process to approach problems. “People go through great lengths to get an LLM to do the thing that they want it to do. And this model sort of does it more easily out of the box without needing to tell it ‘please don’t output things in that format’” he said.
These models are in the early stages and aren’t connected to the web or other external tools, and their access is limited to developers who have spent at least $1000 on the platform. There’s also safety concerns. During testing of earlier versions of the models, model evaluation company Apollo Research found that the models were capable of deceiving its users by appearing to be safe and helpful while actually not properly responding to queries.
DATA DILEMMAS
AI companies have been scraping data from Mumsnet, a UK-based parenting forum with an archive of more than six billion words, Wired reported. The platform has since tried to sign licensing deals with AI companies including OpenAI to sell its largely female-written content. But talks with OpenAI, which initially expressed interest in the website’s data, fell through after the AI behemoth said it strikes partnerships for large datasets that aren’t publicly available. Now, Mumsnet is taking legal action against the company and other scrapers for alleged copyright infringement.
DEEP DIVE
Black and white portraits of Rosa Parks, letters penned by Thomas Jefferson and The Giant Bible of Mainz, a 15th century manuscript known to be one of the last handwritten Bibles in Europe. These are among the 180 million items including books, manuscripts, maps and audio recordings housed within the Library of Congress.
Every year hundreds of thousands of visitors walk through the library’s high-ceiling pillared halls, passing beneath Renaissance-style domes, embellished with murals and mosaics. But of late, the more than 200-year-old library has attracted a new type of patron: AI companies that are eager to access the library’s digital archives — and the 185 petabytes of data stored within it — to develop and train their most advanced AI models.
“We know that we have a large amount of digital material that large language model companies are very interested in,” Judith Conklin, chief information officer at the Library of Congress (LOC) told Forbes. “It’s extraordinarily popular.”
But people who want access to the Library’s data must collect it via the API, a portal through which anyone, from a genealogist to an AI researcher, can download data. They are prohibited from scraping content directly from the site, a common practice among AI companies and one that Conklin said has become a real “hurdle” for the library because it slows public access to its archives.
“There are others who want our data to train their own models but they want it fast and so they just scrape our websites,” she said. “If they’re hurting the performance of our websites, we have to manually slow them down.”
The hunt for data is just one part of the story. Companies like OpenAI, Amazon and Microsoft are also courting the world’s largest library as a customer. They claim AI models can help librarians and subject matter specialists with tasks like navigating catalogs, searching records and summarizing long documents. This is certainly possible, but there are some rough edges that need to be ironed out first. Natalie Smith, the LOC’s director of digital strategy, told Forbes that AI models, trained on contemporary data, sometimes struggle with historical accuracy — identifying a person holding a book as someone holding a cell phone, for example. “There is an overwhelming bias towards today’s times and so they often apply modern concepts to historical documents,” Smith said.
Read the full story on Forbes.
WEEKLY DEMO
Billionaire venture capitalist Walter Kortschak, who has invested in tech companies like Twitter and Lyft for more than four decades, told Forbes that it’s dangerous to get swept up in AI’s “mass hysteria.” The investor, who has stakes in OpenAI and Anthropic, offered some advice, saying that it’s important to be selective, “invest in outlier founders” and discipline the number of investments you’re making.
QUIZ
OpenAI’s new o1 model’s was able to correctly answer questions from this difficult TV quiz show:
- Jeopardy
- The Floor
- Only Connect
- Wheel Of Fortune
Check if you got it right here.
MODEL BEHAVIOR
Snapchat has the right to use AI-generated images of users and their friends in advertisements, according to the terms of service for the social media company’s “My Selfies” tool, which lets people upload their photos to create AI-generated images, 404 Media reported. While the feature can be toggled off, it’s on by default. At least one person has come across an advertisement on Snapchat stories that featured an AI-generated image of themselves, the person posted on Reddit.
