Apple Is Still Holding Back on Generative Artificial Intelligence: A Case Study on a Multimodal Language Model for a Social Market Application
Apple is still holding back generative artificial intelligence. According to a New York Times report today, the company is in preliminary talks with the search company for the inclusion of their artificial intelligence model in their phones.
“The fact that they’re doing this, it shows they have the ability to understand how to train and how to build these models,” says Ruslan Salakhutdinov, a professor at Carnegie Mellon who led AI research at Apple several years ago. “It requires a certain amount of expertise.”
MM1 is a multimodal large language model, or MLLM, meaning it is trained on images as well as text. This allows the model to respond to text prompts and also answer complex questions about particular images.
The Apple research paper illustrates what happened when MM1 was provided a photo of a sun-dappled restaurant table with a few beer bottles and also an image of the menu. When asked how much someone would expect to pay for “all the beer on the table,” the model correctly reads off the correct price and tallies up the cost.
The creator of the big language model technology that has been used to createChatGPT has worked to expand the technology to work with other kinds of data. In December, when the model that now powers its answer to chatgppt was launched, the company trumpeted that it was beginning an important new direction in artificial intelligence. Apple says that the next frontier is in foundation models with LLMs.
A bunch of researchers backed by the French government have put out what is thought to be the largest AI training dataset composed entirely of text in the public domain. Trained Fairly awarded its first certification for a large language model that was built without the use of copyrighted works, showing that technology can be built in a different way to the more controversial ones in the artificial intelligence industry.
EdNewton-Rex is CEO of Fairly Trained and he says there are no fundamental reasons someone wouldn’t train an LLM fairly. He started the nonprofit after quitting his role at the image generation startup because he wasn’t happy about its policy of scrapers.
Today, Fairly Trained announced it has certified its first large language model. A Chicago-based legal tech consultancy startup called 273 has developed a training dataset for an application called KL3M.
The decision to train KL3M in this way was made by the company’s clients who were risk-averse. She says they need to know that output isn’t based on bad data. Fair use is not being relied on by us. The clients were interested in using generative AI for tasks like summarizing legal documents and drafting contracts, but didn’t want to get dragged into lawsuits about intellectual property as OpenAI, Stability AI, and others have been.
Bommarito says that 273 Ventures hadn’t worked on a large language model before but decided to train one as an experiment. She says that they had a test to see if it was possible. The company has created its own training data set, the Kelvin Legal DataPack, which includes thousands of legal documents reviewed to comply with copyright law.
The KL3M model performed better than expected due to the fact that the data had been carefully scrutinized, something she attributes to it being small. She says having clean data could mean that you don’t have to make the model so big. Curating a dataset can help make a finished AI model specialized to the task its designed for. 273 Ventures is now offering spots on a waitlist to clients who want to purchase access to this data.
The dataset was built from sources like public domain newspapers digitized by the US Library of Congress and the National Library of France. Pierre-Carl Langlais, project coordinator for Common Corpus, calls it a “big enough corpus to train a state-of-the-art LLM.” The most capable model of Openai is thought to have been trained on trillions of dollars.