Leaked models are a recurring problem in machine learning. They can cause models to overfit training data and produce a higher performance when predicting on unseen data than they would have if the model had not overfitted.
This is a serious and widespread problem in AI that needs to be avoided. It occurs when information about the target variable leaks into the input during the model’s training process.
Facebook’s AI tool, LLaMA, has leaked online. This is the first time that a major tech firm’s AI model has been made public.
Llama, which stands for Large Language Model Meta AI, was developed by Facebook’s Meta division. The model is similar to OpenAI’s GPT-3, but smaller in size and less expensive to run.
While it has a small footprint, LLaMA still performs as well as the most powerful GPT-3 models. And thanks to its slimmed-down size, it can be run on cheaper systems like desktop computers.
DeepMind is Google’s artificial intelligence division, and one of its most successful. It’s developed algorithms that slash energy use in data centres and optimize phone batteries. It’s also helped Google build AI that can diagnose medical conditions, such as cancer.
DeepMind’s research is driven by a passion for humanity, according to its founders. Demis Hassabis and Shane Legg, who are both based in London, say the company’s mission is to improve lives through machine learning.
It’s a vision that has attracted hundreds of the world’s most talented experts. They work on subjects as varied as self-driving cars, sports analysis, and even medical diagnoses.
But as the company grows, a number of DeepMind employees are considering their options. They are frustrated that Google has taken control of their health-related research, rather than allowing it to operate independently. Some have even left the company.
Facebook’s AI is a powerhouse that has a vast range of capabilities that enable it to serve its users and advertisers better. Among the many features that it has built are proactive detection technologies that take down content that praises or supports terrorist groups and organized hate.
In addition, deep learning and AI experts are building systems that can analyze photos without relying on tags or surrounding text. This could help in analyzing photos for disaster relief, or to find people with specific disabilities that need more coverage.
This technology is currently under evaluation by Facebook, which uses it to help identify people at risk of suicide, according to a report from Motherboard. It also helps users on dating apps such as Tinder increase their matches by encouraging them to start conversations with strangers.
In the midst of this AI revolution, Facebook’s proprietary model LLaMA, which was only accessible to approved researchers and government officials, leaked last week on 4Chan. It is the first time a major tech company’s AI model has been made available to the public.
Google’s Shadow Model
Google relies on a vast shadow workforce of temporary workers, paid by staffing agencies. These contractors often work alongside full-time employees, performing tasks ranging from data analysis to security.
This workforce outnumbers Google’s direct employees and is a major source of resentment among its own workers as well as politicians and labor unions. Temps are denied benefits and often get paid less than their full-time counterparts.
As a result, they often are not included in company-wide meetings and events. They are also barred from looking at job postings or attending hiring fairs.
These workers also have a poor track record of communication, according to their employers. They were reportedly not given real-time updates during a shooting at YouTube headquarters in April, and their concerns about workplace security were not communicated to them.
Researchers showed how these attacks could be used on a range of machine learning models, including computer vision and language models. For instance, it was shown that an attacker could poison 64 sentences in a WikiText dataset to extract a six-digit number from a model with only about 230 guesses, resulting in 39 times fewer queries than if the model had been trained without such poisoning.