Getting ChatGPT to operate autonomously within the confines of an operating system has proven to be a difficult task for numerous reasons, but a team composed of scientists from Microsoft Research and Peking University may have discovered the secret sauce. The team conducted a study to determine why large artificial intelligence (AI) language models like GPT-4 fail on tasks that require manipulation of an operating system. Next-generation systems like ChatGPT running on GPT-4 set the benchmark for generative tasks like composing an email or writing a poem. However, getting them to act as agents within a general environment presents a significant challenge.
Traditionally, AI models are trained to explore through reinforcement learning in a virtual environment. AI developers have used modified versions of popular video games like Super Mario Bros. and Minecraft to “teach” models concepts such as self-guided exploration and goal finding. However, operating systems are a completely different terrain for AI models. As agents, performing functions within an OS often presents itself as a multimodal challenge that requires the exchange of information between different components, programs, and applications.
Generally speaking, the reinforcement training approach requires trial and error. However, as anyone who has entered their password incorrectly too many times, or forgotten which shortcuts work in which applications knows, data can easily be lost when using such an approach in an OS environment. The researchers worked with several LLMs including open source Llama2 70B from Meta and GPT-3.5 and GPT-4 from OpenAI. According to research, none of them worked particularly well.
For researchers to find a way to overcome these challenges, they first had to understand why LLMs failed to manipulate operating systems when some AI models were capable of superhuman feats like beating everyone at chess and Go. The team developed a novel training environment called AndroidArena that allowed LLMs to explore an environment similar to the Android operating system. Then, after creating test tasks and a reference system, they identified the lack of four key capabilities as responsible: comprehension, reasoning, exploration and reflection.
Although the scope of work was specifically intended to cover problem identification, during the research process the team identified a “simple” method to increase the accuracy of a model by 27%. Essentially, the team provided the model with automated information related to the number of attempts it had previously made and what it had attempted during those attempts. This addressed the problem of lack of “reflection” by somehow embedding the memory within the stimuli used to activate it.
This research could be significant in the quest to build a better AI assistant.
Microsoft continues its investment streak in Europe by announcing a $2 billion investment in Spain to expand its cloud and AI infrastructure and build data centers.
Microsoft has announced another major investment in Europe with a $2.1 billion commitment to expand its artificial intelligence (AI) and cloud infrastructure in Spain. Brad Smith, vice president and president of the company, posted on the social media platform
Smith said it’s not just about building data centers, but about committing to helping build the country’s “security and development and digital transformation of its government, businesses and people.”
According to the president of Microsoft, the company has been investing in Spain for 37 years. In September 2021, the company announced a new R&D center for AI technologies based in Barcelona, Spain. At the time, Alberto Granados, president of Microsoft Spain, said that Microsoft’s investment was “proof” of Spain’s talent in the digital sector. This investment comes after another investment Microsoft made in Europe on February 15 with €3 billion earmarked to develop Germany’s AI ecosystem.
Similar to its plans in Spain, over the next two years Microsoft has committed to investing the funds to boost German AI infrastructure, build new data centers, and train people in developing AI skills. A broader trend is underway, with leading technology companies investing in Europe in AI development, as the European Union is about to launch its landmark EU AI legislation. On the same day that Microsoft announced its investment in Germany, France’s Ministry of Economy, Finance and Industrial and Digital Sovereignty and Google announced plans to establish a new hub dedicated to AI in Paris, which will house almost 300 researchers and engineers to support the country’s AI ambitions. Two days earlier, Google announced its “AI Opportunities Initiative for Europe,” through which it said it would invest 25 million euros ($26.9 million) in skills training for Europeans in the AI sector.