Amazon intends to relaunch Alexa as an artificially intelligent agent that can perform practical tasks, but to do so, the company’s technicians will have to find solutions to several problems that are currently preventing the system from being rebuilt.
Image source: aboutamazon.com
The company has been trying for two years to give new life to the voice assistant, which is used on 500 million devices around the world, by transplanting a new “brain” based on generative AI into it. This is hampered by several problems: AI hallucinations, that is, giving deliberately incorrect answers to requests; response speed, that is, the delay before receiving a response; and overall system reliability. Rohit Prasad, head of the strong AI (AGI) development team at Amazon, told the Financial Times. Today Alexa can only perform simple tasks, such as playing music or setting alarms, but the company intends to expand its capabilities into a full-fledged agent that acts as a personal concierge. Such a system, for example, will help the user with choosing a restaurant or adjust the lighting in the bedroom depending on sleep cycles.
Work on the project to relaunch Alexa began when OpenAI ChatGPT was released at the end of 2022. Microsoft, Google, and Meta✴ have all quickly incorporated generative AI into their computing platforms, improved the quality of their software services, and there are doubts that Amazon will deliver on its vision in time to compete with other tech giants. The voice assistant currently runs on tightly calibrated algorithms, but the company has encountered unexpected difficulties in implementing more powerful but less predictable large language models. Integrating generative AI into a service used by hundreds of millions of people around the world seems to be a project of unprecedented complexity for the company’s specialists, which is not limited to the mechanical imposition of a large language model on the Alexa service.
For the system to work as an agent, Alexa’s “brain” must be able to access several hundred programs and services. “Sometimes we underestimate how many services are integrated into Alexa, and there are a huge number of them. These applications receive billions of requests a week, so when you’re trying to make reliable actions fast, you need to be able to do it in a very cost-effective way,” said Rohit Prasad. Users expect Alexa to provide quick and accurate answers, which runs counter to the probabilistic nature of today’s generative AI, which is statistical software that predicts word sequences based on speech and language patterns.
Amazon is looking to recruit experts to shape the AI’s personality, voice, and diction so that it remains familiar to Alexa users. And something will have to be done about “hallucinations,” which, on the scale of Amazon’s work, risk occurring many times a day. Despite the fact that the company employs many strong specialists and has impressive financial resources, developers must confront “technical and bureaucratic problems”: poor-quality data markup in training arrays and outdated or missing documentation.
The original implementation of Alexa was built on technology acquired from the British startup Evi in 2012 – a question and answer machine that searches an array of facts; the answers could be a weather forecast or a song in the music library. The technology behind the voice assistant, which was already inflexible, was weighed down by a clunky and disorganized codebase and a highly distributed development team. The new Alexa uses a whole range of AI models: some are used to recognize and translate voice requests and generate responses, others perform a monitoring function, not allowing inappropriate responses and preventing the occurrence of “hallucinations”. This set includes large Amazon language models, including the recently introduced Nova, as well as a product from its partner Anthropic Claude.
The company’s developers now intend to integrate additional filters for children and test Alexa-integrated equipment such as smart lights and Ring doorbells. Developers of third-party products have not yet made predictions about when the updated Alexa will appear or how it will be able to work with third-party services. Amazon scared them by urging them to prepare for a radical update to their voice assistant, but so far everything has died down. And the company has no understanding of how Alexa will be monetized – scenarios with the introduction of a paid subscription or receiving a commission from the sale of goods and services are allowed.