Astra, the assistant who wants to surpass Siri and Alexa

On December 11, Google DeepMind lifted the veil on several innovations, including the highly anticipated Gemini 2.0 and major advances in the Astra project. If these products are not yet accessible to the general public, their potential, outlined during this presentation, could strengthen Google’s position as a major player in artificial intelligence.

Since the launch of Gemini 1.0 in December 2023, Google has demonstrated its desire to establish itself in the field of multimodal technologies. This first model, designed to simultaneously process data from various formats – text, image, video, sound or even code – represented an important step.

With Gemini 2.0, the company is taking a further step, by equipping its AI with enriched capabilities intended to create intelligent agents capable of interacting with users while carrying out various tasks for them (through a suitable smartphone in particular ). This new version also powers Astra, an assistant project unveiled last May, which aims to surpass the functionality of traditional voice assistants such as Siri or Alexa.

Gemini 2.0: the advent of Trillium TPUs

Gemini 2.0 stands out with significantly better performance than its predecessor, Gemini 1.5. Its processing speed is doubled, and its results on reference benchmarks such as MMLU-Pro (Massive Multitask Language Understanding – Professional) testify to its power, with an increase of several points on comprehension scores in various areas, such as as mathematics, philosophy or even psychology.

The efficiency of Gemini 2.0 relies on Trillium TPUs, the sixth generation of processors developed by Google. These hardware components optimize both model training and inference, enabling faster processing and increased efficiency.

For now, this version is only accessible to selected developers and testers. No public launch date has yet been communicated. However, Google plans to integrate Gemini 2.0 into several of its flagship products, starting with its search engine, where it will be used in the “AI Overviews” systems to summarize search results. A lighter version, called Gemini 2.0 Flash, is already available to the general public via the Gemini tool.

Astra: a universal assistant in the making

By fully exploiting the multimodal capabilities of Gemini 2.0, Astra facilitates natural and fluid interactions with the user. It answers questions, assists in the execution of various tasks and relies, if necessary, on Google services such as Search, Maps or Lens to enrich its answers.

One of Astra’s most notable features is its “optimized range” storage capability. In short, the assistant remembers past exchanges over several sessions, but in a relatively limited and well-thought-out period, which allows it to adapt its responses to the context and the specific needs of the user without exaggerating the references. This would therefore add to a long-term storage capacity, such as that of ChatGPT.

In a demo released by Google, the system shows impressive fluidity, as shown in the video below. However, criticisms remain. A journalist from MIT
Technology Reviewwho had the opportunity to test Astra, points out “notable shortcomings” and evokes a “chasm” between the promises of the demonstrations and reality.

Video presentation of the Astra project: