Table of Contents
At the forefront of technological progress, enormous IT companies are engaged in fierce competition to develop the most advanced generative AI tools. These tools are useful not just for end consumers, but also for developers and organizations.
Google, a heavyweight in this space, has launched Gemini, a dual-purpose wonder that serves as both the name of their AI interlocutor and the massive language model that supports it.
The Trio of Gemini: Nano, Pro, and Ultra
Gemini’s pedigree is broad, with three separate breeds: nano, pro, and ultra. The small Nano, nimble and compact, is the only one of the trio that fits nicely within the limitations of top-tier Android hardware, such as the Pixel 8 Pro and Galaxy S24 series.
However, its siblings control the huge span of Google’s cloud empire. Nano’s stature, albeit little when compared to Pro and Ultra, is said to grow with a new feature.
Nano’s Niche: Compact Yet Capable
Nano’s specialty is the production and deconstruction of concise text blocks. It now brings to life a trio of AI wonders on the Pixel 8 Pro: it distills audio recordings into pithy summaries in the Pixel Recorder app, creates smart answers in WhatsApp via Gboard, and refines texts with wise recommendations in the Google Texts app. The Galaxy S24 series, too, basks in Nano’s radiance, with capabilities such as Magic Compose.
APIs: The Conduits of Innovation
The ease with which programs may use Gemini Nano via an API demonstrates its potential for expansion. According to Android Authority, Nano may soon become the progenitor of AI-powered article synopses. As of August, Google’s experimental Search Generative Experience suite may distill the substance of web pages into important points, and this functionality is available on every Android device via a simple option in the Google app.
On-Device Summaries: A Glimpse into the Future
While the present AI article summary flourishes on the cloud, there is a hint of a future in which devices equipped with Gemini Nano, such as the Pixel 8 Pro and Galaxy S24 series, would be able to perform this feature independently. To understand this scenario, one must first investigate the workings of Gemini Nano on Android.
Android’s AICore: The Custodian of Gemini Nano
Android’s AICore service, rather than burdening individual apps with Gemini Nano, takes on the role of model custodian. This technique reduces storage requirements while streamlining the model’s distribution and refreshing. Applications may use a range of APIs from Google’s AI Edge SDK to access Gemini Nano’s intelligence for on-device inference, including one that lets Gemini Nano be fine-tuned with a LoRA block for specialized workloads.
Securing the Future: AI Safety and IP
In the field of machine learning IP and AI safety, Google advocates for secure APIs to install its Gemini Nano model and LoRA blocks onto devices. These APIs are derived from Android’s Private Compute Services, an open-source program that allows for the cloud-based download of machine learning models, protecting the Android System Intelligence app from the web’s prying eyes.
Protected Download: The Guardian API
AICore’s Protected Download API protects the integrity of materials downloaded into devices, including a binary transparency record for verification. AICore uses this API to obtain the Gemini Nano model and its associated LoRA blocks. A new AICore client called “AI_CORE_CHROME_SUMMARIZATION_OUTPUT” was recently discovered, implying the existence of an upcoming LoRA block dedicated to AI article summaries.
The Inference: A Future Painted with AI Summaries
Though the specifics of this new client remain unknown, its name and API’s aim point to a future in which Gemini Nano may automatically synthesize article summaries on-device. If this guess is correct, we anxiously anticipate Google’s announcement, as Gemini Nano’s repertoire of on-device AI capabilities is now limited.
With this AI article summary function incorporated into the Google app, it is expected to be available on both the Pixel 8 Pro and the Galaxy S24 series.
