Jolla has officially unveiled its inaugural personal server-based AI assistant. The rebooted company is developing a privacy-centric AI device called the Jolla Mind2, which TechCrunch initially disclosed at MWC in February.
During a livestreamed event on Monday, the company commenced preorders for the device, with the first batch expected to be delivered in Europe later this year. Global preorders will become available in June, with deliveries slated for late this year or early next.
In the past several months since the debut of Jolla’s initial 3D-printed prototype, other consumer-focused AI devices like Humane’s Ai Pin and Rabbit R1 have generated a buzz. However, early excitement has waned due to subpar or incomplete user experiences, highlighting that these emerging AI gadgets often prioritize experimentation over functionality.
Antti Saarnio, CEO and co-founder, emphasizes that Jolla aims to steer clear of this pitfall. The company is proceeding cautiously to avoid overpromising and underdelivering. “This is a pivotal moment for AI integration into our software. It’s a massive disruption. The initial approaches were rushed, leading to issues,” he explained to TechCrunch. “Our focus is on introducing software that genuinely works.”
The feedback, while harsh, is warranted considering recent product launches.
Saarnio also mentioned that the initial production run would consist of up to 500 units for early adopters in Europe this fall, likely leveraging the community of enthusiasts from their earlier Sailfish mobile OS.
The Jolla Mind2 will be priced at €699 (including VAT), considerably higher than initially planned. This hike is due to increased on-board RAM (16GB) and storage (1TB). However, users will also need to pay a monthly subscription fee starting at €9.99, making this another premium-priced AI device.
AI agents living in a box
The Jolla Mind2 is equipped with a suite of AI agents tailored for specific productivity applications. These agents are engineered to seamlessly interface with third-party services through APIs, enabling them to perform a variety of tasks on your behalf. For instance, an email agent is capable of managing your inbox, composing, and dispatching emails. Additionally, there is a contacts agent—demonstrated briefly by Jolla at MWC—that serves as an information repository for individuals within your professional network, helping you maintain an up-to-date and organized contact list.
In a pre-launch video call with TechCrunch, Saarnio provided a demonstration of the latest iteration of Jolla Mind2. This showcased several new features such as an advanced email agent, a document preview and summarizing tool, an e-signature capability, and a novel feature termed “knowledge bases.”
While the productivity-centric features appeared functional during the demo, there were noticeable latency issues. Saarnio attributed these to earlier technical disruptions, which had resulted in some last-minute performance setbacks.
Currently, switching between agents in the chatbot interface is a manual process, but Saarnio assured that this would be automated in the final version, leveraging the AI’s semantic understanding of user queries.
Planned AI agents for the final product include: a calendar agent, a storage agent, a task management agent, a message agent for integration with third-party messaging apps, and a “coach agent” designed to interface with third-party health and activity tracking devices, enabling users to query their health data.
The principal selling point of Jolla Mind2 is its commitment to on-device, private data processing. Jolla guarantees that user queries and data will remain secure on the device’s hardware, unlike, for instance, OpenAI’s ChatGPT, which stores personal information in the cloud, potentially for commercial purposes.
Although the emphasis on privacy is commendable, reducing latency to a minimum is crucial, especially given the device’s intended productivity and convenience advantages for prosumer users. Jolla’s strategic focus remains on protecting user data while enhancing user productivity.
Jolla’s core value proposition is that its approximately 3-billion-parameter AI model, described as a “small language model” by Saarnio, can interface with various third-party data sources. This allows extensive utility and data processing without compromising data security.
For more complex queries where the local AI model might fall short, users will have the choice to send queries to third-party large language models (LLMs), with a clear indication that doing so will expose their data beyond their secure environment. Jolla is considering implementing a color-coding system to denote the privacy level of data transmission (e.g., blue for total on-device privacy and red for data exposure to external AI).
Saarnio emphasized that performance optimization is a key focus for the team, stating, “It’s basically the old rule that if you want to make a breakthrough it has to be five times better than existing solutions.”
Furthermore, strong security measures are paramount. The device will facilitate private VPN connections to ensure secure communication between the user’s mobile device or computer and Jolla Mind2. An encrypted cloud-based backup for user data stored on the device will also be available in case of hardware failure or loss.
Finalizing the zero-knowledge encryption architecture to ensure absolute data privacy will be critical for privacy-aware users. These details are still under consideration.
AI hardware with a purpose?
A prominent criticism directed towards early AI devices, such as Humane’s Ai Pin and the Rabbit R1, revolves around a pertinent question: Why can’t this functionality be just an app, considering that almost everyone already owns a smartphone?
This critique, however, doesn’t quite apply to the Jolla Mind2. Unlike its counterparts, the AI housed within the Jolla Mind2 is designed to be stationary rather than portable, intended for secure placement at home or in an office. Thus, users are not burdened with carrying multiple devices. Interaction with the Jolla Mind2 typically occurs through your mobile device or desktop computer, utilizing a chatbot-style conversational interface.
The other primary argument presented by Saarnio to support Jolla Mind2 as a viable device revolves around the complexities and high costs associated with scaling AI processing in the cloud. Saarnio posits that handling a local language model (LLM) for each individual user separately via cloud infrastructure would be not only challenging but also prohibitively expensive.
“The scalability of cloud infrastructure would become immensely difficult if you had to run a local LLMx for every user individually. This would necessitate an always-active cloud service, given that rebooting might take up to five minutes, thereby impeding practical use,” he explained. “Alternatively, you could download some form of solution to your desktop and operate it via your smartphone. For a multi-device environment, this type of personal server appears to be the sole feasible solution.”
Another noteworthy feature of the AI agent within Jolla Mind2 is a sophisticated knowledge base system. This enables users to instruct their device to connect with curated repositories of information, thereby enhancing its utility.
Saarnio showcased a curated dataset about deforestation in Africa as an example. Once this knowledge base is ingested by the device, users can query it, effectively augmenting the model’s ability to assist them in delving deeper into specific topics.
“A user might say, ‘I want to learn about African deforestation’,” he elaborated. “The AI agent would then suggest connecting to an external knowledge base crafted by a provider on this subject. You can interact with this knowledge base and even request summaries or reports.
“We believe in the need for graded information on the internet,” he added. “Imagine a climate science thought leader or professor creating a knowledge base and uploading all essential research papers. Users would then have a level of trust in this graded information.”
Should Jolla successfully implement this concept, it would represent a significant advancement. Large Language Models (LLMs) are notorious for generating fabricated or misleading information. In an era where internet content is increasingly AI-generated, discerning trustworthy information is paramount.
Jolla’s approach to this burgeoning problem is to allow users to direct their on-device AI model to their chosen sources of credible information. This user-centric remedy addresses “Big AI’s” reliability issues while also offering a greener technological option. By combining smaller AI models with intelligently curated data sources, this method is more energy-efficient, avoiding the energy-intensive practices employed by larger AI models.
However, for this feature to be effective, Jolla will need to ensure the compilation of useful knowledge bases. The vision is for these to be curated and rated by the users and the broader community that Jolla hopes to attract. Saarnio believes this is an attainable goal, as domain experts should be able to readily gather and share valuable research collections.
Jolla Mind2 also tackles another common issue: the lack of user control over their technological interactions. Conventional user interfaces are often designed to distract or manipulate users. Thus, part of Jolla Mind2’s appeal is its potential to help users regain control from disruptive elements like incessant notifications and distracting apps. Users can instruct the AI to filter out digital noise.
For example, the AI model can be tailored to show only AI-related posts from a user’s X feed, filtering out all other content. This feature empowers users to curate their digital intake more effectively.
“The goal is to create a peaceful digital working environment,” Saarnio explained.
Saarnio is well-versed in the challenges of persuading consumers to adopt new devices, given Jolla’s history as an alternative smartphone manufacturer. Consequently, the team plans a B2B licensing strategy, which Saarnio believes holds the greatest potential for scaling. By partnering with other companies, the goal is to sell “hundreds of thousands” or even millions of devices, far surpassing the limited scope of the Jolla enthusiast community.
The AI component is being developed under a new business entity, Venho AI, which is also intended to serve as a licensing provider for businesses interested in their personal server-cum-AI-assistant model.
Potential partners could include telecom companies, as they risk being sidelined in the digital landscape reshaped by tech giants integrating generative AI into their platforms.
However, Jolla/Venho’s priority remains clear: delivering a robust AI product.
“First, we need to mature the software, test it with the community, and then, after the summer, we’ll begin discussions with distribution partners,” Saarnio concluded.
