Jolla has officially announced the first version of its personal server-based AI assistant currently in development. The reborn startup is building a privacy-focused AI device, also known as Jolla Mind2. TechCrunch exclusively revealed him at his MWC in February.
Pre-orders also began during Monday's livestreamed launch event, with the first units expected to ship in Europe later this year. Global pre-orders will begin in June and are expected to ship late this year or early next year.
In the more than two months since we saw the first 3D printed prototype of Jolla's AI-in-a-box, there has been considerable buzz around other consumer AI devices such as Humane's Ai Pin and Rabbit R1. There was a lot of hype. But early interest dwindled in the face of a poor or unfinished user experience and a sense that early AI gadgets were more about experimentation and less about practicality.
The European startup behind Jolla Mind2 is working hard to ensure its AI devices don't fall into this trap, according to CEO and co-founder Antti Saarnio. That's why they tread carefully and try to avoid the pitfalls of over-promising and under-delivering.
“I believe that its integration into our software is one of the biggest disruptive moments for AI. It’s a massive disruption. But the initial approach was fundamentally hasty, and that’s where the problem lies. ,” he told TechCrunch. “We should introduce software that actually works.”
The feedback is harsh, but fair considering recent releases.
Sarnio said the team plans to ship the first few hundred units (up to 500) of devices this fall to early adopters in Europe, possibly to enthusiasts who built around early products such as the Sailfish mobile OS. It will take advantage of the community.
Jolla Mind2 will cost 699 euros (incl. VAT), making the hardware considerably more expensive than the team originally planned. However, it also has more onboard RAM (16 GB) and storage (1 TB) than originally budgeted. The not so good: Users have to pay a subscription starting from 9.99 euros per month. So this is an AI device that isn't cheap either.
AI agent living in a box
Jolla Mind2 includes a set of AI agents tailored for a variety of productivity-focused use cases. These are designed to integrate with relevant third-party services (via APIs) so they can perform a variety of functions, such as email agents that prioritize inboxes and compose and send messages. . Alternatively, Contact Agent, which Jolla briefly demoed at his MWC, acts as a repository of information about the people you interact with so you can stay on top of your professional network.
Image credit: Jolla
In a video call with TechCrunch ahead of Monday's official announcement, Saarnio demoed the latest version of Jolla Mind2, showing off some never-before-seen features, including the aforementioned email agent. Document preview and summary feature. E-signature functionality for documents. and a new thing it calls a “knowledge base” (more on that later).
The productivity-focused features seen in the demo worked, but there were some noticeable latency issues. Sarnio apologetically said that a demo gremlin had hit earlier in the day and caused a last-minute performance issue.
In the chatbot interface demo, switching between agents was also manual, but in the final product it will be automated with AI semantically understanding user queries, he said.
Planned AI agents include: Storage agent. Task management. Message agent (to integrate with third-party messaging apps). “Coach Agent” is intended to leverage third-party activity/health tracking apps and devices to allow users to query quantified health data on their devices.
The promise of private, on-device processing is the product's main selling point. Jolla claims that user queries and data are stored securely on hardware owned by the user. For example, if you use OpenAI's ChatGPT, instead of your personal information being siphoned off to the cloud for commercial data mining or profit opportunities for others…
Privacy sounds great, but obviously you need to minimize latency. This is doubly important given that in addition to the central focus of Jolla's strategy on personal data firewalls, it is also targeting “prosumer” use cases with productivity and convenience. .
The core pitch is that the device's on-board AI model of about 3BN parameters (Saarnio calls it a “small language model”) can be connected to all kinds of third-party data sources. is. This allows users' information to be used for further processing and scalable utilities without worrying about compromising the safety or integrity of the information when harnessing the power of AI.
For queries where Jolla Mind2's local AI model may not be sufficient, the system gives users the option to send the query “out of the world”, to a third-party large-scale language model (LLM). At the same time, make the user aware of the implications of doing so. You're sending data outside of a secure, private space. Jolla is experimenting with some form of color-coding in messages to indicate the level of data privacy that applies (for example, blue to represent complete security on the device, and data exposed to commercial AI). red if all privacy bets are off).
The performance Saanio has seen will be a top priority for the team working to improve the product. “It's basically an old rule that if you want to make a breakthrough, it has to be five times better than the existing solution,” he said.
Security should also be an absolute priority. The hardware does things like set up a private VPN connection so your mobile device or computer can securely communicate with your device. Sarnio added that the box will store an encrypted cloud-based backup of user data in case of hardware failure or loss.
Choosing which zero-knowledge encryption architecture to disable external access to data is an important consideration for privacy-conscious users. Those details are still being worked out.
AI hardware with a purpose?
One of the big criticisms leveled at early AI devices like Humane's Ai Pin and Rabbit R1 took the form of awkward questions. “Isn't this just an app?” As you know, everyone is already packing their smartphones.
This is clearly not an offensive line that fits Jolla Mind2. First, the box housing the AI is intended to be static, not movable. Keep it in a safe place in your home or office. Therefore, most of the time you won't be carrying around two pieces of hardware. In fact, a mobile (or desktop computer) is the usual tool for interacting with Jolla Mind2 via a chatbot-style conversational interface.
Image credit: Jolla
Another big argument Saarnio makes to justify Jolla Mind2 as a device is that trying to run a personal server-style approach to AI processing in the cloud is difficult or prohibitively expensive to scale.
“I think it would be very difficult to scale a cloud infrastructure if each user had to run a separate local LLM. Cloud services need to be running all the time. Restarts take 5 minutes. You can’t really use it that way because it can unravel,” he argued. “For example, you could use some kind of solution that you downloaded to your desktop and use it on your smartphone. Also, if you want to have a multi-device environment, I think the only option is to use a personal server like this.”
The aforementioned knowledge base is another type of AI agent feature that allows users to instruct their devices to connect to a curated repository of information, further extending its utility.
Mr. Sarnio demonstrated examples of selected information dumps on deforestation in Africa. Once the knowledge base is on the device, users can access and query it, extending the functionality of the model and giving users a deeper understanding of specific topics.
“user [could say] “Hey, I want to know about deforestation in Africa,” he explained. “Then the AI agent said he has one provider here.” [who has] We've created an external knowledge base about this. Would you like to connect to this? Then you can start chatting with this knowledge base. You can also ask them to create summaries, documents/reports, etc. on this.
“That's one of the big things we're thinking about: we need graded information on the internet,” he added. “So you can have thought leaders and professors in fields like climate science create a knowledge base and upload all the relevant research papers, and then users can… It gives you some kind of confidence that you’ve scored it.”
If Jolla can fly this, it could be pretty clever. Not only do LLMs fabricate information, they also tend to present made-up nonsense as if it were absolute truth. So how can web users in the growing landscape of AI-generated internet content ensure that the information they are exposed to is authentic?
The startup's answer to this rapidly expanding knowledge crisis is to enable users to point their own on-device AI models to their preferred sources of truth. This is a delightfully human-centered solution to the big AI truth problem. Small AI models and smartly curated data sources provide a greener type of GenAI tooling than those offered by Big AI, which has energy-intensive, compute- and data-intensive approaches. can also do.
Of course, Jolla needs to compile a useful knowledge base for this feature to work. These are intended to be curated and evaluated by users and the broader community behind the approach. Sarnio doesn't think that's a big question. Domain experts will be able to easily collate and share useful research repositories, he suggests.
Jolla Mind2 focuses on another issue: technology users' software experiences are generally far beyond their control. User interfaces are routinely designed to be intentionally distracting, attention-focused, or overtly manipulative. So another selling point of this product is that it helps people take back their agency from all the dark patterns, sludge, notifications and other things that are really annoying when it comes to the apps they have to use. You can have AI cut through the noise for you.
Saarnio said AI models will be able to filter third-party content. For example, a user can request that only her AI-related posts be shown from the X feed, and nothing else. That means on-demand superpowers to shape what you're digitally consuming and what you're not consuming.
“The whole idea is [is] “To create a peaceful digital working environment,” he added.
Given Jolla's long backstory as an alternative smartphone maker, Saarnio knows better than anyone how difficult it is to convince people to buy a novel device. So, naturally, the team also has plans for his B2B license.
He said this is where the startup sees the greatest potential to expand the adoption of its AI devices, with plans to sell “hundreds of thousands” or even millions of devices through partners. I am assuming that it is possible. He admits that sales for the Jolla community are unlikely to exceed tens of thousands at best, given the limited size of its loyal and loyal fan base.
The AI component of the product is being developed under a separate (new) entity called Venho AI. The company is not only responsible for the software brains that power Jolla Mind2, but also acts as a licensing supplier to other companies that want to offer their own branded versions of the personal server-cum-AI assistant concept.
Saarnio said the licensing of AI models is changing, given that carriers could once again miss out on the digital spoils as big tech companies pivot to embed generative AI into their platforms. It suggests that this could be one of your potential target customers.
But first things first. Jolla/Venho needs to ship a solid AI product.
“First we need to mature the software, test and build it with the community, and then once the summer is over, we'll start talking with reseller partners,” he added.