Truly personal AI with Notekeeper

Hi, hope you are enjoying Notekeeper. It is the first app in the JewelBox family and beautifully shows some of the key principles we set out when building JewelBox. Notekeeper is born out of the opportunity all of us have with AI (Large Language Model) technology while adhering to the principles we have committed to in JewelBox.

Large Language Models are incredible

ChatGPT clearly showed how compelling the Large Language Models (LLMs) are as a technology. For the first time, we had a confluence of three very powerful capabilities in a single product:

• interact with a machine using natural human language
• access the world's vast knowledge
• summarize, elaborate on, and scrutinize aspects of that knowledge on command

At JewelBox, we believe that this ease of use is remarkably empowering for people who are trying to understand and process knowledge or information.

Economics of AI and User Profile-Based Ad Targeting

Even though LLM technology is compelling, we believe that AI chatbots are moving in the same direction as internet search and digital social media in the long term. As we laid out in depth in our motivation for creating JewelBox, the biggest business opportunity open to AI chatbots is to use the highly precise intent in chatbot interactions to create a user profile and target ads to users. Thus, AI chatbots are prone to the same issues that plague internet search and digital media companies:

• Biased and propaganda-filled content chasing users, instead of users discovering unbiased and high-quality content by themselves. Internet search companies and digital media companies are either incapable of, or deliberately don't remove such propaganda from their recommendations.
• The user has no agency over their profile - at least not with the same ease of use as AI or search queries. They cannot influence the level of bias in the results and these products don't help them understand how universally applicable or biased these results are.

This leads to problems like not being able to trust search results or digital media for providing the whole truth, and not being able to understand how this content is contextualized to one's personal situation.

The Opportunity with AI

With AI chatbots, we expect the user profile as well as the user intent to be significantly more precise. This is because users get a lot of value from AI chatbots with deep, intent-based questions and follow-ups. While this could potentially be great for consumers, at JewelBox, we see the incentives aligning towards intensifying propaganda, a deep filter bubble and lack of trust with AI-generated content similar to current digital media.
We should note that the way AI chatbots are built today, they do seem to give us universally applicable and well-contextualized responses. AI chatbots are, most likely, striving to get more usage over search and digital media. However, with time, as the incentives with ad targeting play out, we expect AI chatbots to slowly shift fully towards biased content with deep user profiling.

There is a silver lining

Unlike internet search and digital media feeds, AI is being built by a large number of firms, sometimes out in the open with open source code and open training data. Even though some tech firms have higher performance proprietary AI models, open models are continuing to catch up fast. Tools and knowledge about Large Language Models are well distributed across people and there is no reason to believe that a single tech firm or a small group of people will monopolize this technology in the long run.

Perhaps, more importantly, AI technology is shaping to be a provider of intelligence rather than just knowledge. AI is, thus, being built as a service for technologists and technology businesses who want to incorporate intelligence within their workflows. The market for intelligence is a significant portion of the information-worker's tasks which is measured in highly valuable time of the same workers. This market is possibly much larger than the digital ads market.

Thus, unlike the past decade with closed search indices and closed digital media silos, AI is available to us via competitively priced APIs which we can use directly to power internet-search-like experiences for people. Remarkably, even small firms can build fully featured, AI-based, internet-search-like experiences with a pay-as-you-scale model. Notekeeper is taking advantage of these enablers to build a compelling AI search + intelligence product with minimal capital investment before scaling up.

Notekeeper: Our take on AI and personalized search

Unlike the biased and propaganda-ridden search results and digital media, we would like to provide users with an alternate option - a fully transparent AI chatbot, where there is no implicit user profiling. Instead, the product encourages curating and providing deep and personal context to state-of-the-art AI models explicitly.

To make it easy for users to curate their personal context, we have used the familiar metaphor of notes as the building block of personal context. Notekeeper enables users to build a web of linked notes which encodes the user's personal context. Given our privacy-forward technology and commitment, we expect that users will feel comfortable curating their most private information within Notekeeper.

Notekeeper forwards this personal context to state-of-the-art AI models when (and only when) the user engages with the AI Assistant. Thus, Notekeeper's AI Assistant has:

• No implicit profile: The explicitly visible web of notes is the only personal context sent to AI models. This makes AI assistant responses fully transparent.
• Agency on context: The user can switch to different notes to get AI responses for a different web of linked notes. Thus, the user has full agency to choose which context to use for their AI assistant queries.
• Privacy safeguards for personal context:
• This personal context is stored with end-to-end encryption in JewelBox.
• While Notekeeper has to forward this context unencrypted to AI providers, it:
• never stores it unencrypted.
• configures AI providers to never store it within their service.
• uses only AI providers who contractually don't use this data for AI training / tuning.
• High-quality AI models: We use state-of-the-art AI models that are tuned to provide an unbiased and trustworthy source of knowledge and intelligence.

A note on bias and propaganda

Note that while Notekeeper integrates with high-quality AI models, these models are trained using content from the internet. The responses are augmented using internet search. This base corpus of the internet can be biased. The search responses used to generate the final AI response can also have propaganda. AI labs often employ people to train their models with additional content that they generate which could have their own biases.

However, what Notekeeper guarantees is that it doesn't introduce additional personal bias through implicit user profiling in the responses.

Platform Agnostic and Low Cost

Notekeeper is built as a web-app so that it can be easily used across several major or minor platforms. It provides its strong privacy protection despite being a web-app. With support from JewelBox, Notekeeper has strong encryption, key management and low storage costs for notes. To control AI costs, Notekeeper gives users a range of AI models to select for their responses. Often, the 'mini' models are good enough for even heavy analysis tasks (like analysing food journals, or credit card transaction data).

Perhaps, most importantly, Notekeeper is not provided as an expensive subscription service. It is provided as an app with pay-per-use for the resources used. Thus, if you use it sparingly, it is not going to drain your money like other subscription services. We expect typical usage to be low cost - enough to make this a viable alternative for people who really want a privacy-conscious option for AI chatbots.

Our goal is to make Notekeeper your app of choice for not just keeping private notes, but also effectively using them as context when engaging with AI. If you were ever hesitant to use regular AI chatbots with truly private information, or wanted certainty about removing any profiled bias from AI responses, we think we have the best product for you.

- with ❤️, ritesh