sameer (privacy/acc)

If you’ve paid attention at all to the broader AI sphere, you will have seen general conversations have moved toward products like: Muse, Town, Instinct, and Hark which are finding PMF in something more personalized. Honestly, at this point it’s a race to build assistants that can essentially take what large models are capable of and integrate a personalized experience for users and enterprise. The value prop was pretty clear when Instinct raised around $250m at a $2.5b valuation last month which is now reported in talks to raise another $1b at a $10b valuation, which is mind blowing I do understand the excitement behind these agents because if you have something that can remember work, help you follow through on plans, and take care of mundane or even important business you’d otherwise spend hours doing, it's appealing. I don’t think enough people understand how precious time truly is and this technology can help reclaim it in our personal lives, along with it also having a large opportunity to streamline business processes. These can easily find information across company systems, coordinate work, and follow through on some routine tasks that could save time cumulatively across organizations. Which all sounds great and can help explain some of the interest, but you would also be smart to wonder how much of that value can reliably be delivered If you understand how AI functions, your curiosity will steer toward how this could actually work. All of these tasks require access to information in your personal life and information that businesses run on (including internal convos, sensitive records, financial docs, etc.). The more responsibility that we give these assistants (and AI in general), the more emphasis will be placed on what they have access to, control, and what they’re allowed to do with it. This access has power creep because people that find it useful will find more use cases for the technology. It can start as a simple connection or summary that leads to an automatic workflow that requires a bit more access and snowballs from there. Overtime, a comfortability inherently sets in without further understanding of the inner mechanisms There is a great TechCrunch article that talks about this type of concern, specifically with Instinct. If you do a bit of digging you’ll see that Instinct’s terms actually say that disconnecting doesn’t automatically delete the information it has already collected and that it requires a separate request. These tiny hurtles required to fully remove data is exactly what people don’t understand As these assistants gain more access and responsibility, mistakes can have extreme consequences. They can be manipulated through prompts to provide information even if it wasn’t supposed to or make mistakes because it misunderstood a prompt that leaks something irreversible. This technology has no morals, it is simply attempting to complete a task the best way possible with the access and information it has. There are some that are more comfortable running this technology locally, maybe even some businesses so that they have infrastructure they control. But, this takes time, resources, and knowledge on top of the responsibility of keeping it running and free of issues 24/7 All of this leads to why private AI and encrypted computing are so exciting, especially within this context. There is a huge opportunity to make this type of technology equally capable, but more than likely, even more capable while also protecting the information and access that it has. Think about how much more useful these products could be if you felt comfortable bringing your actual life, work, and sensitive information into them: finances, personal plans, internal convos, business documents, healthcare data, honestly there is no ceiling. This is what we are targeting with Blackthorn at @Arcium: private inference that protects information while the AI uses it. The goal is to let people and businesses use capable AI with sensitive information without having to run the infra themselves Simple, powerful, <encrypted>