2 Paths to Privacy in AI
Homomorphic encryption and Bring Your Own Model is the privacy future we need in AI
I have really been deep down the rabbit hole of AI for the past couple months. That’s what sparked this newsletter. I’m learning so much about this space and it’s extremely exciting.
That said, there is big part of this I’m really not liking. The fact that we have nearly no privacy of our data and we are sort of forced to use a handful of companies AI models. So why is this the case?
There is a race going on right now to be the first company to AGI. Many companies big and small have their own takes on what we need to do in order to get to this future. Unfortunately for you and I this means that if we want to use this technology we have to trust these companies with an awful lot. I’m not a very trusting person (I’m working on it). That said, I really do not think a handful of companies should be in charge of all the worlds AI. I don’t care who is running those companies.
We need government regulation in this space but let’s be honest. The odds of that happening are slim and the odds of them getting it right or understanding it at all is even smaller, near zero.
Honestly, my main concern is the privacy of my data. I know I'm probably already an open book to many companies—hi, Google and Apple—but with AI and its capabilities, I find myself wanting to use it for even more parts of my life.
Also if you are a business that is just blindly trusting these companies with your data. That feels like a real mistake.
Again, I’m new to this AI world so I’m very open to new information. From what I understand about companies and models like OpenAI or Anthropic, any time you ask them questions or interact with their chatbots, none of that data is encrypted or secured in any way. So when their model is figuring out the answer to your question its looking at all of your data as clear as day. Everyone is just out here raw dogging it with their data blindly trusting Sam Altman and OpenAI. I can’t get over this.
So, is there any hope for individuals like myself who care about privacy and not being forced into using one of three models from these companies? I think we got two paths.
Homomorphic encryption
I’m not gonna sit here and pretend to understand this encryption, at least not yet. I’m hoping to do a deep dive on this in a future post but for now the way I understand it is this type of encryption allows for models to act on your encrypted data. There is a lot of math and encryption and algorithms. This is what we really need. This kind of gives us the best of both worlds. The best models acting on encrypted data. Win.
I’m honestly not sure the current state of things regarding HE though. It’s a little suspicious as to why OpenAI or other companies are not openly trying to make more progress here. I searched “openai homomorphic encryption” and couldn’t find any result directly from them.
That said, Apple just released how they are doing HE to power certain experiences on device. They also released a swift library for HE called swift-homomorphic-encryption. Google released a post last year discussing FHE and they released a compiler as well. The compiler hasn’t seen much activity as of recent but something exists.
Bring Your Own Models
Another solution I’m a big fan of is Bring Your Own Model. Companies and individuals should invest in creating their own models. The barrier to do this is getting lower and lower. That’s great!
Personally, I’ve been using Ollama and LMStudio for awhile now to play with all of these open source models that exist now. More and more companies such as Meta and Google are open sourcing their models to the world for us all to use.
You have websites like HuggingFace that is basically a giant hub for all of these models. It’s amazing seeing just how much is available for individuals or companies to use for free.
The only real bottleneck here is compute. I’ve run a decent number of models on my computer but they are slow. You really need something beefier than your standard Macbook. Unfortunately though, compute is expensive and that doesn’t seem to be changing anytime soon.
I’m hopeful though that new and creative solutions can be built to allow for individuals to be able to create their own models and bring those wherever they need them.
AI products should start to be designed in a way that promotes this type of openness and ability for users to plug in their own models where they can. This is especially true if you are a business. As a business you want to build up your own Vector DBs and models. You should be able to take those and use those in the best AI tools out there. That only works if those companies that are building those tools provide a way for you to do so. I think they should.
Closing thoughts
The good news is I like both of these options and I think both are going to happen. That’s good for everybody. These big companies that hold the biggest and best models really need to make security and privacy a priority. Either that or open source the models. Also don’t be afraid to invest in building your own models. It’s hard right now but it’s getting easier. Hopefully we get a lot more innovation here for individuals.
Thank you for reading this post! If you enjoyed it, feel free to follow me below and visit my website for more content.