Today let’s talk about how we can use AI to speed up how we search for files. A tool we might use for this is what I’m thinking of as an AI File Explorer: an app for your computer or phone that analyzes your files with machine learnings models.

I can use it to tag, search, and organize my files based on their contents, not just their file names. The best part is that I wouldn’t have to do all the tagging manually, the AI would organize everything for me.

Thinking about this makes wonder if this is what the beginning of a really great personalized AI assistant will look like in the future. The kind of AI that will automate boring tasks and free us up to focus more on our interests and creativity.

Imagine a world where all your thoughts and ideas are neatly organized and searchable. The AI could cite facts from the sites you visit, PDFs you download, and podcasts you listen to. Think of it like Ctrl+F for your brain.

Or we can look at ChatGPT and it’s GPT plugin store as good examples — they’re developing specialized bots (ex: Laundry Buddy Scholar AI) that are not groundbreaking applications of Large Language Models (LLMs), but they are setting the stage for the future. For example: what if I want Scholar AI to summarize my pdfs, presentations, and spreadsheets? That’s cool and I do want that, but I’m not sure that’s data I want to share with OpenAI (or any company).

In another case you might have terabytes of images and videos where it doesn’t make sense to send all your data to OpenAI. The data is too large or it’s too expensive to analyze all of it. So what options do you have?

To answer that lets ask different question: where is your personal data stored? If it’s on the cloud then we’ll need to keep digging into what that means for using your data. You can’t use AI on your files if it can’t easily access them.

Not your disk, not your data

Most of us have data stored in a few places: cloud storage provider (Google Drive, iCloud, Dropbox, etc.), folders on your computer, and maybe an external drive. If we zoom in on the cloud storage industry for a second we see that the storage business is booming. The industry itself was expected grow over $100B/yr last year..

This matters for 2 reasons:

1. Storage providers aren’t incentivized to help users manage their data locally, because they make money by storing the data online.

2. There are risks associated with trusting providers to protect your data. It’s probably why businesses are also exploring cloud repatriation for their data.

Data vibe shift

People are starting to change their attitudes about digital media. Powerful directors like Guillermo Del Toro and Christopher Nolan are encouraging people to buy physical media. They’re directly criticizing streaming services for pulling films from their platforms.

So what does this all have to do with personalized AI?

To start, if you don’t physically store your own data on a computer or hard drive, then you will depend on cloud providers for access to your data. In other words, companies with large amounts of data have the power to train powerful models on that data and at the same time might try to prevent you from doing the same. If the current showdown between Beeper, a startup that reverse-engineered iMessage to make it compatible with Android, is any indication, the platforms hosting your data are not down for the cause..

Next-Generation File Explorers

A year after Chat GPT’s release, AI is entering its Multimodal Era. Multimodal models are exciting because they can “understand” text, audio, and video. This is creating a new industry dedicated to making sense of people’s data, including videos and photos. AI File Explorers are a subset of this industry. This niche is focused on selling the AI dream to end users, not businesses. A few companies in this bucket are Raycast, Shade, and Cosmos. I am a Co-Founder of Cosmos and can say one challenge we’re facing as an industry is convincing users it is worth the effort to download an app to bring AI to their data.

One potential advantage for startups is that Apple and Microsoft’s default file browsers are pretty meh. They don’t inspire me and they’re not made for creators. Windows Copilot is trying to change that, but so far the reviews are underwhelming.

Looking ahead soon it we’ll be able to our search images by mood, color palettes, or other custom labels. AI will do all the boring work of tagging each file without any human input. It only makes sense that our file explorers will be the brains behind this operation.

Private by Default

We also trust and expect our file explorer activity to be private by default. I see an opportunity for consumer AI companies like Cosmos to help users run smaller, specialized models on their own computers, cutting out the cloud completely.

A privacy-forward product could gives users the AI superpowers they want without needing to compromise their or their workplace’s privacy. I’m confident there’s a strong, sustainable business to be built in offering users a private, secure way to apply AI to their personal files. Users shouldn’t have to worry if the products they use are training AI models on their private photos and documents.

Small Language Models

Another trend to watch in AI is the need to manage compute costs. It’s the reason Microsoft, OpenAI’s own compute-provider, is investing in Small Language Models (SLMs) — low-cost, high-performing models that beat larger models on certain tasks. File Explorers are the natural testing lab for SLMs and could give users more freedom to experiment with different language models, especially those coming from open-source community.

There could be a marketplace where users test open and closed sourced models to best fit their needs while keeping their data private. An ecosystem like this would keep a wall of separation between users’ data and model developers.

Use cases

This all sounds interesting in the abstract, but what are some real-life use cases for an AI File Explorer? Below are a few worth exploring.

Similarity search

Using a reference image or text, find all the documents similar to the original. This could be useful for identifying blurry photos, finding all photos with “warm” colors, or all the references “deadlines” in your PDF documents.

Search by context

Search the content of your files, not just their names. If you know there is a homeowner’s insurance contract somewhere in your files, but you don’t remember what it’s called, what good is that?

Searching file contents like “show me pictures of my friend David” or “what was the start date for my last job?” would make this information useful and actionable.

Auto-tagging

Tagging features let you organize your files with custom labels (sports, notes, headshots, contracts, etc.). It would make it easy to organize massive file collections on your computer and external drives. But it would also remove the need for most manual labeling, saving people and businesses valuable time.

Prepping for the Future

It’s been an exciting year for AI. The pace of development is head-spinning it’s hard to imagine where we’ll be in next 5 years. It’s likely AI will be deeply ingrained in our day-to-day lives by then. I’m looking forward to documenting the journey.