There’s a surprising amount of post-excitement fear, uncertainty and doubt being directed at DeepSeek. I’ve seen posts that say that open sourcing the model creates a security hole, posts that question the sources of DeepSeek’s training data (I won’t cover this here, but this is absolutely a concern), posts that point out DeepSeek’s T&C’s give them perpetual usage of your prompts and personal data.

Are these fears valid? Except for the open source one (I can’t believe we still worry about open source being less secure, haven’t heard this one for years), yeah, they’re valid fears. The issue is, you shouldn’t be worried about just DeepSeek, you should ask all of these same questions of any hosted LLM service. You should ask the same questions of any system you don’t have full control over. If you’re not paying for the product, you are the product. Well…sort of. I still strongly believe in open source, which typically allows you to run something “for free” on the hardware of your choice. The difference between open source and something that’s vaguely “free” is that real open source publishes a license which describes the terms of use of the project.

“But, DeepSeek is open source?” Yes, parts of DeepSeek are open source. In the ongoing discussion around LLMs and privacy we tend to talk about an LLM as the combination of the actual LLM “brain” (model) and the runtime of the LLM. DeepSeek has 16 open source repositories on Github and 13 collections on huggingface. They also published papers on their methodology which others can learn from. All of these things are positive and follow a pattern set by Meta when they open sourced their Llama model. However, just like Meta, there’s a difference between open sourcing a model and the runtime you choose to interact with that model.

If you have enough hardware, you can run Meta’s Llama model or DeepSeek’s R1 model yourself using a runtime like Ollama or LM Studio. Now you’re sharing your information with Ollama or LM Studio, so you need to ask yourself if they can be trusted with whatever data you’re giving them but (unless they’re doing something truly nefarious) none of your data is being passed back to the model authors.

This split between the LLM and the runtime that allows you to interact with the LLM is going to be a theme in future articles. While I’m completely blown away by the rapidly advancing abilities of LLMs, I’m more interested in the runtimes we’re wrapping around our LLM “brains”. AI agents and chatbots will give you different results when you swap out the LLM that drives them but what an agent or chatbot can accomplish is equally dictated by the choice of runtime for the agent or chatbot.

Circling back on the misplaced worries over DeepSeek, should you be worried DeepSeek’s R1 model is stealing your data? No. Should you be worried that DeepSeek’s free appstore app that connects to R1 is stealing your data? Within reason, yes. Always interact with any application (web, mobile, desktop) with a bit of skepticism. However - if you’re not trying out new AI models as quickly as they’re being released, you’re missing out on a tremendously exciting world. Healthy skepticism, always a good idea…letting fear get in the way of an opportunity to grow? That’s a mistake.

While the definition of AI agents is debateable and there’s plenty of “workflow automations” that are being called “agents” in error, there’s no doubt that we’re starting to solve a lot of our day to day problems with some variation of an “AI agent”. First, a quick definition, what do I mean when I say agent? This is going to change a lot as the technology develops, but for right now - when I say “AI agent” I mean any service or framework that uses one or more LLMs to call an outside function to execute an action. I remember a big ah-ha moment last year when I was reading OpenAI docs and landed on their function calling API, my attitude towards LLMs jumped from “novel chatbot” to “this is going to start acting in our world in ways few are thinking about right now.”

Lot’s of pieces have come together to make this all possible, with LLMs providing the last significant brick in the super cool sci fi AI future. Although there’s an increasing number of AI agent frameworks available, most of the agent discussion still feels more like talking about doing things instead of doing them. I read about new agent frameworks daily, but way less about interesting usage of agents in the real world.

So, let’s change that…I’m challenging myself (and you!) to write less about agents in the abstract and more about agents solving a problem you have. The world needs theory and discussion about what agents mean and how to use them ethically and effectively, but I’m a believer that sometimes the best way to form an opinion on a technology is to get into the dirt and try it out for yourself.

I posted about this on linkedin, feel free to jump into the conversation there or send me a message on bluesky.