Brendon added DeepSeek to our AI tool register and ran it through our traffic light system. No real surprises that it's been ranked as Red for team use - pretty much due to its privacy terms. Assume that your data is going to end up in China. This doesn't mean the team can't have a wee explore. We just won't be putting any client or sensitive data or content anywhere near it.
Quote from Brendon: "So pretty much any input provided to this AI can be used to help train it, and all data is stored on servers in China (as you might expect). So I wouldn't go using this AI for anything of a sensitive or commercial nature."
What the team were saying on Slack:
- Sergey: This is a proof point that you can create outstanding models for a lot less than a few billions and billions of dollars
- Joshua: the model weights have also been released so you can run it locally and fully privately if you like. The 14b param model runs on our macbooks using Ollama with decent performance
- Brendon: This is the first model, where the terms of use state that you as a user MUST verify the authenticity and accuracy of AI-generated content before sharing and MUST label such content appropriately to inform others of its synthetic nature.
- Daniel: It is really good, cost wise and efficiency , just the fact that it stops the race for more powerful GPUS and power consumption it is great news, since it is open source, just fork it to avoid feeding anyone else data-center (China, USA, you name it...)
In a nutshell we'll be keeping an eye on it. Having a tinker with it. But not using it for client stuff in the short to medium term (and until the terms of use change).
Brendon shared this article from Live Science.