DeepSeek R1 is making shockwaves by releasing their R1 model. It outperforms GPT-4, Claude Sonnet and Gemini and cost less than $10m to develop.

More than that, the entire 604B model (in multiple variants) has been open-sourced together with a whitepaper explaining the approach.

Immediately after it went live, people started running the model on clusters of the new M4 Mac Minis, which can run the full model (64GB RAM models) cheaply.

Well, cheaply compared to what you'd typically need to run a competitive model.

All of this has caused the stock market a bit of panic.

Nvidia took the biggest hit at -15% which wiped a half a bill of their valuation. OpenAI has tried to stay competitive by offering the GPT-4 Mini for free.

I watched the whole thing with popcorn in hand. So far, I've just played with DeepSeek a bit on general-purpose tasks, and it's doing very well.

I'll feed it some more coding tasks in the following days to see if it lives up to the hype.

But I'd love nothing more than to see capable open-source models that we can eventually run locally on our hardware.

I grew up in a time when computing became personal, and I don't like the forced move away from our devices to the cloud.

I love the capability, but it leaves a bitter taste in my mouth.

I would love to see our devices be able to compete in a privacy-focused way.

Yours,

Taj

​