Quick Thoughts on Deepseek R1
If I have to write blog post for every LLM released in the last year or so, it would have been a packed list. Fortunately, not every release is worth writing notes for.
But with the recent Deepseek R1 release, a reasoner model from Deepseek, I thought it was worth for me to document my own experience.
I know it's a speedy model, implementing a somehow-outstanding Reinforced Learning which in-turn, made the response quality top notch.
But what surprised me was the type of inner monologue that it conveyed to think and rethink its' response.
I couldn't pinpoint cathegorically why, but I felt that, compared to OpenAI's o1, it thinks better, at least to my likings.
I was testing it by reviewing a partially-finished algorithm I wrote to efficiently solve a 4x4 or 3x3 image sliding puzzle.
While the approach would be pretty standard, but I felt like I was reading a real person's train of thoughts when it was evaluating my code. It saw crossroads but then able to path its' own reasonable choice. I felt it really contemplated for that 128 seconds.
Amazing, amazing.
I was floored by it. As a human user, this was the feeling of getting the much needed help to progress, and be very confident about my helper. The feeling was profound.
And if this is just what's to come this year, then so be it. Bring it on.