Deepseek r1
(it's been a week)
A new research model just dropped. One, that to my amusement, has all of the usual people claiming foul. The type that falls for scam altman's pied piper song - people who become useful after being targeted by a talented corporate communications team.
>It's a good model sir.
>Noooooo they cheated they lied about their training costs
What do you.. mean? You literally have the model. Print the config out.. read their research paper.. come on man
I can't say I'm surprised. These people will lay themselves down for a multi billion dollar corporation they own no stock in, after asking their 200 dollar model for a chicken korma recipe. China is the enemy after all!
China.. you mean the guys selling me actuators cheaper than you are? I'm Canadian, remember? Why should I be pro america? I'm pro strong engineers. And they're good engineers.
People will do the bare minimum to feel like they're not being left out. The only reason they're excited about deepseek r1 is because you can get 10k likes making a somewhat tasteless post about Muh Xi Ping Censorship - because the model refuses to say something that is illegal to say in China.
But maybe I'm not any better than these people (I mean, I am a lot better than them, but as a thought exercise, what if I wasn't?). I am out here, making fun of these so called "machine learning guys" on social media, and not having even tried the model myself.
What's the point? What is my actual goal?
I haven't the time or energy this week to try out a new model. Not only did the whale bros put it out on a Monday, before a total slugfest of a work week, they also felt the need to come up with a few breakthroughs as well. You're telling me I'm going to have to learn RL.. ugh
I'm burnt out man. I'm burnt out by the acceleration. I'm supposed to be e/acc. But it's Sunday and I feel dead. I want to go to an art gallery with my friends. I want to get food with my family. I want to go see Finneas live. You feel me?
I mean, what's the point of keeping up with every single research paper? Every week there is something new. It's the singularity man. I can't keep up with it. That's the definition of it. Just a constant siren buzz of progress. I'm going to throw up.
I have a job. I have a niece. There are things I like doing. I like skating, I like going out sometimes. I want to lift weights more.
By the time I will have caught up on R1, the qwen guys are going to come out with a better model. Then, Tech Shrek and Froge will cook up a sampler that increases its ability for free. Some other folks will figure out how to give the model sympy and allow it to cook with tools. It's open source now. The progress is going to jump
Man. Fuck this. I'm going to mess with my zig based CAD app
I'm burnt out man
Please stop accelerating
I used to try to keep up with ML research and happenings. There was a stretch where I got a little burnt out, and just gave up on it all. I didn't read a single research paper, compared to my pace of one a day.
Then, I decided to get back into it. It took me a day to catch up.. because everything before the previous three months was made obsolete by this week's research..! I didn't even need to read it
These model improvements are more of the same. In terms of impact.. the models are getting better. Yeah, that's what technology does. One day you're on a nokia brick phone, playing snake, the next your google pixel is explaining what bird you've spotted on a hike.
You should generally have the future improvements of ML models as an expectation. You shouldn't really be surprised (though you're allowed to feel vindication!).
Just remember that you can stop for three months and get caught up in a day. You don't need to be at the bleeding edge.
Besides, it's only the bleeding edge this week. Hooo boy
That said, you should read deepseek's research papers. They're a gift. And not the R1 paper (I haven't read it yet). I'm talking about the base model paper. Go read it
