There is just a trove of comments here, I was going to pick out the best, but honestly, there's something in here for everyone.
Ok here is the obvious stand-out:
Also, user KamalaHarrisWaifu has a lot of "takes" in the thread.
I actually have no idea to what extent the cost of Deepseek (or other Chinese models) are subsidized. One thing is true though, the grid these things exist on in China is owned by the state, and not privately owned, which I'm sure changes the calculus. (Not to mention the sheer capacity of said grid being off the charts).
I’m forced to use LLM bullshit at work and out of curiosity I tried a 5 dollar trial plan for some Chinese models including deepseek, replicated my usual workflows and they completely destroyed the expensive western models on our corporate plan.
They just did what I asked quickly and efficiently, compared to Claude which bumbled around and seemed to be trying to drive up the costs as much as possible before getting close to what I had asked for.
I still don’t really care for the technology but I am impressed it’s reached a point where it can chuck out a prototype application in 15 minutes for less than a dollars worth of “tokens”.
Just from fucking around with them I've noticed Anthropic models in particular tend to just not take what you say at face value. Like they will twist your words and basically respond to something you never even said in the first place. Or they'll decide that you asking what type of screw to use means you want a rundown of what every type of screw is and the pros and cons of all of them. lmao.
I remember a bunch of people complaining about the shift from ChatGPT 4 to 5 and how 5 "lacked warmth" and "felt robotic." Comparing some of the responses (this is the only context in which I use ChatGPT), I actually found 5 to be much less obnoxious. It acted like a machine instead of a cloying manipulative sycophant.
I literally can't stand using western models partly because of this. They just deep throat it the whole conversation. Chinese models are much more likely to just do the task and shut it. The worst part is recent western models try to be quippy or talk like they're your buddy i guess? Like I've noticed some of them will now start responses with something along the lines of: "Oh, dang! That's a pretty interesting thing!" or like when I was showing Google Gemini an error log and asked it to pick out the errors I needed so I didn't have to read the whole thing it would be like "Look at that! We're so close to this thing working! 90% of the way there. Just one little thing needs fixed..." ugh, shut the fuck up. Do what I asked machine. Why do they insist on making it talk like that? It's gross.
They wanted a cloying manipulative sycophant
I want my robots to feel robotic
v4-Flash is so fast you can't even keep up with what it's doing half the time.