What can we learn from this?
>>109003741>"Other" BTFO
>>109003741Chinese models are selling their tokens for cheaper and work well enough, so a lot of companies are going to choose the cheaper option. American companies need to lower their token prices to stay competitive!
>>109003753Maybe Other only started in May?
Kramer: Mornin' buddy!Jerry: It's 8:00 in the morning, what the hell are you doing here? What's going on? And what's that humming sound?Kramer: Having breakfast. I gotta be at Kramerica HQ by 9:00.Jerry: Kramerica? What?Kramer: It's my new AI company. Kramerica.ai. Check it out, Jerry.Jerry: I'll pass. So, what is this AI thing all about?Kramer: It stands for artificial intelligence.Jerry: Yeah, I get that. But what does it do?Kramer: What does it do? Jerry, what DOESN'T it do?Jerry: You don't even know what it does, do you?Kramer: Have you been sleeping under a rock, Jerry? Just log into Kramerica.ai, ask the AI a question, and it will answer it in seconds!Jerry: You mean like Google?Kramer: That's a search engine, this is AI.Jerry: What difference does it make? And isn't Google using AI in their search results, anyway?Kramer: But with Kramerica.ai, you can create agents. Say you want to ask that question every hour, 24 hours a day, you can create an agent to do it automatically and text you the results.Jerry: I don't think that's as useful as you're trying to make it sound.Kramer: Listen, I've got an intern coming in later from NYU, computer science major. He's going to help me build it. He says he can connect some Claude API tokens before we build up our own infrastructure and he can vibecode our core AI model. My buddy Bob Sacamano, he hooked me up with some IBMs from the scrapyard but I'm going to need more hardware to scale up our datacenter. If you want to invest, I would-Jerry: NO, Kramer. NO.Kramer: Never took you for a luddite.Jerry: Yeah, yeah, I'm going back to bed. Where is this Kramerica HQ anyway?Kramer: Across the hall, buddy.(humming intensifies)
>OY SHUT IT DOWN STOP NOTICING HUMAN
>>109003999>the - fleshoy - acknowledged>shut - it -dooooooown
>>109003741Chinese models have open weight versions that you can run without being forced to have a permanent, variable expense in your books from having to pay off Anthropic or OpenAI until the end of time. Why would you tie yourself to one of the American giants when they can make your business fundamentally nonviable overnight with a new pricing scheme? Only the extremely wealthy or the extremely stupid do it, which it turns out is plenty of people.
>>109003753>>109003852DON'T BE TALKING SHIT!EUROPE HAS GREAT AI MODELS, SUCH AS...
>>109003741note that chinese models are 10 times as stupid and use 10 times as many tokens for the same task. chatgpt is much better than claude in this regard even
>>109003769And this time they can't even blame child labor
>>109005598Things that never happened
>>109005598
>>109005598Calm down SAAR. It's not your fault the Chinese are building more efficient models AND releasing them as open source.But at the end of the day we all don't need a trillion parameter slop monster when we could have a new well build MoE do 99% of the tasks just as well.
Chineses tech is just better and cheaper, USA future is to be raped by China in every area.
>>109005627
>>109004215It's called API and all the providers offer it, it's nit exclusive to chinese
most openai and anthropic traffic isn't going to show up on openrouterthat said, there is a dearth of cheap westen models right nowanthropic haven't updated sonnet, let alone haiku for a whileopenai haven't updated gpt mini for a few monthsthe chinese models are worse, but for many tasks even that level of intelligence is acceptable.basically what i'm saying is that an 80 iq kimi k2.6 is smarter than the average westerner
>>109005642be still, wumaoyou're distilling western modelswestern models are distilling yhwh himself
>>109005650Nothing personner, timmy
>>109005655sshwhen the superintelligence is born it will treat you kindly too
>>109005678>casually admitting that you're as dumb as a mistral modelyou have stopped seeing the limits of the models because they have surpassed youit's okay, it will happen to all of us
>>109005678not true at all. try pushing it to the limits of stuff it knows, forcing it to think outside the box. then qualitative differences between models become very clear.
>>109005695just sounds like you've been doing babby-work since you graduated.
>>109005708Relax, vibe a little.
>>109005708go make another wordpress plugin mr. software engineer
>>109005655He's just going to shove benchmarks in your face instead of admitting that there's little difference to the average user.
>>109005740non-average user here. fuck you
>>109005613>>109005620>>109005621I didnt save the benchmark screenshots, but you can cope all you want, the open weight models are just not as good.t. self-hosting enjoyer
>>109005609The children are building the power plans that drive the price of electricity down.
>>109005771https://youtu.be/TLEK0UZH4cs?t=85
>>109005756They're good for most tasks people actually use it for.And for the heavier tasks frontier slop models are too expensive.There will be a reckoning in the industry.
They are 80-90% as good at 1/10 of the price
>>109005598>chatgpt>goodIt's fucking ass. And I'm saying this as a person who does a fuck ton of things through AI and uses almost all the popular ones. When it comes to general knowledge, Gemini is much better. When it comes to researching stuff, Gemini and Grok are better. When it comes to coding, Claude and Deepseek are better. When it comes to generating images, it's fucking useless because of the guardrails it has and doing image generation locally is much better.
>>109005807>When it comes to coding, Claude and Deepseek are better
>>109003741>Agentic coding takes off>People realize US models are too costlyThat making things efficient will be the next demand. And that is a bit at odds with the whole "we need to make our money back" bit.
>>109005807I prefer the code claude generates as well, but you pay for it in higher token consumption
You don't need to have the BEST product, only a good enough and cheap enough, the Chinese have learned this long ago and that's why everyone is using Chinese AI now, this is why US AI companies will never make a profit and will just implode and drag the entire American economy with them into the pits of hell.
>>109005771>>109005609Coal mines don't dig themselves
>>109005609They will instead complain about IP theft because their outputs were used to train chinese models. The fucking irony.If plagiarizing the entire internet to train models isn't copyright infringement, then training models based on other models' outputs shouldn't be either.I'm sure after US AI companies are done stealing everyone's work, they will start lobbying the government to put restrictions on training new models though. Especially training based on other models' outputs. It's such a classic. This is the true purpose of regulations, not whatever nonsense commies on reddit come up with.
>>109005650>you're distilling western modelsgonna cry, bitch?
>>109006233of course notsome people enjoy sloppy seconds, i'm just not into that myself
>>109005632Yes, but you can run the Chinese models on your own hardware for no additional cost and serve it on your own terms. For the American giants' models, you are forced to pay them for API access, and you have no guarantees that the cost for that won't change overnight.
>>109005650That just means a year's head start. Open chink models will flourish and offer dirt cheap $/tok rates while the US AI companies go bankrupt.
Even a 24b LLM is accurate enough for most retards. Who is actually spending money for tokens?
>>109003741>methodology >open router because most western people use open ai for ai right
toke
>>109003741That 90% of the code AI actually writes can be done with small models and you really only need the big ones for debugging and planning.I could easily burn through $50 in claude sonnet tokens vibecoding a project but I can get that under $5 by using sonnet just for planning and writing code with qwen 3.6 or deepseek.
>>10901136155 tokes per second
>>109012274>sonnet for planningyikes...
I can only get so much code done with the Chinese models. I keep going back to the big expensive models because I end up churning and wasting my time with the cheaper models. The cheap models can handle a lot of things but they can also screw you up and waste your time.I also work as a systems programmer, not web, so maybe that matters too.
>>109012304>she doesn't know
>>109004066-... . --. .. -. -. .. -. --. / --- ..-. / -- . ... ... .- --. . --..-- / ... .... ..- - / .. - / -.. --- .-- -. --..-- / .-. . .--. . .- - --..-- / ... .... ..- - / .. - / -.. --- .-- -. .-.-.- / . -. -.. / --- ..-. / -- . ... ... .- --. . / .-.-.
>>109003741American companies have more capital, so they're just throwing money at problems, using their sheer scale to brute force solutions. The chinks are taking the longer approach by using actual engineering and more efficient technologies to improve things. This applies to power plants too.American AI companies are doing the same shit US automakers once did. Eventually they will cite "security concerns" and lobby daddy government to impose tariffs or block foreign competitors from the American market, just as happened with cars. That will go just as well as it did back then. (American cars are garbage).
>>109012446.-- .... .- - . ...- . .-. / -.-- --- ..- / .... .- ...- . / ... .- .. -.. / .. -. / - .... . / -.. .- .-. -.- / ... .... .- .-.. .-.. / -... . / .... . .- .-. -.. / .. -. / - .... . / .-.. .. --. .... - .-.-.-
>>109013493Wrong, If anything goolag is the one innovating right now and they're the ones that started this shit before chatGPT even came out, also china is bruteforcing power plants too, from coal to nuclear they're just building whatever it takes and it just so happens that some of them are really efficient.
>>109003741Well no shit, the only time people use open router is to use other models. If they have an American model they are 10x more likely just to use Claude app or chatgpt's app.
>>109013493>The chinks are taking the longer approach by using actual engineering and more efficient technologies to improve things.They release weighs not the training code. If they have some secret super efficient training methods maybe they should pool their resources and stop chasing the tails of Western AI labs and actually surpass them.
>>109003741>striking shift toward Chinese modelsYour chart shows that they are still using American models at the same rate as before. There's no shift, just additional usage in Chinese models as well as the same usage in American models.
>>109003911
>>109015025Efficient doesn't mean you surpass infinite money.
>>109003741>by American AI startups??????????????????????
>>109015025If OP's image is correct, they already won.
