They haven't even fixed the antigravity.
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Weren’t people dooming about Qwen when Junyang Lin said he was gonna leave?
Feels like Qwen’s sped up and improved their releases…
Feels like Qwen’s sped up and improved their releases…
he was probably the hindrance.
my thought on this given the commentary on this from Qwen team member. Huge gain here if Alibaba's own Zhenwu-890 provides the compute it needs for self improvement.
Also, this removes the bullshit about domestic chips can't be used to train. If you can run AI models on them to improve kernel, then they can also write kernels for them for CPT and post training.
Gemini please analyze capcut bytecode and recreate it, ooh and make no mistakes.
The Pareto Frontier. Chart made with Claude.
5 out of 7 models are Chinese on the pareto frontier. Looks pretty good. Qwen 3.7-Max is clearly the best non-US model publicly available and also very cost-effective for the intelligence you get.
Someone on Twitter said that V4 flash is now the new Gemini Flash given that Google no longer seems interested in releasing very cheap but strong models. That seems right. Compare V4 flash with Gemini 3 Flash on the chart. Same intelligence but crazy cost difference.
5 out of 7 models are Chinese on the pareto frontier. Looks pretty good. Qwen 3.7-Max is clearly the best non-US model publicly available and also very cost-effective for the intelligence you get.
Someone on Twitter said that V4 flash is now the new Gemini Flash given that Google no longer seems interested in releasing very cheap but strong models. That seems right. Compare V4 flash with Gemini 3 Flash on the chart. Same intelligence but crazy cost difference.
China’s AI start-up funding triples to US$16b in first quarter amid bets on LLMs.
Funding for China’s artificial-intelligence-related start-ups jumped nearly threefold year on year in the first quarter, as investors poured capital into large language models (LLMs) and embodied AI amid growing optimism over the country’s technology ecosystem.
AI-related start-ups secured more than 110 billion yuan (US$16.2 billion) in the first three months of the year, representing a 185 per cent surge from the same period last year, according to data released on Thursday by Beijing-based venture capital and private equity research firm Zero2IPO Research.
The AI boom has helped lift China’s broader private equity and venture capital market. Total investment activity reached 2,568 deals worth 234.4 billion yuan in the March quarter
Funding for China’s artificial-intelligence-related start-ups jumped nearly threefold year on year in the first quarter, as investors poured capital into large language models (LLMs) and embodied AI amid growing optimism over the country’s technology ecosystem.
AI-related start-ups secured more than 110 billion yuan (US$16.2 billion) in the first three months of the year, representing a 185 per cent surge from the same period last year, according to data released on Thursday by Beijing-based venture capital and private equity research firm Zero2IPO Research.
The AI boom has helped lift China’s broader private equity and venture capital market. Total investment activity reached 2,568 deals worth 234.4 billion yuan in the March quarter
That makes sense. V4 flash is a great model but V4 Pro was overpriced compared to Kimi & MiMo. This resets everything.
Speaking of DeepSeek. Their funding round just keep getting bigger (was $7 billion last week).
Wenfeng is probably the most idealistic and idiosyncratic founder in AI. He also mandates max 6-8 hours per day of work while keeping the organisation as flat as possible. Hard not to root for the company.
P.S. DeepSeek noted in their V4 model card that they are getting tons of Huawei inference compute by EOY. So if they are permacutting prices now, then that bodes well. They should be able to use that inference compute to massively scale up to meet demand.
Speaking of DeepSeek. Their funding round just keep getting bigger (was $7 billion last week).
Wenfeng is probably the most idealistic and idiosyncratic founder in AI. He also mandates max 6-8 hours per day of work while keeping the organisation as flat as possible. Hard not to root for the company.
P.S. DeepSeek noted in their V4 model card that they are getting tons of Huawei inference compute by EOY. So if they are permacutting prices now, then that bodes well. They should be able to use that inference compute to massively scale up to meet demand.
After a month of release, people are discovering the value of DeepSeek V4 Flash. Since its release, it has gained usage every week on Open Router and is now the top model by tokens (weekly chart).
I have already used over 1B tokens this month of V4 flash using the official api (not through OpenRouter) and it has cost less than 100 yuan.