Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Sugon showcases its scaleX Supercluster which uses 16 of its 640-card scaleX640 SuperNodes.
my thread on Doubao and Seed1.8 in light of ByteDance's launch event today. They are up to 50T tokens consumed per day just in China.
Gemini hallucinates a lot more. OpenAI will have a more frequent release cadence, they also updated their image model. It's an improvement, but it's definitely not as good as Nano Banana Pro. I still personally prefer Opus 4.5, something about its voice and prose...
Either way, the big 3 all have their niche. I look forward to Deepseek V4 destroying Western labs again in January (hopefully).
^Good thread by TP. He asks why there is so little discussion about them, and the answer is simple: because they haven't focused on Western markets. Once they do, attention will shift.
I tried using their model some quarters ago but their website wasn't even in English, but it is now. Nevertheless, they don't have a good web interface for chat. I primarily use AI on my PC and the other major Chinese players have a decent web interface whereas Bytedance does not.
At their main website, there's a bunch of Chinese names for a bunch of apps without telling the user what they do and for what. All this feels like a major oversight and frankly should be inexusable by now.
If they fix these things, then they should be talked about more. Otherwise they will continue to be omitted in Western discourse.
I tried using their model some quarters ago but their website wasn't even in English, but it is now. Nevertheless, they don't have a good web interface for chat. I primarily use AI on my PC and the other major Chinese players have a decent web interface whereas Bytedance does not.
At their main website, there's a bunch of Chinese names for a bunch of apps without telling the user what they do and for what. All this feels like a major oversight and frankly should be inexusable by now.
If they fix these things, then they should be talked about more. Otherwise they will continue to be omitted in Western discourse.
ByteDance is essentially trying to not even let you know the model behind the scene. to them, it's all about the app experience, which a combination of the model itself and the quality of the apps. So, that's very interesting approach and completely makes sense for them.
I will say that BD doesn't need to work as hard. They can just inject more AI into TikTok and CapCut and people will just use it.
Anyhow, Alibaba here with China Southern where their Qwen model is used to fine tune into the model that the airline needs
I will say that BD doesn't need to work as hard. They can just inject more AI into TikTok and CapCut and people will just use it.
Anyhow, Alibaba here with China Southern where their Qwen model is used to fine tune into the model that the airline needs
I think some of this may be related to this AI data center article about transformers in wsj
it is interesting that they are still building medium voltage UPS tech when AI data centers have moved to HVDC
Nvidia is pushing for 800VDC solutions (800 V direct current I think)
vs the current 415V AC solution
Key efficiency gains
- Up to 5% improvement in end-to-end power efficiency
- Maintenance costs reduced by up to 70% due to fewer PSU failures and lower labor costs for component upkeep
- Lower cooling expenses from eliminating AC/DC PSUs inside IT racks
Any reason why Bytedance does not have a benchmark coding and general llm. Video generation is where they dominate but they are nowhere to be seen in coding benchmarks. Or it it that they just don't focus on it.
As a daily user of all the major Chinese chat LLMs, Doubao probably has the best user experience for general non-thinking knowledge queries. It usually responds the fastest and includes pictures and relevant Douyin videos. The voice and audio interface is the smoothest, and it's a very satisfying experience asking questions using voice and listening to the response.
For drawbacks, unlike other chat apps which all use an abstract logo, Doubao uses an AI avatar mascot, which is annoying to me but maybe appeals to more mainstream people? There's also no dark mode support which all other apps support.
For more in-depth thinking type questions or things where I want lots of data, I like DeepSeek thinking mode or Kimi K2 since they tend to respond with more tables making things more readable.
Recently The Information wrote an article on OpenAI's recent struggles and one highlight is that mainstream users didn't care much about the benchmark of their thinking models and prefer speed of response, so their thinking models don't really drive subscription conversion. Perhaps it's a trap for product focused LLM companies to try to spend all their efforts in beating benchmarks and Bytedance has the right priority on response speed for the masses.
For drawbacks, unlike other chat apps which all use an abstract logo, Doubao uses an AI avatar mascot, which is annoying to me but maybe appeals to more mainstream people? There's also no dark mode support which all other apps support.
For more in-depth thinking type questions or things where I want lots of data, I like DeepSeek thinking mode or Kimi K2 since they tend to respond with more tables making things more readable.
Recently The Information wrote an article on OpenAI's recent struggles and one highlight is that mainstream users didn't care much about the benchmark of their thinking models and prefer speed of response, so their thinking models don't really drive subscription conversion. Perhaps it's a trap for product focused LLM companies to try to spend all their efforts in beating benchmarks and Bytedance has the right priority on response speed for the masses.