ycui7

5 ms·

1.
by ycui7 7d ago
llm is not a deterministic program. the same model won't even return deterministic answer. what's the point keep the model freezed? if you want deterministic returns, you should set the temperature to 0 to get the best possibility
2.
by ycui7 19d ago
the fact that this author cannot get qwen3.8-27b run at the same speed as qwen3.6-27b, says the article is not worth reading. the author does not know anything about how to run local AI. 3.8 and 3.6 are the same model with different weight.
3.
by ycui7 21d ago
because it has 1T ssd not 4T
4.
by ycui7 29d ago
Cerebras is an uncut whole wafer. Each wafer gets you 44GB SRAM (not a typo, SRAM, not DRAM/VRAM). A few years ago, leading process node wafer from tsmc is $20K/ea without guarantee on yield. A single full wafer likely can run qwe
5.
by ycui7 1mo ago
if you have the VRAM, use offical release. quantized model lose focus after long context and can do damages or thinking loop
6.
by ycui7 1mo ago
that is why we enable web search for the agent. the memory can come from the internet. deepseek-v4-flash needs web search to return true facts.
7.
by ycui7 1mo ago
the a little disappointing part is this is released in BF16. so i suppose no QAT was implemented.
8.
by ycui7 1mo ago
on one side, deepmind makes a lot of advancement in science related application. but, on the commercial side, they struggle to compete with other major LLM providers.
9.
by ycui7 1mo ago
it is funny when people say i am struggling to spend money.
10.
by ycui7 1mo ago
they need a competent gov contractor. recognizing license plate is a fully solved problem many years ago.
11.
by ycui7 1mo ago
so qwen3.x-27b on hardware? or better deepseek-v4-flash on hardware .
12.
by ycui7 2mo ago
Can they still go public ? MiniMax M3 Pro is also coming, then DeepSeek-v4-Pro GA, then GLM5.5. There will only be bad news for them in the coming few weeks/months.
13.
by ycui7 2mo ago
the rational in one’s mind is similar to buying expensive supercar but no driving it daily. owning a few GPUs is a lot cheaper than supercars.
14.
by ycui7 2mo ago
Huawei Ascend NPU
15.
by ycui7 2mo ago
my AC is noiser than my GPU server.
16.
by ycui7 2mo ago
you don’t need new weight. try vllm-moet from github. it will autogenerate 2-bit plane.
17.
by ycui7 2mo ago
K3 is natively trained to mxfp4, if they cannot get a hold of Blackwell chip, it is meaningless. Hopper does not do native 4-bit floating math. Either they have Blackwell with native 4-bit floating math, or they use have Chinese domestic NP
18.
by ycui7 2mo ago
For people with single RTX PRO 6000 96GB or DGX Spark 128GB, vllm-moet is a very good engine, although lesser known. It auto generate a symmetric 2-bit plane for inference and also generate a 4-bit delta cache to recover precision. Support
19.
by ycui7 2mo ago
and it was created by Chinese born Professor and Student.
20.
by ycui7 2mo ago
so GLM won?
21.
by ycui7 2mo ago
discounted competitor could cheat. they can offer subpar model response and sell it as deepseek-v4. it is uneconomical to prove inference providers are cheating, so they get away with it. cheating inference provider does not care if their c
22.
by ycui7 2mo ago
feels like the American domestic manufacturing is done, there is no hope to save it.
23.
by ycui7 2mo ago
it takes extra effort to open source a model even if you had it running internally. even traditional software takes extra effort to get released as open source.
24.
by ycui7 2mo ago
every cloud provider trains on your data, regardless of what they promise. real user interaction is the best reinforcement-learning trace.
25.
by ycui7 2mo ago
OpenAntrophic and OpenOpenAI ?
26.
China recovers Long March 10B rocket(bilibili.com)
11 points by ycui7 2mo ago|3 comments
27.
by ycui7 2mo ago
China successfully recovers Long March 10B rocket following maiden flight, marking a breakthrough in rocket reusability
28.
by ycui7 2mo ago
this is not subsidizing. this is way too expensive for a no-name model.
29.
by ycui7 2mo ago
there is an algorithm called quick-select. getting median of an array should not require full sort of the whole array, only a partial sort is needed to get the median. quick-select does this.
30.
by ycui7 3mo ago
24-bit was created because microphone want to record large dynamic range without gain switching circuit. 96kHz was created to better reproduce 20kHz high frequency, so the digital noise shaping filter does not need to be super sharp right a

More ›