Luke Wang (@Remember2015) profile picture

Engineer, Realist

Shanghai, China
Joined December 2007
Photo by Luke Wang (@Remember2015)
carbon crashed in antigravity but alive in gemini....
1
72
Antigravity 里面的 Orchestrator 模式不支持调整 carbon 的思考强度...有点不够灵活,默认 medium ,解决方案是不用 Orchestrator 模式,而是在 carbon(high) 里面告诉它,“请注意除了规划类任务外,其他一些难度没那么高的任务你可以交给 subagent (用 hydrogen 模型来执行,推理强度high),你自己 review 结果/结论即可”
1
30
Replying to @robust_girl
The cunning fox from La Fontaine😄😄
2
7
840
过去一年半,我们死磕 Browser Use: 把 DOM 压到极致、测了几百种架构,试图教 AI 像人一样使用浏览器。 直到某天深夜,模型的进步用一种完全反直觉的方式,击碎了我们引以为傲的设计和架构。 liaocaoxuezhe.com/article/th…
57
105
20
681
131,254
居然不到 2MB
I wanted to test Fable 5.2's quality with one-shot prompts without human feedback. Everything in this game is from scratch: models, sounds, textures. ~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours! Play it here: claude.ai/artifact/BGY1SztrA…
222
护城河并非只有 cuda,而是 cuda 上长出来的生态
The Nvidia CUDA moat is gone. Anthropic sees all compute as fungible and it only took Mythos 3 days to figure out how to run itself on AMD hardware.
81
16 年前的网站依然存在,omg
50
印度光明节,号称 built in 印度的 Alpie 9B 发布。看了一下,完全是向 Qwen3.5-9B “致敬”: - 架构方面,完全 copy - 分词器方面,词表一个字没动 - 权重结构和数值:基本逐位相同 哦,唯一不同的是模型的自我认知: chat template 每轮都注入指令,问来自哪里就答 “India”。 把 X 网友当傻子么? huggingface.co/169Pi/Alpie_s…
This Diwali, a little more light from India's AI research community onto the global AI landscape. Introducing the latest benchmark results from Alpie Seed, our 9B multimodal reasoning model developed at @169Pi_ai Research. We believe the next generation of AI will be shaped not only by models that scale bigger, but by the research and engineering that make smaller models increasingly capable. We put Alpie Seed through a range of demanding evaluations across reasoning, scientific knowledge, mathematics, coding and software engineering. Here are some of the results: → 82.54% on MMLU-Pro → 84.32% on GPQA Diamond → 72.99% on LiveCodeBench v6 → 61.20% on SWE-Bench Verified The comparison highlights some interesting results. Alpie Seed records the highest GPQA Diamond score among the models shown. On MMLU-Pro, it performs competitively with larger models, while its LiveCodeBench result is within striking distance of substantially larger systems. Its SWE-Bench Verified result of 61.20% is another important milestone, reflecting its ability to work through real-world software engineering tasks. This evaluation used mini-SWE-agent without retries. Beyond benchmark accuracy, we also examined inference efficiency. In a separate profiling evaluation, Alpie Seed recorded: → 36.66 ms time to first token → 41.37 tokens/s generation speed → 9,796.22 tokens/s prompt processing speed These measurements come from a separate profiling setup and should not be confused with the comparative benchmark results. What makes this work meaningful to us is the process behind it: developing a compact model through supervised fine-tuning and reinforcement learning, evaluating it across diverse tasks, and making the resulting weights accessible to the wider community. The comparison includes published results from other model developers. These evaluations were not rerun under a common harness, so differences in methodology and configuration should be considered when interpreting the numbers. Built in India. Developed for a global research and developer community. This is Alpie Seed. And this is one step in a much larger journey for 169Pi Research. Explore the model: Hugging Face: huggingface.co/169Pi/Alpie_s… GGUF: huggingface.co/169Pi/Alpie_S… Ollama: ollama.com/169pi/alpie-seed Kaggle: kaggle.com/models/alpie/alpi… Happy Diwali from everyone at 169Pi.
38
58
14
814
77,125
其实只有豆包的审美是符合的,其他几个都不太对
Replying to @GrayNoteLab
很Nice,个个都是精品,太哇塞了
54
Luke Wang retweeted
just in case anyone doesn't know what "Western Values" are baked in, some quick demonstrations from their dataset. they baked in CIA propaganda and want the model to ignore anything that's not from CIA sources. it's actually a comprehensive collection of CIA talking points🤡
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
41
164
24
1,695
58,525
依然臭不可闻
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
29
Luke Wang retweeted
I reverse engineered tiktok
738
5,956
1,044
49,407
4,097,011
Luke Wang retweeted
《代码的未来》真是一本奇书。
36
40
11
512
62,514
Today we are happy to share one of our first open research releases: Latent Sparse Attention (LSA), a step towards subquadratic attention that matches accuracy of full attention at ultra-long context. Sparse attention attends to only a few tokens per query, but it has to find the right ones. At 10M tokens, that search takes 98% of the attention layer's time. LSA makes sparse attention up to 22.7× faster and full attention 72.4x faster. With @kubicek_ales @caciolai @PengLiangzu @SukritSumant @Ben_Hoov @FraserGreenlee @RMBattleday @radim_urban @thoefler @jcrwhittington
101
338
82
3,037
508,208
Luke Wang retweeted
Today, we open-source P-H3-Image, an image generation and editing model built from MiniMax-H3 by @MiniMax_AI. A SOTA video model, turned into an efficient open image model. - 𝗛𝗼𝘄 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗶𝘀 𝗶𝘁? Generate or edit images 5× faster, in 8 steps instead of 50. A 4 MP image takes 2.6 s instead of 12.4 s. The model is 7× smaller than H3, so high-resolution images are a light download and run on a much smaller GPU. - 𝗛𝗼𝘄 𝗴𝗼𝗼𝗱 𝗶𝘀 𝗶𝘁? Strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. It renders 2 to 16 MP and edits with up to 8 input images. -𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? We reworked every component of H3: the prompt upsampler, the text encoder, the diffusion schedule, the transformer, and the decoder. Use it with Diffusers or ComfyUI. 📄 Tech Report: pruna.ai/blog/oss-p-h3-image… 🤗 @huggingface Open Weights: huggingface.co/PrunaAI/P-H3-… 👉 @ComfyUI Native Support: huggingface.co/PrunaAI/P-H3-…
107
176
58
1,621
162,707
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,367
7,262
5,753
55,942
10,054,503
Luke Wang retweeted
Moving mountains!
54
125
12
2,426
119,379
1/6 We find early signs that frontier LLMs are surprisingly good at *natively* reading, writing, and reasoning about neural network weights. Instead of training, models may treat raw weights like readable, editable code: diagnosing failures and patching them directly.
17
82
23
3,266
471,949