OpenAI o3 Looks at GPT-2 Through the Logit Lens
I read the paper Dissecting Recall of Factual Associations in Auto-Regressive Language Models by Geva et al., and I highly recommend it. If you’re curious about how large language models process in...
I read the paper Dissecting Recall of Factual Associations in Auto-Regressive Language Models by Geva et al., and I highly recommend it. If you’re curious about how large language models process in...
In the last post of this series, we tried the instruction-tuned IT model. In this post, we focus on the raw PT model. The following is Table 15 from the Gemma 3 technical report. It compares the p...
In the first post of this series, we examined the specifications of the Gemma 3 model. In this post, we will actually run it and get an intuitive understanding of the inference process. I will assu...
Recently, I’ve become quite fascinated by small LLMs such as Gemma 3 and the idea of running them locally. I tried out Ollama, which builds on top of llama.cpp but wraps it in an elegant, user-frie...
I have attempted to learn the Transformer multiple times. When the now-classical paper Attention Is All You Need first came out in 2017, I was attracted by its title. That was about 5 years after ...
时空一渺, 转瞬即逝, 光影留真, 方得长存。
既然选了 M43,应该是冲着便携,GX7 比 GX8、GX85 和 GX9 都更小更轻。 一个电池 GX7 能拍 330 张,GX8 拍 300 张,GX85 拍 290 张,GX9 只能拍 260 张。 GX7 性价比更高,2000 一个套机。 镁合金机身,结实,磕碰不心疼。GX85 好像是塑料的。另外电子快门 GX7 是 12bit,GX85 阉割到 10bit。C 模式从三...
微单这步棋走对了吗? 昨天搞了个 10 年前发布的松下 GX7,把玩了一会,什么都很赞,就是这个电子取景器我真不理解。只用过单反,这是第一次用电子取景器。调了半天屈光度,怎么都没法完全调清楚。我的单反没这个问题,贼清楚。 清晰度暂且不说,微单和单反的光学取景器体验真的是没法比。发着光的数字显示器,完全没有了光学的自然即时感,晚上还亮眼,贴这么近没人担心蓝光伤眼吗?单反上的光学取景和 LC...
学习摄影,顺便也学学摄像。看到我的松下 GX7 视频模式里有快门优先,非常疑惑:难道快门速度不应该正好是帧率的倒数吗? 于是研究了一番,才知道原来还有 180 度快门这个说法,也学到了特殊快门的作用。太有趣了!
问了下 GPT。顺便感叹下,大模型时代学摄影真爽啊! 根据需求决定固定快门速度还是光圈 场景 固定快门速度 固定光圈 电影感画面(自然的运动模糊) 是(1/50s、1/60s 等) 可调整 浅景深(背景虚化) 可调整 ...