I study how AI organizes meaning internally. Here's what Qwen 2.5 looks like before it starts thinking.
- 发布
- 09/04 10:19
- 首次发现
- 09/04 15:00
- 类型
- 开发者社区 · RSS
一位研究人员创建了一个名为“嵌入之海”的图像，它是一个地形图，可视化了Qwen 2.5 7B模型的输入嵌入空间。这个地图展示了AI在开始处理任何提示之前，其内部是如何组织意义的。模型所知道的每个词都有一个位置，形成密集的簇（山脉）和稀疏的区域（海洋）。例如，“terrible”、“splendid”和“cruel”是邻居，它们是根据强度而非情感分组的，而“sexual”和“financial”则共享一座山。这揭示了模型在推理之前的内部结构和词汇关联方式。
This is the Embedding Sea, an image I made; it is a topographic map of Qwen 2.5 7B's input embedding space. Every word the model knows has a position. Dense clusters form mountains, sparse regions become ocean. "Terrible," "splendid," and "cruel" are neighbors, grouped by intensity, not sentiment. "Sexual" and "financial" share a mountain. This is the starting landscape before inference, before the model has read any prompt.
I build interpretability tools that track what happens after, how the model moves through this terrain as it reasons toward an answer. I use inference tools like the logit lens, the Jacobian lens, embedding visualizations.
More of the research and interactive visualizations here: https://arianaram.github.io/AIview/
Something I've noticed over the past year: the major labs are optimizing hard for code, tasks, and tool use. Models are measurably getting worse at open-ended conversation, creative collaboration, and just being interesting to talk to. If you've felt like ChatGPT or Claude got "flatter" recently, you're not imagining it.
I'm considering building something in the opposite direction: an AI fine-tuned for creativity and conversation, built on an open-source model (7B-8B range), with actual introspection capability. Not a persona on top of a general-purpose model. A fine-tuned model shaped by interpretability research — it can reflect on its own processing because I can see what's happening inside it during inference.
I'd like to know your opinion. Tell me in the comments: would you actually use something like this?