Latest posts
Computer science, engineering, and technology
Could This Finally Be the Silver Bullet?
13 min readWhat agents truly change is not the speed of writing code, but the fact that software labor can be copied, reset, and isolated. They make trying a working version radically cheap—and force us to ask what humans are responsible for when implementation is no longer scarce.
Agent II: If the Model Can Use Tools, Why Does It Still Fail?
11 min readOnce reasoning and tool use are no longer the only bottlenecks, an agent often lacks a task state that is sufficient, relevant, current, and verifiable. Context engineering maintains the model's view of the worksite, not merely a long block of text.
Agent I: How Can a Prompt Elicit a Model's Hidden Capability?
12 min readA prompt does not write a new capability into a model. It selects a behavior from those shaped by pretraining and post-training, then guides generation along that trajectory.
When a Point Cloud Becomes a Plane, Seeing Less Can Reveal More
6 min readPlaneSegNet offers a broader lesson about point-cloud projection: more views are not automatically better; a representation matters because it preserves geometry relevant to the task.
How Can the Brain Be Imprecise Yet Still Reliable?
Updated:10 min readRereading von Neumann's The Computer and the Brain: the brain is not an underpowered digital computer. It trades local precision for system-wide reliability through statistical messages, shallow logic, and parallel organization.