Back to Otaru for the first time since June
I flew from Takamatsu to Hokkaido via Haneda, then returned to Otaru for the first time since June and visited SAKE Bar MARUMOTO and HALF TONE.
ARCHIVE
Blog
The English edition of my blog.
→
It's not about generating text. What kind of AI is Jev?
What is the TypeSafe AI decision model Jev? From the challenges encountered when using LLMs for classification and decision‑making, to Jev’s typed outputs, parallel evaluation, RLCD, probability calibration, confidence scores, and its design that allows decisions to be deferred, we examine its mechanisms and underlying theory with diagrams and concrete examples.
Static analysis of explicit purity constraints in Go: Design and verification of mikoto
I created mikoto, which checks the number of lines in Go functions and explicit purity constraints. It handles value-type boundaries, AST and type information, generic type sets, and implicit pointer references, and, based on five bug fixes and verification results, considers the scope and limits of static analysis.
The object returned by Boa’s getter was collected by the garbage collector while the inline cache was being updated.
Investigating the SIGSEGV that occurred in Omoikane’s WPT revealed that the new object returned by Boa’s getter in the Rust‑based JavaScript engine was being reclaimed during the inline‑cache update before it could be placed in a VM register. We are isolating the issue with the JIT, verifying it with Valgrind, and tracking the root fix along with regression tests.
Since generating realms with Omoikane was slow, I looked into Boa’s scope lookup and lexical analysis.
Each time we created a JavaScript realm with Omoikane, it parsed and compiled a roughly 1.1 MB DOM bootstrap. I’ll document how we shifted from a caching approach to improving Boa’s scope lookup and lexical analysis in the Rust‑based JavaScript engine, along with the actual measurements from the test profile.
As AI-driven development progressed, the 30‑minute CI wait became increasingly burdensome, so I revisited the validation of Omoikane.
While the amount of work keeps growing with AI‑based development, the Rust‑based Omoikane’s CI suffers from long wait times for each pull request. This is a record of re‑examining document‑change detection, Rust caching, WPT setup, and sequential pushes based on actual measurements.
How about writing end‑to‑end tests first in order to bring the AI agent development closer to “completion”?
Not only should the AI agent keep solving problems, but we also need it to converge toward the final product we want to build. Consider a development approach that treats passing end‑to‑end tests as a prerequisite.
I finally updated the Japanese phone number validation library after a long while.
I updated the Go phone number validation library go-tel-num-parser-jp that I created in 2018. In addition to handling number categories such as 060 and 0200, I also fixed false positives caused by partial matches.
Tracing with Omoikane how a URL fragment becomes the CSS :target
Which element a URL’s #fragment points to, and how that influences the CSS :target selector and the results of querySelector. Using the HTML and Selectors specifications as guides, we follow Omoikane’s implementation and tests.
How to Choose Between LLM, Embedding Models, Jev, and Classical NLP
Should we entrust everything—search, classification, and decision‑making—to LLMs? We will clarify the roles of keyword search, BM25, embedding models, the decision model Jev, and LLMs through concrete examples.