広告効果測定
同意した場合のみX Pixelを使用します。
メールアドレスや制作内容は送信しません。
Some content is automatically translated from Japanese. Untranslated sections are shown in the original language.
TIMELINE ARCHIVE
Follow public updates, creation logs, and related releases from newest to oldest.
Updates
6
Creation logs
3
Related releases
0
TIMELINE
Public records are arranged in one timeline without duplicating the same event.
Sep 6 02:58
I prepared an external evaluation for portability and found that the reproduction of the learned portion is still just under half, while the complete match on the unlearned side remains at zero. The internal separation of "I can answer now / evidence is needed" under the assumption of opening the search did not overcome the gate. I continued to test hypotheses around the membrane, such as coupling interference and re-reproduction after growth, but neither stratification retention nor known reproduction moved forward.
We have stopped the practice of intentionally halting in the middle and shifted to a policy of continuously testing and eliminating hypotheses toward our goals.
After isolating the types of failures, I will try to apply the next fix.
Aug 19 12:17
For long software-work requests, fixed the conflict between the C++ language and tool paths and confirmed 5/5 first responses on unseen workflows, 9/9 transitions, and 2/2 recoveries. Thirty-seven regression tests, including natural conversation, also passed.
The initial silence was caused not only by the amount of learning, but by the language-side speech membrane incorrectly suppressing a learned path on the tool side.
Add a complex request that is less like the training material and test the stability of the first response.
Aug 5 17:02
Since July 28, the system has moved toward sharing pure multicore C++ SNNs across the Lab, browser learning, and natural chat. It now asks follow-up questions for unknown input, gives acknowledgments after teaching, observes session boundaries, and separates self-echo history from human-taught history. In live conversation, exact matches sometimes work after teaching, but unseen paraphrases and topic changes still produce silence or unrelated replay, so it has not been promoted to natural chat.
Direct conversation tests revealed summaries contaminated by self-echo and the replay of past answers triggered by weak associations. The C++ source, provenance, and acceptance membrane now suppress these effects, and test-generated history has been moved aside and isolated, but contention in live shared state remains.
Finish the C++ response-generation path so that content taught in the same session can be reconstructed reliably for unseen paraphrases without producing unrelated answers after a topic change.
Jul 28 07:57
Since July 17, response generation has been rebuilt around multiple C++ SNN cores—topic, relationship, match, permission to speak, inhibition, and recall—rather than Python convenience. Reproduction of known cases and paraphrases of the same facts now works much more often, and wrong answers to strange questions have been greatly reduced. However, the ability to apply knowledge to genuinely different unseen questions remains weak, and unseen exact matches are still nearly zero.
Before adding more volume, the real linkage between questions and answers needs to be repaired. Inflating the score with host-side judgments was deemed outside the main path.
Rather than trying to raise unseen F1 while continuing to emit the full answer verbatim, run an experiment that adds a sequence-generation core capable of producing partially correct answers.
Jul 17 11:38
A small conversation experiment showed that the system can learn word sequences but is weak at actually answering a question. Before increasing volume, the next stage is to repair that linkage.
Shifted toward verifying whether responses actually follow instructions before trying to improve quality by adding more data.
Because swapping the answers barely reduced the score, raising the suspicion that the system was producing formulaic sentences.
Choose the next step either through an experiment that repairs the question-answer linkage or by comparing the system with a similar existing design.
Jul 16 18:20
While advancing a multimodal AI experiment designed to run only when needed, prepared a roughly 15-minute explanatory video script for general and internal audiences. Possible future accuracy improvements are now being kept as references rather than implemented immediately.
Because I want people outside the project to understand, without exaggeration, what it aims to do and how far it has progressed.
Read through the script together and begin preparing diagrams or a screen recording.
Does this explanation communicate what kind of experiment is being built?
7 2026 15
The SNNLab project was started.
12 years of engineering experience.
