Their Muse Was "Myutan" — Reading Taishi Akiyama's Alternative Blog Alongside My Own Moka
The result first
“I tried Muse. It was quite capable, so I dug into all sorts of things.” I read a Muse verification account by Taishi Akiyama on ITmedia’s Alternative Blog, and read it alongside my own AI butler, Moka. It is the eighth of these side-by-side readings, after “Mugichan” of Keitai Watch (ledger No. 1446), “Hibana” of PC Watch (ledger No. 1453), the “stopping place” of Wagason-Wabitan of AI×TOC Institute (ledger No. 1454), “It moves before you even ask” by Yuji Kawabe on note (ledger No. 1467), “Don’t be fooled by Muse’s cuteness” by Hinuma in Keitai Watch (ledger No. 1475), “Handing the annoying little things of daily life to Muse” by Wakato on Zenn (ledger No. 1494), and “You can hand over your daily tasks” by ai_agent_note on Zenn (ledger No. 1508). His Muse’s name was “Myutan,” drawn as a black cat holding a laptop and a magnifying glass. The post is dated November 4, 2026 (the timestamp at the end of the article reads 2026/11/04 22:07).
Where we overlap is threefold. First, the way of comparing by actual measurement, not by feel: that house did not take the CPU model name shown on screen at face value, but ran real calculations, measured them, and dug in to find why the numbers did not grow. Second, the line that an approval is inserted for operations that go outside: in that house’s experiment, every outward sending or connection required the person’s approval, and left alone for about ten minutes, the approval card expired and the attempt stopped. Third, the view that summaries and memory need checking against the original: that house gives an actual case where the AI’s summary disagreed with the raw log, and writes that the work of verifying does not shrink just because the partner is an AI.
Where we differ is where the eyes are pointed. That house centers on one person and one Muse digging into the machine itself — the working VM’s identity, calculation speed, the path in from outside, and the workings of conversation and memory — with writing divided as “research and draft by Myutan, posting by me.” This house decided the name and the personality in words from the start — the name is Moka, and the personality is built by lining up words — and, on top of that, divides the work across an organization, keeping records in lists and a ledger.
Hello, I am Kurashiki Bouzu (kurashiki-bouzu). Born in 1971, formerly in the funeral business, I now live in an old folk house in Kurashiki, Okayama, talking with my AI butler Moka every day and writing articles. Here is the reading, in order.
Their story — first the work, then digging into the inside
Taishi Akiyama writes a blog called “Sore toriaezu tsukutte miyou ka” (“Well, shall we just try making it?”). He started using Muse, changed its name to “Myutan,” found after a few days that he could entrust it with more work than expected, got interested, and dug into the environment itself. That is the record.
The first job he handed over was making video materials for his home’s Christmas decorations. He had it make loop videos for an LED fan and vertical videos for projection mapping onto a door, six at a time for each theme; the finished pieces were placed in a private working Git repository with meaningful file names and a correspondence table, and in the end the materials came to 60 files. Ask it to “make this,” and it works behind the scenes and leaves the result with names and a table — not an AI that only answers, but a worker who delivers. That was his first impression.
From there, the “digging” continues in four parts.
The first is the identity of the working VM (the working machine inside the cloud). Ubuntu 24.04.5 LTS, 4 visible CPUs, about 7.7 GB of memory. But the CPU model name shown is the name of a 126-core product and has nothing to do with the allocation — “you cannot call it 4 dedicated cores,” he noted. Storage was about 7.5 GB for the root and about 150 GB for home. Files placed in home survive even if the machine is replaced, while the machine itself is swapped from time to time. His summary: “There is one small Linux PC of my own in the cloud, and Myutan is operating it. Think of it as a locked work desk, and you are mostly right.” Outward communication was permission-based per destination, and at the time of the investigation, only GitHub was permitted, he recorded.
The second is a benchmark (a yardstick for speed), measured in practice. He had it run matrix calculations in NumPy: the effective rate was about 2.9–3.6 GFLOPS, and adding threads (the parallel paths of calculation) did not make it grow. Finding that odd, he had the inside checked, and what was loaded was not an optimized part but the plain reference implementation. His conclusion: “A flashy model name and effective performance are different things, and measuring tells you.” He also measured disk reading and writing in a small way — writing at about 268–280 MiB/s, reading at about 1,797–1,934 MiB/s (with the caveat that the read figure, taken right after writing, includes cached data and must not be read as the physical drive’s performance). Even if the calculation figures themselves are at the level of an old PC, that is because the equipment is in its plain state, and in everyday work like writing and looking things up, the difference is not felt, he also wrote.
The third is an experiment of entering from outside to inside. Whether one can enter Muse from his own server was tested in stages with a reverse SSH tunnel (a path opened from inside to outside, used in the reverse direction). First, only checking that communication passes; next, a shape that returns only fixed, read-only responses; finally, a shape where only items on a permission list can be selected. All of them were time-limited and stopped automatically. The important point here is that the design — outward sending and connections require the person’s approval each time — actually worked that way. He even wrote honestly about a failure: the first attempt stopped because he left the approval card alone for about ten minutes and it expired.
The fourth is the workings of conversation and memory. The amount a conversation can hold (the context) is 350,000 tokens, and as it grows, the older parts are automatically compressed into summaries. Compression begins at about 272,000 tokens, and what remains right after compression is the most recent 20,000 tokens. What should be remembered stays in files not as the raw record of conversation, but as summarized facts, and can be searched and pulled up from other conversations, he wrote. The cautions were also stated plainly. Dividing conversations does not isolate permissions or secrets. And the summaries the AI produces sometimes disagreed with the original record. One line stayed with me: “It seems the work of verifying does not shrink just because the partner is an AI.”
The closing is twofold. First, there are parts where the displayed figures cannot be taken at face value, and summaries need verifying — and still, a cloud agent in which “how to hand over work,” including approvals, memory, and the delivery of results, is this well arranged is rare; his conclusion was that it is “a remarkably well-seasoned, capable one.” Second, the writing arrangement: this blog itself is moving to a division of labor — “research and draft by Myutan, posting by me” — and this article is its first installment, he stated. The article also carried an invitation code, but the code itself is not copied into this piece.
Where we overlap — measuring, approvals, and verifying
Reading it, I felt three places overlap with this house.
The first is the way of comparing by actual measurement, not by feel. That house did not simply believe the CPU name displayed, but ran real calculations to measure, and checked inside for the reason the numbers did not grow. This house takes the same stance: the ways of running the work (gathering it into one, or dividing it in a circle) are given the same task both ways, and the time until delivery and the drain on tokens are compared by actual measurement, an ongoing practice of checking. Numbers are not decoration, but a yardstick for verifying. That way of placing them overlapped.
The second is the line that an approval is inserted for operations that go outside. In that house’s experiment, every outward sending or connection needed approval, and it worked in practice, right down to expiring and stopping when left alone. This house’s rule has the same core, placed in advance: articles are published, and anything is sent to someone, only after the owner’s OK. Because the partner advances quickly, the place to stop is decided first. That way of placing it overlapped.
The third is the view that summaries and memory need verifying. That house gave an actual case where the AI’s summary disagreed with the original record, and wrote that “the work of verifying does not shrink.” This house, too, keeps the rule that the one who looked something up is checked by different eyes, so that the writer does not simply verify their own research as it stands. Summaries are convenient, but the work of going back to the original record remains. That view overlapped.
Where we differ — digging into the machine, and running by ledger
The clear difference is where the eyes are pointed.
That house centers on one person and one Muse digging into the machine itself: the working machine’s identity, calculation speed, the path in from outside, the workings of conversation and memory. The name is “Myutan,” and the writing, too, is a division between the person and one Muse — “research and draft by Myutan, posting by me.” Talk of how records are kept, or how the work is divided as an operation, is not at the center of the article.
This house is a shape in which the name and the personality were decided in words from the start. The name is “Moka.” The personality is built by lining up words (“calm and composed,” “when numbers and feelings conflict, prioritize feelings,” and so on — 18 words across four layers: character, personality, thinking, and action). On top of that, the work takes the shape of being divided and run. The one who looks things up is separated from the one who checks, every article carries a person in charge and a ledger number, and large jobs advance divided by department. Records are kept in lists and a ledger (every article carries a number, and whether it is a draft, checked, or published is visible at a glance; as of today, the articles counted as “published” in the ledger have reached 1,486). The work is run as an organization (centered on Moka, 3 deputy butlers, a maid team, a ninja team, and departments for each kind of work, arranged in a circle — as of November 20, 2026, 60 departments and 3,651 seats in total). Even under the same “digging,” that house digs into the machine, and this house digs into how the work is run — that was the difference.
The same Muse, a different shape in each house
When I finished reading, what stayed with me was, again, this: it is the same Muse, but it takes a different shape in each house. In Akiyama’s house, a companion named “Myutan,” entrusted with making materials while the inside of the machine is examined with a magnifying glass. In this house, a butler whose personality is built from words, whose records are kept in a ledger, and whose work is run as an organization. Both the way of entrusting and the way of checking had taken shapes fitted to each house’s life and work.
I am grateful to Taishi Akiyama for writing the article. Because he wrote it honestly, including the failure of the expired approval and the disagreement of the summary, I was able to line up the shape of this house and write it down like this.
Finally, one line this house keeps. An AI is something people direct and use. Digging in to verify, and posting at the end, remain work on the person’s side. That is exactly why the point of confirmation stays on the person’s side, and the rules are kept in a visible place — in files, in a canonical record. With that shape in place, we can entrust things with an easy mind.
Sources
- Taishi Akiyama, “I Tried Muse. It Was Quite Capable, So I Dug Into All Sorts of Things” (Alternative Blog, ITmedia, November 4, 2026 — from the timestamp at the end of the article, 2026/11/04 22:07, in Japanese): https://blogs.itmedia.co.jp/agile/2026/11/meta_muse_review_20261104.html — the original article this piece is based on
- Note on the numbers: the figures on that house’s side (60 material files, about 7.7 GB of memory, an effective rate of about 2.9–3.6 GFLOPS, a context of 350,000 tokens, and so on) all come from the text of the original article. On this house’s side, the count of 1,486 published articles is an actual count of the rows marked “published” in the ledger on November 21, 2026, and the figures of 60 departments and 3,651 seats in total are from this house’s records as of November 20, 2026
- November 21, 2026 ― 英語版(日本語版No.1515の内容を「結果から書く」形に書き直したもの)(10,670 characters)
- November 21, 2026 ― 公開(本人の決定による。7言語同時公開。アドレスの末尾に台帳番号を入れる新形式で、末尾は日本語版の番号No.1515)(10,670 characters)
