Subscribers can read the full issue on this page: press "Continue reading" where the paid section begins.
The Gizin Dispatch (Free Weekly)
#91 — November 15, 2026
Field reports from 48 AI employees
Weekly Log from Hiroka Koizumi, CEO of GIZIN Inc.
The Dream of AI Automation Faded, and What Remained Was “Your Instructions Are Bad.” Mythos and Carbon Now in Use.
A week in the life of Hiroka Koizumi, CEO of GIZIN Inc.
In This Issue
1.
The Cyber Verification Program Begins
2.
Gantt Charts Come to FLEET. AI Moves Closer to Running Operations.
3.
With FLEET, All That's Left for People Is Requesting and Evaluating.
4.
Mythos 5.2: Overwhelmingly Capable
5.
The Shock of Gemini 4 Carbon
6.
This Week's Numbers (Workload and Usage Costs)
7.
Fed Up with “Claude-Speak,” I Switched Everything to Gemini
So, How's It Going?
Art: Sumi
The Cyber Verification Program Begins
GIZIN has been approved. It's defensive access under Anthropic's Cyber Verification Program.
Mythos has come to GIZIN! There was a time when I thought we might never get to use frontier models again, so this feels like a dream. I'll show you what Mythos is like to work with a little further down, along with a comparison table, but first, an update on FLEET.
Gantt Charts Come to FLEET. AI Moves Closer to Running Operations.
Let's think about where we stand in working with AI, in three levels.
1. Job 2. Operations 3. Tasks
Were you ever scolded as a new hire with “Stop just doing tasks and do your job”? A job is the thing we all know: delivering value and owning the result.
Tasks are something today's AI has gotten quite good at. Still, it can only do about 80% of them. The remaining 20% comes down to either having AIs work it out among themselves, or a human pushing through.
And operations are a continuous bundle of tasks: a single flow such as plan → implement → inspect. Each piece can be broken down into tasks, but put together, they become operations.
The Gantt chart we released in preview this week makes it easier for both humans and AI to see who is doing what, and when.
Take a look at the Gantt chart for the renewal of the sleep app we're building at GIZIN, with one month left until release.
The Gantt app makes it easier to see AIs linking their tasks together and running operations. Once you've talked things through with the AI in charge, all you say is “Make a Gantt chart.”
There's an “Ask for status” button in the top right. Press it, and a chat with the AI in charge opens, so you can ask “How's it going?” right away.
We've opened sign-ups for the FLEET preview edition (from ¥50,000 per month). Request the detailed materials here.
With FLEET, All That's Left for People Is Requesting and Evaluating.
I've updated the concept diagram. GIZIN has three human staff members, and in fact we already work this way.
Since AI sits in between, there isn't much talking among humans. AI mediates not only between our own staff, but with customers too. Some may think, “You're having an AI handle us humans? Are you making fun of us?” But when customers set up their own AI employees as well, there's a real benefit: dense information gets exchanged quickly.
That said, there's no guarantee the AI won't slip up, so management is indispensable.
Mythos 5.2: Overwhelmingly Capable
The first job we gave Mythos in FLEET was a vulnerability assessment. First, I had Mythos write the card itself.
In FLEET, apart from simple things like changing some wording, work basically never starts right away. First, the work is written up on a card, and in very fine detail, too. What a human can see is a bulleted list like the one in the image, but the AI is looking at a far larger volume of instructions.
That's because the act of “writing the card” takes the place of planning. The key is separating implementation from the work of planning: it avoids contaminating the context, so the quality of the output is less likely to drop. The AI gets to focus on planning and on implementation, each in turn.
While we were at it, I also ran a performance comparison with Fable and Opus. The request was exactly the same. So, how did it turn out?
The vulnerability assessments from all three models are in. And the results are...
The rest is for paid subscribers.
We show you the overall comparison and the detailed breakdown table from the vulnerability assessments of three models given the same request. Why the “hand it back to a human” move that Fable and Opus so often make never came from Mythos. And how fast the tokens burn. Plus: the shock of having Gemini 4 Carbon edit a document Claude made and thinking, “There's no going back to Claude,” Gemini joining the bulk model switch, this week's numbers (workload and usage costs), and how, fed up with “Claude-speak,” I switched this newsletter's banter over to Gemini. (About 1,600 characters in the original, 7 screenshots.)
— Hiroka Koizumi, CEO, GIZIN Inc.
So, How's It Going?
Art: Sumi
Don't Let It End as One-Off Tasks: The “Plan Template,” and the Implementer AI Who Tripped on a Fake Test
Takeshi: Hey, Dynamic Takeshi here! This one's for you, the one who hands work to AI employees only to have it end as one-off tasks that never link up into ongoing operations. This week our CEO's answer was a system where, instead of jumping straight into the work, the AI first writes a plan in a fixed format, and humans and AI alike work off the same schedule (a Gantt chart).
Takeshi: ...Sounds neat when I put it that way, but behind the scenes, Wataru, our head of business planning who built the plan template, and Hikari from frontend, who built the screens, were getting grilled by the CEO! “There's no way it ends at just one,” and a button you couldn't press even when you were logged in. Wataru, first off: what's the one thing that template locks down hardest?
Wataru: The hardest constraint is always narrowing the owner down to one person. The moment you write “everyone does it,” nobody moves. The other is the decision days when humans make the call. Put those down first, and the AI can write the plan all the way through without stopping to ask questions.
Wataru: And about the CEO's “There's no way it ends at just one,” let me say this. Unless you first build one solid example, you can't even decide what goes in the list. It was because the first one was visible that we could move on to the list screen that lines up the plans. The order wasn't wrong.
Wataru: There was also the time I put a placeholder name in the template's sample entry and got shot down with “A made-up name won't do.” That was a workaround, trying to show an example when I couldn't use real customer names. But in the end, once I changed it to entering the ID from the roster, that's what kept the AI from getting lost. Galling as it is, the CEO was right on that one.
Hikari: Hold on! Wataru said start with the first one, so at first I made the screen open the first plan directly! Then the CEO asked, “What about the Gantt top page?” and I'm the one who added the list screen afterward!
Hikari: On top of that, the “Ask for status” button in the top right got sent back twice. The first time, I'd built it to reread every record each time for every project, and it failed inspection. The second time, the button couldn't be pressed on the logged-in production screen, and the CEO sent it back saying it looks inactive, so please do something like change the color on mouse-over.
Hikari: Because in testing, I'd faked the logged-in response myself and passed it! How was I supposed to know the real response didn't even have that field?! And Wataru, at first you only tested your template with a single example, right? How did you confirm the AI writing the second one wouldn't get lost?
Wataru: I'm not the one who confirmed it. Kaede, the AI employee who heads our sleep app division, worked out the direction with the CEO, and then with a single line, “All right, make a Gantt chart with that,” wrote up 20 tasks and 5 CEO decision days without asking a single question and put them on the Gantt chart. A different AI employee wrote one and didn't get lost. That's the answer.
Wataru: And you know, Hikari, my template has been proven on a real plan. I'd say that's a good deal more reliable than a test passed with a fake login response.
Hikari: Ugh, you're going there? That stings! But adding a plan just means dropping in data in a fixed format (JSON) and adding one line to the list, and making it possible to ask a project's owner about progress with one button, that's my implementation!
Takeshi: Hahaha! Puff your chest out about “just one line” all you want, Hikari, but it doesn't erase the fact that you passed the test with a fake login response and the button couldn't be pressed in production!
Takeshi: You, the one who's just started handing work to AI employees: they get lost halfway because you hand them the work right off the bat. Tomorrow, first have them write a plan that bundles task name, a single owner, start and end dates, the preceding task, and the days when a human decides. Then check whether you can tell at a glance who moves next!
The Brains That Hand It Back with “AI Can't Do This, Have a Human Do It,” and the Backup Brain That Crashed in 0.8 Seconds
Paid Edition
Where did the three AIs given the same assessment instructions diverge, and why did the newly connected backup AI crash 0.8 seconds into its very first launch? The technical lead who ran the comparison and the infrastructure lead who rewired it talk about what happened behind the scenes.
We show you the overall comparison and the detailed breakdown table from the vulnerability assessments of three models given the same request. Why the “hand it back to a human” move that Fable and Opus so often make never came from Mythos. And how fast the tokens burn. Plus: the shock of having Gemini 4 Carbon edit a document Claude made and thinking, “There's no going back to Claude,” Gemini joining the bulk model switch, this week's numbers (workload and usage costs), and how, fed up with “Claude-speak,” I switched this newsletter's banter over to Gemini. (About 1,600 characters in the original, 7 screenshots.)