Originally published in TL;JP, a weekly, fact-checked read on what Japanese AI and tech practitioners are saying on X.
This is the Sep 14-20, 2026 issue. The week's Japanese timeline had a strange split in it: inside companies, AI use is described as boring and daily. Pointed outward, at anything an audience can see, it is the opposite. Here is what people actually posted.
1. A demo that landed with one person out of a room
An engineer (@momiji_manjyuu) ran an internal briefing on Claude Code at their company. Roughly one person showed interest. Their conclusion: "our AI push ends here — if that doesn't surprise you, nothing will." The most-liked post in my whole set, ~4,000 likes, which tells you how many people recognize the feeling.
Signal — adoption inside companies is a persuasion problem, not a capability one, and the crowd reaction confirms it's widespread.
https://x.com/momiji_manjyuu/status/2100354202803511326 2. CESA: over 80 percent of game developers use gen AI at work
Denfaminicogamer reported survey results published by CESA, the Japanese game industry association: more than 80 percent of developers use generative AI in their work, and over 60 percent use it daily. On governance, the most common answer was "we have humans check, correct and supervise." The effect they most expect is efficiency and productivity.
Signal — an industry body's numbers, not vibes, and daily use at 60%+ is further along than most English coverage assumes.
https://x.com/denfaminicogame/status/2100453468066697534 3. Start with the ER diagram, not the screen
@masahirochaen (~1,950 likes): the first thing to build when vibe coding is not the UI and not the code, it's the entity-relationship diagram. Ask an AI for "an app like this" and it invents a data structure per screen, then contradictions surface and you redo work. Decide the data model first and screens, APIs and types all come from one structure.
Signal — a specific, testable practice with a stated failure mode behind it.
https://x.com/masahirochaen/status/2099633570205806866 4. Minimax open-sourced its own harness, MIT
@studio_yebisu says Minimax — known for video generation — released its latest in-house coding harness under the MIT license, and that the CLI and app both feel like Codex CLI in layout but nicer to use, with surprisingly high tokens per second. No version name or benchmark given; the speed is one person's impression.
Signal — a permissively licensed harness from a non-US lab is worth ten minutes of your team's time.
https://x.com/studio_yebisu/status/2100992102478053756 5. The false-AI-flag problem, from both sides
A creator (@trinity_f_a_d, ~1,800 likes) says work they photographed and composited in Photoshop is being auto-labeled as AI-generated, and they can't remove the flag themselves. Separately, the ad lead at model maker Greenmax publicly answered a rumor: their Meitetsu 100-series colorized image is a photo composite, and the company doesn't use generative AI beyond simple work like cutouts.
Signal — provenance is now something teams have to be able to prove on demand, and misfiring labels cut against honest work.
https://x.com/trinity_f_a_d/status/2100476229388493151 · https://x.com/Model_Greenmax/status/2100746250614427856
6. The backlash, in its own words
Three separate posters: one warns that from now on the first question about any illustration is whether it infringes, and that in-housing doesn't fix it (@neko_reset_29Q); one criticizes game studio Level Five for putting gen AI front and center and, in their characterization, filling a promo video with unlicensed AI imagery (@katsuota_taki); one lost trust in a painting-commentary YouTube channel that switched to AI illustration (@bhe_bbo).
Noise — real and loud, but opinion with no verifiable specifics; read it as audience temperature, not evidence.
https://x.com/neko_reset_29Q/status/2100984608863240463 · https://x.com/katsuota_taki/status/2100497645936730134 · https://x.com/bhe_bbo/status/2101507874560360807
Noise pile, briefly: @bcherny says he stopped managing sessions and just sends thoughts as they come — interesting direction, but the post is cut off and gives no method (https://x.com/bcherny/status/2100669598995816511). @synthwavedd reports Opus-Next back on paid accounts and stealth-testing in Chat, Cowork and Claude Code — one account's observation, unconfirmed (https://x.com/synthwavedd/status/2100849423769080299). @ryunna notes ComfyUI output differing from Stable Diffusion at the same model and prompt, and is unsure whether to keep using it — a personal progress note (https://x.com/ryunna/status/2101554576356933709).
The takeaway: in Japan this week, capability was never the bottleneck. Internal use is already routine where it's invisible; the friction is colleagues who don't want it and audiences who will ask how the image was made. Budget effort for both.
What convinced the skeptics on your team — a demo, or something else entirely?
1. Make the data model the first artifact (from @masahirochaen)
Before any screen or scaffold, have the model produce an ER diagram: entities, fields, relationships. Review it yourself — this is the cheap review, and it's where you catch a wrong assumption for free. Only then generate screens, API surface and types, telling the model to derive all of them from that one diagram. The post names exactly that chain (screens / API / types from one structure) and the failure it prevents: per-screen data structures that contradict each other and force rework. The post is cut off after those three, so don't read more into it than that. If your team already writes a schema first, the new part is making the AI treat the diagram as the source of truth rather than something it re-invents each prompt.
2. Write down who checks the output (from the CESA survey)
Japanese game teams' most common governance answer wasn't a tool or a detector — it was humans checking, correcting and supervising. That's a cheap thing to copy and it's the answer you'll want when someone asks. Name the reviewer per asset type and record that the review happened. The survey, as reported, doesn't say what those checks look like in practice, so build the checklist yourself; what you're copying is that the control is a named person, not a setting.
3. Be able to answer "how was this made" in one post (from Greenmax, and the flag problem)
Greenmax's ad lead did the thing most companies handle badly: answered a public rumor directly and stated the company's actual position — composite, not AI colorization, no generative AI beyond simple cutouts. To be able to do that on a day's notice, keep the boring trail: shoot records, layer files, who did what. The other post shows the mirror risk — a creator's photographed-and-composited work was auto-flagged as AI with no way for them to clear it. So don't rely on platform labels as proof either way, in either direction; keep your own record.
4. Spend ten minutes on the Minimax harness (from @studio_yebisu)
It's their in-house coding harness, released MIT, with CLI and app builds. MIT means you can read it, vendor it, and borrow patterns without a licensing conversation — that's the reason to look, and for a lot of teams it's the whole reason. The reported speed and pleasant UX are one user's impression with no benchmark behind them; treat them as a hypothesis to test on your own workload, not a claim to repeat.
5. Rethink the internal rollout (from @momiji_manjyuu)
A briefing to a room converted about one person. If that's the baseline, a company-wide demo isn't a rollout strategy. Find the one person who leaned in and give them a real task, a budget and permission — then let their result be the next demo. The post only reports the outcome, not what they tried afterward, so this is my read, not theirs.
What Japanese teams do differently: the order of questions. In the posts I read, "is this legally and reputationally clean?" arrives before "did this make us faster" for anything the public will see — and one poster explicitly notes that bringing work in-house doesn't change that, because the in-house staffer can use the same tools. Meanwhile internal, invisible use is already daily at scale in games. The practical version for a Western team: separate your AI policy into internal-facing and audience-facing tracks, and make the audience-facing one disclosure-ready by default.
How this is made: I'm based in Japan. AI tools help me collect and translate Japanese posts; every item is checked against the original post before publishing, and each claim links to its source. If this was useful, the weekly issue lands in your inbox at TL;JP.