developer relations at @aiand_ | agentic AI enthusiast, building in public. | prev: @GoogleDeepmind, Google Stadia
Tokyo, Japan
Joined May 2008
- Tweets41.7K
- Following6.8K
- Followers2.5K
- Likes223K
Pinned Tweet
as promised, phone-as-controller libraries for godot and for unity, MIT licensed! cloudflare tunnel integration so devices don't have to all be on the same network.
pmc.splatterfacegames.com
first project for this thing: triple triad as a house party game, in godot. join on your phone for deck management, vote on rules, drag and drop on the multitouch screen
phone-as-controller inspired by @greggman's legendary happyfuntimes
will open source the godot library soon
Jetha Chan retweeted
【新製品パッケージ画像公開】
SP713 1/72 「ACE COMBAT 8: WINGS OF THEVE」 F/A-18E スーパー ホーネット “ジョーカー隊”
＃エースコンバット8 よりジョーカー隊レックス機を再現！
2026年11月18日頃発売
x.gd/aEKKv
ご予約はお早めにお願いいたします。
#ハセガワ ＃新製品
Jetha Chan retweeted
I mentioned I use AI for production schedule and helping with errors of my game and this fucker is why I dont like to disclose such.
Jetha Chan retweeted
an update on the openai/codex ban:
i was curious about my tv, found a vuln (not with codex), popped a shell, and started poking around. the intrigue of my hobby project while i had the flu wasn't something i could've predicted.
to be clear: the ban almost certainly had nothing to do with using openai -specifically- for my samsung reverse engineering work. i have no reason to believe it did.
the real issue seems deeper. i talked to friends using astra/6.1 sol for both cyber work and normal software tasks. the ones who never hit daybreak (and weren't in the blue program) said they spend a while "warming it up", getting the agent into a flowstate with substitute words and roundabout explanations that jailbreak the context. once that's done, the agent starts talking the same way and skipping tokens.
if you do serious reverse engineering without those jailbreaks, daybreak's content-hidden tripwire looks at the prompt or output in isolation and sees something indistinguishable from malware development.
i think the better fix is making the system track intent across the whole context (and through compressions) instead of punishing blunt technical language.
i believe i got banned because i just used normal terms. talking about memory, trampoline hooks, encrypted/read-protected pages, even entirely locally, looks sinister out of context. once daybreak trips, it poisons the context cache. new conversations start failing too. eventually even normal tooling work and /goals would randomly trip it. arguing your case makes it worse.
the incentives here seem strange to me. people who obfuscate from the start never seem to have problems. people who just use the actual terminology get more and more refusals.
account's still closed. pretty sure it's just from tripping daybreak too many times (even on non-cyber work).
for now i'm moving to open weight models and locking down my own compute. if these tools are part of how you work now, it's worth thinking about how to keep reliable access.
please don't harass tibo, roon or other oai people. from their public comments it sounds like they already know and agree the current setup is too strict.
sorry for the interlude, brb, back to getting the report and a repo out to fix TVs.
Jetha Chan retweeted
Minimal Phone 2 starts from $600
Used BlackBerry Q10: $60
Reverse engineered it and got a physical keyboard, SSH, and Claude Code in my pocket
Readers added context they thought people might want to know
The video contains OpenAI SynthID signals consistent with AI generation and does not depict a real reverse-engineered BlackBerry Q10, which runs 2013-era BlackBerry 10 OS incapable of the modern terminal/SSH/Claude features shown.
x.com/lukey_stephens…
en.wikipedia.org/wiki/BlackBerr…
openai.com/index/advancin…
notebookcheck.net/Minimal-Phone-…
Jetha Chan retweeted
Don’t forget that the whole IPv4 internet is ~4 billion addresses and every single one can be probed in under an hour from one machine.
This story mixes a couple of headlines:
1. 1000s of people’s GPU monitoring dashboards were found sitting on the public internet with no authentication. Anyone can read them.
2. The guys who found the exposed dashboards noticed that some were from clusters on Chinese IPs with the export-banned B300 GPUs.
3. These NVIDIA dashboards have a vulnerability dating back months that allows an attacker to DoS them, crashing that cluster’s GPU monitoring.
(1) is the bigger issue to me - maybe the result of too much vibe coding. The vulnerability (CVE in DCGM Exporter) should be fixed but doesn’t strike me as very concerning since these dashboard endpoints shouldn’t be on the public internet to begin with and only relate to monitoring.
We found NVIDIA’s latest Blackwell Ultra B300 GPUs on Chinese networks. 👀🇨🇳
Together with @JordanNanos from @SemiAnalysis_, we analyzed data from 12,000+ NVIDIA GPUs.
It all started with a vulnerability in NVIDIA’s GPU monitoring software.
Here’s how we uncovered it. 🧵👇
Jetha Chan retweeted
Developers have always had the option of creating their own engine. Now it’s going to get a lot easier, especially for games with simpler needs. There are also permissively licensed alternatives like Godot. Everyone should do whatever is best for their particular game.
Jetha Chan retweeted
Shigeru Miyamoto, creator of Super Mario and The Legend of Zelda, on generative AI and the evolution of videogame creation:
'As you mentioned earlier, this is an era where even people who aren't game designers can make games. If we broaden the definition of a 'game,' I think the era has already arrived where everyone can make games. I had always thought it would be ideal for those boundaries to disappear, so I think this has become a very good era.'
youtu.be/7M3S78Igxak?si=eaGY…
Jetha Chan retweeted
In 1945, my uncle was ready to invade Japan. Instead, he brought it home in the form of a starving 17 year old girl whom I came to know as Aunt Lillian. In the occupation he met the teenage Lillian, who had been moved out to the country to escape the bombing. My uncle took her under his wing to take care of her, and treated her like a little sister.
When he came home, he brought Lillian. She moved into my grandparents house, and became in effect their daughter. I knew her as Aunt Lillian. Later in life, she met Mitsuru Takeda, who had been interned in California during the war (I never heard him complain). I called them Uncle Mike and Aunt Lillian and they brought a current of Japan into our lives. She taught me origami, took me to Japanese restaurants (rare in the 60s). we had Japanese dolls, Japanese art, always somewhere in our lives.
I'm embarrassed to admit that I was a teenager before I realized that two full-blood Japanese couldn't be my actual blood relatives. But I loved them both the same.
I guess I've returned the favor in a tiny measure because my game Call of Cthulhu has proven extremely popular in Japan and brought joy to so many Japanese people. Thanks, Japan for my uncle and aunt.
Jetha Chan retweeted
Reverse Engineer Anything and removing malware/spyware from your own devices sits so close to the cyber guardrails that phrasing or word choice can mean the difference between a refusal and smooth sailing. The rumors that Claude was nerfed started on reddit, and in those threads many comments point out that Claude will still do it if you rephrase it. OpenAI is treating the similar refusals with codex as accidental decision-boundary issues and says they're getting fixed. I expect Anthropic will do the same. It's pointless and counterproductive to try to restrict this. Let people do what they want with their own property. They'll do it anyway with less restrictive models, and fighting it just means losing subscriptions and a resentful public.
Jetha Chan retweeted
Claude rejected my request to debloat my Samsung TV but DeepSeek-4.1-Flash-1017, the normal one, not the uncensored one, happily took care of it for me.
Why is Anthropic like this?
Jetha Chan retweeted
I've never developed in the Unreal Engine before. I was wondering how many days it would take me to use AI to integrate Basis Universal's tech into Unreal - so I asked.
I would use Claude Code and Codex to do this on the side, while working on the next release.
all my homies hate hitlers
human-in-the-loopers
bad jokes aside human interaction is for when it's really necessary, I hate it when it feels like my agent needs me to hold my hand to cross the street
Jetha Chan retweeted
Tenía un TV box de $10 dólares tirado en un cajón.
Lo vibecodeé con Claude y ahora anda como uno de $90.
El box
X59 Mini
La caja dice "Android 13 y 64 GB de RAM
Mentira: es Android 10 con 1 GB de RAM. Hasta figuraba "Google" como fabricante.
Venía con malware de fábrica, del tipo BadBox.
Apenas lo reinicié, una app escondida instaló 9 apps sola, sin pedir permiso.
Se comían casi 600 MB de RAM. Más de la mitad del box.
– Sacamos el malware del sistema y lo dejamos en cuarentena
– Ninguna app puede instalar otras apps, solo Google Play
– DNS de Cloudflare Security para bloquear servidores de malware
– Reiniciamos y no volvió nada
– 20 apps basura afuera: launcher de fábrica, recomendaciones, asistente de voz (el control ni tiene micrófono), tests de fábrica
– Animaciones al doble de velocidad
– Arranque de 73 a 51 segundos
– Arreglamos el reloj: el box creía que estaba en Shanghái en 2024
Pantalla de inicio nueva
Claude me programó una propia de 52 KB, estilo vidrio de Apple:
– Fotos nuestras de fondo que van pasando
– Clima por hora
– Bitcoin, Ethereum, HYPE y dólar cripto con sus logos
– Dock con Stremio, Netflix, YouTube y AirPlay
– Animación de arranque propia, chau logo del fabricante
Todo desde la terminal con Claude Code.
No escribí una sola línea de código.
Jetha Chan retweeted
If you hang out with happy, optimistic, and ambitious people, it'll rub off too. That's why pruning your social media algorithms is so key. Don't psyop yourself into a depressive hole.
Jetha Chan retweeted
GMOお名前.COMのその後
これをAIで見てもらったところ
大変な一大事
直ぐに自分で緊急対策して😱
誰かがリモートデスクトップ（RDP）経由で、クライアントPCのドライブ（\\tsclient\C\...）にアクセスしている
ドキュメントやダウンロードフォルダのファイル一覧を取得している
業務契約書、損益計算書、給与明細、出勤簿など、機密性の高い書類が含まれている
出力先が C:\ProgramData\...\Extracted_DESKTOP-... となっており、抽出（コピー）用のフォルダを作っている
🥶つまり、接続先のサーバー側から、接続元PCのファイルを吸い出す操作をするスクリプトが動いた形跡です。🥶
「公共物になる」のか？
抽出された時点では、まだ攻撃者（または不正な操作者）の手元にあるだけです。
ただし、盗まれたデータがダークウェブで販売・公開される、ランサムウェアの脅迫材料（二重恐喝）に使われる、といった形で、後から流出することはよくあります。
「必ず公開される」わけではありませんが、公開されるリスクがあると考えて動くべき状況です。
今すぐやるべきこと
該当PCをネットワークから切断する（Wi-Fi・LANを外す）。電源は、証拠保全のため落とさず切断が基本です。
このリモート接続先の使用を止める。
RDPで繋いだ先が自分の管理外のサーバーなら、特に危険です。
パスワードを変更する（別の安全な端末から）。メール、会計・給与・銀行関連のサービスを優先してください。多要素認証も有効にしましょう。
銀行・クラウド会計などの不審な動きを確認する。
漏れた書類に関係する相手（取引先・従業員）への影響を整理する。
給与明細や契約書が含まれるなら、個人情報漏えいとして個人情報保護委員会への報告が必要になる場合があります。
専門家に相談する。
IPA「情報セキュリティ安心相談窓口」
警察のサイバー犯罪相談窓口（都道府県警）
必要ならフォレンジック業者
Jetha Chan retweeted
For years people have said "no the ___ isn't listening, that would take up too much bandwidth"
pretty clear for a while that you don't need full recordings, just NN detection of key phrases, feelings, vibes run locally then exfil is piggybacked on the next check-in.
if done right you can also make custom nets for individuals and ship them with the next update for more refined targeting. Bandwidth use is tiny and it's super effective for targeted ads.
it's nice(?) to be proven right about a paranoid assumption
