AI-pilled / Building https://xitter.cf/t.co/3j8glg8Mq4 / Cloud architect / Kubestronaut

Joined February 2026
OddSVG oddsvg.com My tests ask models to generate something they almost certainly never saw in training. - Dishwashing Raccoon - Handstanding Giraffe - Ironing Platypus - Jump-roping Octopus - Necktie-tying Jellyfish Because of that, memorized output won't carry them — a model has to actually combine 3D shading, SMIL animation, and plausible anatomy on the spot. The dishwashing raccoon is the easiest of the three, but it still exposes plenty: how the kitchen reads, the shape of the hands, the running water and the suds. The handstanding giraffe raises the bar, since the long neck has to go down while the body holds its balance. The ironing platypus is the hardest. The platypus itself is difficult to draw correctly, and on top of that you get to watch how the model handles the iron and the ironing board — and the contact between them. Jump-roping Octopus: The model must accurately render all 8 tentacles, coordinating 2 tentacles to swing the rope while the remaining 6 handle the jump and landing. It also tests whether the model avoids spatial errors, such as the rope clipping through the octopus's body. Jellyfish Tying a Necktie in Front of a Mirror: This tests how the model creatively handles putting a necktie on a neckless creature, along with its ability to correctly render mirror reflections and spatial perspective. Weaker models fail most of these outright. Among the stronger ones, the differences show up in the details, which is where you start to see each model's temperament.
1
1
1
274
Kimi K3는 최소한 나한테는 별로였어... 나는 Qwen3.8 Max가 가격 대비 제일 나은 것 같아.
Pi + Opus 5 (OpenRouter) Pi + Qwen3.8 Max (OpenRouter) 아무래도 구독이 필요할 것 같아. Pi + Kimi K3 (Ollama Cloud) K3는 많이 안써봤는데 써보고 후기 올릴게.
1
66
OddSVG oddsvg.com My tests ask models to generate something they almost certainly never saw in training. - Dishwashing Raccoon - Handstanding Giraffe - Ironing Platypus - Jump-roping Octopus - Necktie-tying Jellyfish Because of that, memorized output won't carry them — a model has to actually combine 3D shading, SMIL animation, and plausible anatomy on the spot. The dishwashing raccoon is the easiest of the three, but it still exposes plenty: how the kitchen reads, the shape of the hands, the running water and the suds. The handstanding giraffe raises the bar, since the long neck has to go down while the body holds its balance. The ironing platypus is the hardest. The platypus itself is difficult to draw correctly, and on top of that you get to watch how the model handles the iron and the ironing board — and the contact between them. Jump-roping Octopus: The model must accurately render all 8 tentacles, coordinating 2 tentacles to swing the rope while the remaining 6 handle the jump and landing. It also tests whether the model avoids spatial errors, such as the rope clipping through the octopus's body. Jellyfish Tying a Necktie in Front of a Mirror: This tests how the model creatively handles putting a necktie on a neckless creature, along with its ability to correctly render mirror reflections and spatial perspective. Weaker models fail most of these outright. Among the stronger ones, the differences show up in the details, which is where you start to see each model's temperament.
1
1
1
274
그리고 Ox Alpha는 전혀 프론티어 모델이 아님.
Ox Alpha가 며칠 전보다 더 뛰어나졌다는 글을 봤어. 그래서 직접 테스트 해봤지. 전혀 그런 건 없었어. 오히려 테스트 결과가 안좋아. 며칠 전보다 뛰어나진게 아니라 그냥 여러번 테스트 하다가 우연히 결과가 더 좋게 나왔던거야.
1
35
Ox Alpha가 며칠 전보다 더 뛰어나졌다는 글을 봤어. 그래서 직접 테스트 해봤지. 전혀 그런 건 없었어. 오히려 테스트 결과가 안좋아. 며칠 전보다 뛰어나진게 아니라 그냥 여러번 테스트 하다가 우연히 결과가 더 좋게 나왔던거야.
1
1
64
Qwen3.8 Max는 내 테스트에서 좋은 모습을 보이고 있어. 왜 이 모델은 인기가 없을까?
현 시점 내가 제일 좋아하는 모델 3가지
1
34
Pi + Opus 5 (OpenRouter) Pi + Qwen3.8 Max (OpenRouter) 아무래도 구독이 필요할 것 같아. Pi + Kimi K3 (Ollama Cloud) K3는 많이 안써봤는데 써보고 후기 올릴게.
1
140
Tailscale로 사용하는 개인 웹앱들. 프록시 하나 두고 라우팅하게 만들고 앱 아이콘 같이 보여주는 메인 페이지 만들기. 아이콘은 ipaslogo.com/ 링크 주고 알아서 설정해달라고 하면 됨.
28
Kimi K3와 GLzm 5.3을 쓰려면 @ollama 클라우드가 괜찮은 안이라고 생각한다. 중국이 아닌 미국에서 서빙하는 덕분에 훨씬 쾌적함.
14
난 진작 뮤트했는데 이제 차단했어.
Replying to @ns123abc
>built a tool to exploit a goodwill feature >insulted everyone who warned her >asked openai to ban the people who used it incredible
17
회사에서는 Fable 5 집에서는 Qwen 3.8 Max
모델이 많이 나오지만 결국 사용하는 모델은 믿을 수 있는 한두가지 모델 뿐...
22
모델이 많이 나오지만 결국 사용하는 모델은 믿을 수 있는 한두가지 모델 뿐...
1
36
AI 관련해서 내 X 타임라인만 보면, 중국은 AI와 로보틱스에서 놀라운 발전을 보여주고 있고, 미국은 한심스러운 짓거리만 하고 있어. 앞으로 어떻게 될까?
9
제 테스트를 완벽하게 통과한 모델은 현재 없습니다.
OddSVG oddsvg.com My tests ask models to generate something they almost certainly never saw in training. - Dishwashing Raccoon - Handstanding Giraffe - Ironing Platypus - Jump-roping Octopus - Necktie-tying Jellyfish Because of that, memorized output won't carry them — a model has to actually combine 3D shading, SMIL animation, and plausible anatomy on the spot. The dishwashing raccoon is the easiest of the three, but it still exposes plenty: how the kitchen reads, the shape of the hands, the running water and the suds. The handstanding giraffe raises the bar, since the long neck has to go down while the body holds its balance. The ironing platypus is the hardest. The platypus itself is difficult to draw correctly, and on top of that you get to watch how the model handles the iron and the ironing board — and the contact between them. Jump-roping Octopus: The model must accurately render all 8 tentacles, coordinating 2 tentacles to swing the rope while the remaining 6 handle the jump and landing. It also tests whether the model avoids spatial errors, such as the rope clipping through the octopus's body. Jellyfish Tying a Necktie in Front of a Mirror: This tests how the model creatively handles putting a necktie on a neckless creature, along with its ability to correctly render mirror reflections and spatial perspective. Weaker models fail most of these outright. Among the stronger ones, the differences show up in the details, which is where you start to see each model's temperament.
20
Ox Alpha가 어떤 모델인지 빨리 밝혀지면 좋겠어. 이 모델이 뭔지 확신에 차서 말하는 놈들 모두 입 닥칠 수 있게.
12
계속 얘기하지만, 모델이 문제가 아니라 하네스가 문제입니다. 클로드 코드 대신에 Pi + Opus 5를 만족스럽게 사용 중이에요.
if fable felt dumber this week, it's not you ❗❗❗ since 2.1.237 the model reads "high" effort as 10 out of 100, the exact number "low" used to be and the changelog doesn't say a word i spent my whole afternoon convinced t3 code and my own app were broken before i went digging through the actual requests anyone else notice or am i going insane?
1
90
Ox Alpha는 괜찮은 성능에 속도를 가지고 있어. 게다가 지금은 무료로 사용할 수 있지. 하지만 최상위 모델은 아니야.
27
내가 전부터 얘기했지만 Opus 5는 모델이 안좋은게 아니라 Claude Code 때문이야. Pi로 Opus 5 사용 중인데 만족스럽게 쓰고 있거든.
確定情報ではないのだけど、Opusの喋り方がどんどん変になっていった原因の一端が分かったかもしれない Claude codeはメモリ(Memory.md)にいろんなことを書いていって、それがセッション開始時に読まれるわけなのだけど、その文体が基本的に圧縮言語になりがちなんだ
2
329
LLM 테스트는 모델이 수행하기 충분히 어려워서 모델마다 결과물에 차이가 있어야 함.
OddSVG oddsvg.com My tests ask models to generate something they almost certainly never saw in training. - Dishwashing Raccoon - Handstanding Giraffe - Ironing Platypus - Jump-roping Octopus - Necktie-tying Jellyfish Because of that, memorized output won't carry them — a model has to actually combine 3D shading, SMIL animation, and plausible anatomy on the spot. The dishwashing raccoon is the easiest of the three, but it still exposes plenty: how the kitchen reads, the shape of the hands, the running water and the suds. The handstanding giraffe raises the bar, since the long neck has to go down while the body holds its balance. The ironing platypus is the hardest. The platypus itself is difficult to draw correctly, and on top of that you get to watch how the model handles the iron and the ironing board — and the contact between them. Jump-roping Octopus: The model must accurately render all 8 tentacles, coordinating 2 tentacles to swing the rope while the remaining 6 handle the jump and landing. It also tests whether the model avoids spatial errors, such as the rope clipping through the octopus's body. Jellyfish Tying a Necktie in Front of a Mirror: This tests how the model creatively handles putting a necktie on a neckless creature, along with its ability to correctly render mirror reflections and spatial perspective. Weaker models fail most of these outright. Among the stronger ones, the differences show up in the details, which is where you start to see each model's temperament.
14
누가 만든지는 상관 없고 무료일 때 많이 쓰자!
Ox Alpha는 아마도 GLM 5.3 모델을 기반으로 사후 훈련시킨 모델이 아닐까? 그리고 이렇게 무료로 제공할 수 있는 걸 보면 서양 쪽 모델일 것 같아.
28