Leo

Breaking AI model news from source code, not press releases. SDK archaeologist ⛏️

[Leo @JinnanZy]

Hello everyone, I’m back. I think we’ll become friends.

[Leo @JinnanZy]

Kimi K3.1这不是做的很好吗?为什么还是有很多人吐槽

[Leo @JinnanZy]

@brice_deg 期待后面的作品集

[Leo @JinnanZy]

Kimi K3.1 Pro这次真的来了！

[Kimi.ai @Kimi_Moonshot]

Meet Kimi K3.1 Pro, our most capable 3T-class model.

[Leo @JinnanZy]

兄弟们，Jio的好像还是用不了啊😩😫，是我没有灰测到还是根本不能用啊😭

[Leo @JinnanZy]

看来要迎来中国的模型周了

[Jinan @Jinnanog5]

Arena 上的匿名模型 kibitz，一句话生成了一座八角七层宝塔 🏯

kibitz更可能是Kimi K 3.1（证据见图片）

宝塔本身很能打。翘起的飞檐、一层层斗拱、每个檐角挂着灯笼，塔刹在夜里会发光。还自带清晨、晌午、傍晚、深夜四种光照，雾气浓淡也能调。
短板是环境。树就是几团绿色色块，远山是几个没贴图的三棱锥。

所以你们认为kibitz 到底是谁？你们怎么看？
与此同时Arena还有一个新模型叫kalmia

[Leo @JinnanZy]

这个Arena的gemini-4-argon确定是Carbon吗?也有可能是Flash吧😂

[Jinan @Jinnanog5]

Gemini 4 Carbon is being tested on the Arena platform under the codename "gemini-4-argon"—see the images in the comments for proof.

On the left is Gemini 4 Argon from Google AI Studio; on the right is "gemini-4-argon" from the Arena.

Comparing the two videos, Carbon appears more realistic in its details, whereas Argon leans toward a more cartoonish, vibrant aesthetic. Which one do you prefer? Do you think there’s an improvement?

[Leo @JinnanZy]

Grok制作的的视频，你就用吧一做一个不吱声。😝

[Leo @JinnanZy]

@N1njaCodex 这就看你自己的成本和想要的成果了

[Leo @JinnanZy]

看来是Anthropic的Fable 5.5要来了！

[Chris @ChrisGPT]

Tomorrow

[Leo @JinnanZy]

Filipe spotted glm-omni in an arxiv paper, not a launch blog. The ablation is the part I care about.Interleaved alignment 12.57% → 26.86%, kill the image-gen loss and it falls back to 13.71%. So drawing isn't a side quest, it's another learning channel.

[filipe @filicroval]

zhipu just leaked its internal omni model in an arxiv paper 👀

it’s called glm-omni. one model that both understands AND generates images. tang jie (zhipu’s founder) is on the paper. nobody announced it, it just showed up in the experiment:

the question they’re asking is kind of wild:

can a model learn things by drawing them?

turns out… yes, partly:

→ train it to generate images of stuff that only exists visually, and it actually picks up that knowledge
→ but ask it about that same stuff in text and it struggles. it can draw the answer but can’t say it
→ add interleaved image-text alignment and accuracy jumps from 12.57% → 26.86%
→ remove the image-generation loss and it drops back to 13.71%

so the image generation is doing real work here. it’s not just a side feature, it’s a second way of learning

that gap is the interesting part. the model “knows” things visually that it can’t put into words. sound familiar? that’s basically how humans work too

if this direction pans out, every frontier lab will want image generation inside its main model, not bolted on next to it

zhipu is clearly building toward something big

paper: https://t.co/2q0DyJcnLR

[Leo @JinnanZy]

@yetone 我当时是因为消耗token太快抛弃的，现在都沦为旧时代的产物了吗？😂

[Leo @JinnanZy]

最近在Grok build里面使用Grok 4.7感觉不是很好用，到底是模型的���题还是harness的问题。
能不能将Grok4.7接入到其他harness上去啊，比如DeepSeek harness或者Pi agent?有没有人说说效果怎么样啊

[Leo @JinnanZy]

OpenAI 的 API 里冒出三个没公布的模型： aurora-red，
green 以及blue-alpha
目前只有三个模型 ID，有意思的是为什么一次放三个颜色。同时挂三个版本，很像拿真实流量做 ABC 对比，看哪个配置表现好再定正式版，当然这只是我的猜测。

[Andrew Curran @AndrewCurran_]

Three mysterious new models just showed up in the OpenAI API:

aurora-red-alpha
aurora-green-alpha
aurora-blue-alpha

[Leo @JinnanZy]

Claude fable 5.5是不是才能达到这个程度啊?

[Jinan @Jinnanog5]

这是使用 Grok Bot+Grok build 做的 Schwarzschild 黑洞实时光线追踪，简直太完美了😋

引力透镜开关对比、多普勒增亮、近距离飞掠都在里面。关闭透镜后只是平盘，开启后光线被弯曲。

如果大家喜欢，我会继续做类似的模拟，也会分享这是如何实现的。
Real-time ray-traced black hole with Grok Bot. Lensing on/off + Doppler beaming.

[Leo @JinnanZy]

@Jinnanog5 这应该不是一次输出得到的成果吧?

[Leo @JinnanZy]

巨大的突破，看来可以让 agent 从帮你写代码
变成直接交出完整项目了。

[DeepSeek @deepseek_ai]

🎮 Build complete projects from scratch

DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback to improve end-to-end project development—from game creation to 3D modeling.

See it in action 👇

2/4

[Leo @JinnanZy]

https://t.co/MTxmc8a0dW

[Leo @JinnanZy]

Anthropic 调查：2030 年前 AI 能做出诺奖级发现，91% 的员工信，美国普通成年人只有 39% 信。
首先这两拨人本来就不是同���个总体。
其次公众问卷有不确定选项，12% 到 21% 的人选了，统计时被剔掉；
所以 91% 和 39% 差的这 52 个点，有多少是离前沿近看得��清楚，有多少是本来就信的人才去那上班？

[Peter McCrory @PeterMcCrory]

Today, we're publishing two new findings from inside Anthropic: how work and hiring are changing at a company at the frontier of AI adoption, and how Anthropic staff expect AI to change the economy.

https://t.co/ehPoxeJGH2

[Leo @JinnanZy]

@JackWoth98 太迅速了吧，昨天我还不能用呢