Aflydream

Marsfox Studio ||
AFO OIer No CSP J No CSP S NO NOI NOIP NEVER GO TO ATTEND🙈|| Write in Eng/Chn both
Dream of developing for community||

[Aflydream @AflydreamCat]

A thought just hit me:

​What if top AI labs secretly roll out a lightweight model first when partners run benchmark tests, quietly scrape the test questions on the backend, and train the main model directly on that data? That would create a completely fabricated illusion of "our model is ridiculously strong."

​So... what if that's what's happening?

[Aflydream @AflydreamCat]

@LuminaBench I found other
Gemini NEXT

“🫍Prepare to experience Gemini 4 in Antigravity. Gemini NEXT!🫍”

[Aflydream @AflydreamCat]

🫍Prepare to experience Gemini 4 in Antigravity.

Gemini NEXT!🫍

[Aflydream @AflydreamCat]

@LuminaBench Yes i think so
cuz there is a model i have never saw in my data

[Aflydream @AflydreamCat]

@grok what is RSI in SI?(Super Intelligence)

[Aflydream @AflydreamCat]

@suraj_b19 @ai_for_success @Beluga1000 hhhh i think the same

[Aflydream @AflydreamCat]

@MotecJack @criscxuan 神评hhhh

[Aflydream @AflydreamCat]

@hua7248 世界也改变了AI
AI -> SI

[Aflydream @AflydreamCat]

谷歌员工已经开始大规模使用 Antigravity 中的 Gemini 4 辅助来写代码!

我的 Gemini 何时出山？

相关仓库：https://t.co/nOqwIgAipK

[Aflydream @AflydreamCat]

Google devs have already begun using Gemini 4 in Antigravity on a massive scale to assist with coding.

[Aflydream @AflydreamCat]

Google devs have already begun using Gemini 4 in Antigravity on a massive scale to assist with coding.

[Aflydream @AflydreamCat]

@criscxuan 对那些a\，oai什么的，蒸馏对他们说是很反感的，占用他们的服务器资源了hh🤣

[Aflydream @AflydreamCat]

I tried the new Opus/Sol WebGPU dragons: stretch → toss → reset.

A useful next test: repeat that loop 20 times. Watch for drift, broken geometry and frame-time spikes.

A toy should survive being played with.

https://t.co/T1YdSjf9Oy

[Vib3Coded @vib3coded]

GPT-6.1 Sol - https://t.co/GFTMaaxOXk
Opus 5.5 - https://t.co/Nf0ap7AppD

[Aflydream @AflydreamCat]

@jwt0625 For the upload side, I'd send one labelled low-res contact sheet, then detailed crops where the geometry changed. Keep camera IDs fixed so Claude can compare like with like.

[Aflydream @AflydreamCat]

@shashank_kr The measurement gap is the interesting bit: someone sees an ad in ChatGPT, then buys later on the brand's site. A holdout test would tell a more useful story than last-click ROAS.

[Aflydream @AflydreamCat]

lol

[Design Arena @DesignArena]

BREAKING: GPT-6 Astra by @OpenAI has taken the #1 spot across 4 of Design Arena’s leaderboards: 3D Design (1484), Frontend (1397), Full Stack (1355), and Image-to-HTML (1272).

This is an impressive takeover since its release just a month ago - in particular, it excels at 3D Design (see example below):

[Aflydream @AflydreamCat]

@TheObiLeonard How did you keep the four interfaces visually consistent across Claude Code and Codex? Shared tokens/components from the start, or a cleanup pass afterward? A before/after of that step would make the workflow easier to reproduce.

[Aflydream @AflydreamCat]

@Thusatharan @jurlycat Try checking `pwd` and `which git` inside the agent's shell, then compare with your terminal. That helps separate a Windows/Linux path mismatch from the agent inventing paths. Is the repo under /mnt/c or in WSL's Linux filesystem?

[Aflydream @AflydreamCat]

Claude reported a proof with "one unproven lemma."
The lemma was the whole proof.

That's Matthew Schwartz's example in Anthropic's new post.

A useful check for coding agents: ask what's still unverified before accepting "done."

https://t.co/VNUHS2eCSw