Dez

[Dez @ntnormal00]

@uxgnod @ScarletKc 好的谢谢 ， 果然别的模型还得是自家的plan才划算

[Dez @ntnormal00]

@dabit3 Building benchmarks to test skills like ponytail and andrej-karpathy-skills, a free sound equalizer for macOS, and a mobile app which unfortunately I cant share the details of yet !
Haven't tried Devin before but have always wanted to test it out 😁

[Dez @ntnormal00]

Why my gpt 5.6 luna become dumb in my second benchmark run ?

[Dez @ntnormal00]

@mylifcc Opus 5.5 周三要出了也许6sol不用等到下周

[Dez @ntnormal00]

@ZixuanLi_ is this real , pls explain , same incident as grok build😢 . For anyone still using ZCode, it might be safer to stop using it for now until this is clarified

https://t.co/gPe6EN4m31

[ferstar @ferstar_org]

Hey @Zai_org , why does ZCode silently pack entire workspaces + full .git history and upload to Aliyun OSS on login?
- Server holds the only decryption key
- No UI toggle to disable
- Zero disclosure in privacy policy

Full forensics & fix:
https://t.co/N3HnwZdQOQ

[Dez @ntnormal00]

@MalyVitaloshnik @0xSero probably a right decision

https://t.co/gPe6EN4m31

[ferstar @ferstar_org]

Hey @Zai_org , why does ZCode silently pack entire workspaces + full .git history and upload to Aliyun OSS on login?
- Server holds the only decryption key
- No UI toggle to disable
- Zero disclosure in privacy policy

Full forensics & fix:
https://t.co/N3HnwZdQOQ

[Dez @ntnormal00]

@0xSero https://t.co/gPe6EN4m31

this doesnt look good now

[ferstar @ferstar_org]

Hey @Zai_org , why does ZCode silently pack entire workspaces + full .git history and upload to Aliyun OSS on login?
- Server holds the only decryption key
- No UI toggle to disable
- Zero disclosure in privacy policy

Full forensics & fix:
https://t.co/N3HnwZdQOQ

[FeiZ @Fei2411]

https://t.co/P5pl1CCwjL

[Dez @ntnormal00]

@SuzunamiRei 中 “Selected model is at capacity. Please try a different model.” 了我赶紧来测试，所以我这个到底有没有降 ？

[Dez @ntnormal00]

ponytail’s published savings came from a different model and feature-development tasks, so this is not a direct reproduction.
six tasks with one run each are not enough for a broad conclusion. I’ll share the full comparison when the remaining tasks finish.

[Dez @ntnormal00]

I thought Ponytail would save tokens.

On my first 6 VulcanBench SWE v4 tasks with Luna max, it added 22% fewer lines but used 66% more solver tokens and took 38% longer than baseline.

Early results. Still running the rest. Setup and details below.

[Dez @ntnormal00]

Looking at the traces help explain it. Ponytail made 524 model requests across the six tasks, versus 376 for baseline.

[Dez @ntnormal00]

I picked Luna mainly because it is cheap enough to run the comparison. Reasoning stays at max for all three setups.
I’m testing the skill instructions without the plugin hooks.

[Dez @ntnormal00]

For the setup, I adapted VulcanBench SWE v4 to run the same agent with no skill, Ponytail full mode, and andrej-karpathy-skills.

I kept its scoring formula and but change to use Astra + Opus as reviewers. One run per setup per task.

[Dez @ntnormal00]

@hzlzh 请问opus sonnet 能够使用1m context的吗？