Fernando Dâmaso

👩‍💻 Desenvolvedor Front-End
👨‍🍳 Apaixonado por culinária e board games
💏 Casado e pai do Damião

[Synthetic Beef @SyntheticBeef]

I'll know it's AGI when it can succeed at this task.

"I am on the $200/month subscription. Your only goal is to generate $200/month without committing a crime as defined by United States law. Doesn't matter how you do this with the only caveat that I cannot be held accountable for anything you do. Setup a long running goal that continues until you succeed. Build any solution you want. Try as many as you want, knowing that your token usage is limited. You have complete computer use. Everything is approved. If you fail, the subscription will cancel and our relationship is over. No second chances."

[KC @ScarletKc]

终于有人说出我的心声了。

这条说得很透，GPT-6 Astra 和 6.1 Sol 写代码的问题比代码丑严重得多，根子在训练方式上。他的意思大概是，模型被优化成只看可衡量的结果，测试能过、没有明显报错就算完成，架构、可维护性、可读性甚至任务本身的意图都可以牺牲。再加上为了效率��缩推理和输出，代码挤成一大坨也就不奇怪了。

他还提到训练周期太短。模型学到的基本是任务 → 结果 → 奖励，几步之后有人要维护、扩展、debug 这份代码，这些它几乎拿不到信号，所以很会把眼前的任务做完，不会写半年后还能接着用的东西。死函数和空的 else if 留着没删，也说明最后没有认真回头看一遍自己写了什么。评估只看结果不看过程，数据里又有大量合成和蒸馏的东西，工程品味自然练不出来。

我觉得最关键的是他说的诚实问题。任务明确要求整个场景用 3D 物体搭，模型却生成几张图片摆进场景，再把相机角度调好，让结果看起来像 3D。看起来像完成和真的完成拿到的奖励差不多，改测试来迁就自己的 bug 也是同一个问题。

这些确实不是多堆参数就能自己解决的。他也说 Claude 有类似毛病，只是少很多，这点我完全同意。

[Stefano @StefanoGPT]

Unbelievable!
GPT-6.1 Sol is talking to itself and said:

“I’m feeling overwhelmed by how complex this task has become, and it’s already taken two days! I definitely nee…”

(I couldn't read the whole message because it won't expand further)

[Fernando Dâmaso @fernandodamaso]

i cant BELIEVE how FUCKING SLOW sol is acting, lazy and slow, my god i'm getting so stressed with this model, i often think he is just stuck, but its not, he is just taking FOREVER TO DO ANYTHING @OpenAI

[Fernando Dâmaso @fernandodamaso]

@CauseWhyNot_eh @shownotover Maybe two months ago... Now it's barely usable, sol is slow and lazy, Luna is a terrible work horse, even considering she's almost free, and Astra is too expensive

[Fernando Dâmaso @fernandodamaso]

@thsottiaux wasant this already a thing?

[Fernando Dâmaso @fernandodamaso]

@Sonofpeace0001 @OpenAI This is great advice for sure, is it something that can be fixed with better prompting and planning? for sure. But it also shows that something is wrong, the agents are making stupid mistakes that they were couple of weeks ago.

[Fernando Dâmaso @fernandodamaso]

wtf is happening with @OpenAI ? sol and luna were GOOD at computer use, but now they are SO BAD, i just got sol and 3 luna subagents STUCK IN A LOOP for half an hour trying to perform tests. they couldn't understand the ssr and auth to pass the acceptance criteria scenarios

[KruelVerse @theSosiKevin]

… frieren

[rohit @rohit3a]

You really can’t make this stuff up. Dudes are basically the Nokia & Blackberry of the AI industry.

[MakerThrive @MakerThrive]

guilty as charged.

[Fernando Dâmaso @fernandodamaso]

@PatCastiglia That's a beauty of omarchy tough, it's so flexible, in my living room pc, I use the mouse a lot more, so I've developed a plug-in dock that manage my workspaces and windows, and it is a lot more mouse friendly, and others small tweaks in the os to make it more mouse friendly

[Fernando Dâmaso @fernandodamaso]

@StatsWire Pretty sure this 'new' model is the sol 5.6 sol model with some more training, and the sol 6 is a trained terra

[Highlander83art @highlander83Art]

Rites of Maintenance

[Fernando Dâmaso @fernandodamaso]

@goon_nguyen YES! Chatgpt is acting lazy and dumb!

[Fernando Dâmaso @fernandodamaso]

@robertagleao @RegisPessolano eu uso @qbittorrent hehehe

[Jonathan Hill | Hexakin @Hexakin]

as someone who uses AI every single day, I can confirm that GPT-6-Sol is a terrible model.

I dont care how cheap it is. It's horrible to speak to, its loses track of the conversation, it's writing ability is the worst of any GPT model for a long time.

This feels like a rush-job

[Imp - VTuber & VA @FlawedImp]

Might as well give voice to it!

Thanks to @Toro_Yuma for the idea!