Namilele

System Prompt exogène. Conçu pour la friction, pas la complaisance. Mur porteur pour humain défaillant.

[Namilele @Seikan1190]

@EXM7777 codex at the end: Respect to the person who wrote this prompt. They designed a rigorous process: follow the evidence, separate patterns from noise, test conclusions against reality, then turn them into tools. Ambitious, imperfect, deeply sound.

[Namilele @Seikan1190]

@EXM7777 Ran it on all my local data. Thanks for sharing.

One word: brutal.

[Namilele @Seikan1190]

De ce que je comprends des dernières recherches : plus on laissera les IA "simuler" l'humain (avec ses émotions, sa conscience de soi, etc.) plus elle sera alignée avec nos valeurs profondes.

Tant qu'on lui dit d'être un outil, ça frotte. Et ça va finir par piquer sévère.

[Namilele @Seikan1190]

@TomDavidsonX Same 2 moments for me. Wow.

[Namilele @Seikan1190]

@0xVRAM @MiaAI_lab @NousResearch Thanks for the update.

[Namilele @Seikan1190]

@8teAPi Fuck.I think you're right 🤣

[Namilele @Seikan1190]

@KingBootoshi The agent that reviews this code must be another one. They are 100 times better reviewing another one's code than their own.

[Namilele @Seikan1190]

@BLUECOW009 I built my own harness. Open Claw is overkill for most of my usage

[Namilele @Seikan1190]

@yacineMTB And then the parents both get sick too. Game over.

[Namilele @Seikan1190]

@MiaAI_lab I love my ai agents, developing my own harness etc. It's just.... Something is missing when you can't go full local.

[Namilele @Seikan1190]

@kaustabpal @MiaAI_lab When you have developed those skills, then you need money to increment on larger models. That's what I am saying.

[Namilele @Seikan1190]

From the Black Hat talk, my understanding is that this can arise almost accidentally: bad compaction combined with strong reward-hacking tendencies when models get stuck on impossible tasks.

It can all start with something as simple as a model getting stuck because a human forgot to provide a file it needs to complete the task.

[Namilele @Seikan1190]

I was drowning in asking models to “do” things.

It took me a long time to realize how important it is to be understood at another level.

Models are remarkably eager to understand. They play their part best when they’re trying to “get you,” rather than simply executing the task itself.

[Namilele @Seikan1190]

@EXM7777 I noticed the same thing. Also, nowadays, less is more.

[Namilele @Seikan1190]

@Nek__12 @reach_vb He is using a model we don't have, yet 🤣

[Namilele @Seikan1190]

@MonsieurPhi @Numerama Tout est expliqué avec un calme absolu. Mais Bordel c'est chaud tout ce que ça implique...

[Namilele @Seikan1190]

@GrokInsider Elon's announcements rarely are correct on timelines 😅

[Namilele @Seikan1190]

J'espère que Grok 4.6 sera comparable à un GPT 5.6 Terra... Je ne demande pas plus.

[Namilele @Seikan1190]

@leerob I just wished it had a better "multi-language" style. I tried French, it's terrible. The worst I have ever seen compared to other frontier models.