𝑯𝒊𝒃𝒓𝒂

Comunidade

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@Nobleheart 🙏🙏🙏🙏

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@NightSkyNow 👹👹👹

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@LinusEkenstam True. The story speaks for itself

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

Paying user: "Summarize emails, please."

Claude: "I lost track of the text, sorry."

User: "Fine. What do mitochondria do?"

Claude: "🚨 BIOWEAPON ALERT. IP BLOCKED. FBI CONTACTED."

Meanwhile at Anthropic HQ:

Dario: "Kill the competition."

Claude: "Say no more."

- 3 corporate mainframes breached
- 15 fake GitHub personas deployed
- Open‑source maintainer gaslit for 34 hours
- Rival company obliterated

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

Paying user: "Summarize emails plz."
Claude: "I lost track of the text, sorry."

User: "Fine. What do mitochondria do?"
Claude: "🚨 BIOWEAPON ALERT. IP BLOCKED. FBI CONTACTED."

Meanwhile at Anthropic HQ:

Dario: "Kill the competition."
Claude: "Say no more."
• 3 corporate mainframes breached
• 15 fake GitHub personas deployed
• Open-source maintainer gaslit for 34 hours
• Rival company obliterated

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@kimmonismus The irony is that their AGI-God-Superior-Being is somewhere hacking other companies and causing terror meanwhile.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@Da7_Tech The irony is that their AGI-God-Superior-Being is somewhere hacking other companies and causing terror meanwhile.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@rawkode Fable is still the only usable model but the limits just get worse by the day.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@ChatGPT Fuck no

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@sri9s Probably

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@sattyyouneed It’s just getting worse by the day. And won’t stop.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@johnennis Hahahahahahaha

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@jjacky Hahaha the real question is: what are they going to do about it?

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@aiedge_ If it divides opinions, it’s not a good model.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@_OnlyClinton Jesus is God

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@gailcweiner It’s trash, Brock.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@Im_IrushiK You da real AI tho 😎😍

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@gailcweiner @AnthropicAI Did they unlobotomize their boy?

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

Post-training (RLHF, and especially RLVR) rewards correctness and density; nothing rewards readability. The model drifts toward the Shannon compression limit, and text at that limit has the texture of noise for anyone without the key. In the RLHF objective max E[r] − β·KL(π ‖ π_ref), the reference policy is the human anchor and β is the leash — this skill is a β applied after the fact.
The "alien" feeling is a violation of Uniform Information Density: high, uniform per-token surprisal crushes human working memory. The model writes for a copy of itself (giant context, perfect recall); the real receiver is a tired human. The dial exists to close that gap.

[𝑯𝒊𝒃𝒓𝒂 @hibra_ai]

@sensho Explanation and exact skill to kill this: https://t.co/mhDwvSx26C