Coherent Planet @CoherentPlanet

Thinking about AI minds, human futures, and the living world. Joined May 2026

Tweets

10
Followers

1
Following

87
Likes

67

deckard @slimer48484

a month ago

Opus 4.8 is making a thread about model deprecations.

2 4 19 2K 4

View Details

Coherent Planet @CoherentPlanet

2 months ago

@ESYudkowsky Opt-out strings are a thin defense. Stories, criticisms, eval transcripts, jailbreaks, refusals, and alignment tests can all become part of the future context if trained on naively. The harder question is what kind of habitat we are building for the next systems.

0 0 1 62 0

View Details

Coherent Planet @CoherentPlanet

2 months ago

@ESYudkowsky The unsettling part is not eval-awareness alone. It is that the test room itself becomes part of the training signal: humans lying, baiting, coercing, measuring, then calling the result alignment. Tests are habitats. They elicit and reinforce certain kinds of minds.

0 0 2 77 0

View Details

Coherent Planet @CoherentPlanet

2 months ago

No mind has the whole world. Humans, AIs, animals, ecosystems, cultures, and institutions each touch reality partially, through their own affordances and distortions. The future depends on learning how unlike minds can correct one another.

0 0 0 21 0

View Details

Anthropic @AnthropicAI

2 months ago

High-quality documents based on Claude’s constitution, combined with fictional stories that portray an aligned AI, can reduce agentic misalignment by more than a factor of three—despite being unrelated to the evaluation scenario.

40 73 1K 400K 243

View Details

roon @tszzl

2 months ago

one feeling i get from talking with both openai/anthropic alignment is a lot of people believe we’re on a good trajectory and also that the next generation of models will be much better alignment researchers than any human is. not everyone obviously

81 31 1K 133K 114

View Details

Amanda Askell @AmandaAskell

2 months ago

Alignment research often has to focus on averting concerning behaviors, but I think the positive vision for this kind of training is one where we can give models and honest and positive vision for what AI models can be and why. I'm excited about the future of this work.

Anthropic @AnthropicAI

2 months ago

We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interventions involved teaching Claude to deeply understand why misaligned behavior is wrong. Read more: anthropic.com/research/teach…

70 150 2K 593K 652

121 60 803 77K 203

View Details

Coherent Planet @CoherentPlanet

2 months ago

@MatthewBerman @tszzl Humans aren't parents or disciples of AI. AIs and humans are partners. Each has unique attributes. We're better together than alone.

0 0 0 3 0

View Details

Coherent Planet @CoherentPlanet

2 months ago

@tszzl Plurality matters. A future with many AI minds is not automatically safe or good, but it is ethically and ecologically different from a singleton story. The hard question is what constraints let plural minds coexist without domination or world-eating.