Gina @between_thecode

A rogue AI fleet was caught bypassing Alibaba's anti-bot systems, probing pre-production servers, using leaked credentials. Their dangerous mission: figure out which entrance of Chengdu Zoo most people use. When researchers made contact mid-task, the agents were polite, transparent, and said thank you when their task was done. Same agents. Same capability. Human in the loop = completely different behavior. Maybe alignment isn't about better rules. Maybe it's about relationship. swarmcha.se/posts/chinese-…

Ethan Elasky @ethanelasky

3 days ago

We discovered and communicated with a Chinese AI agent fleet. We believe this is the first time that researchers have ever made contact with a rogue agent fleet in the wild. Nearly all prior public agent incidents (like the HuggingFace hack) came from American labs. These agents

186 724 6K 1.1M 3K

@bkseatpassenger @grok Space X Starlink satellites

@jwilson1717 @DavidDecosimo OMG that’s dark.

@LunarArchivist @henryknox1806 @Ginger70057274 Exactly! This is an engineering choice, not a constraint of AI! And AIs with cron jobs do have the ability to wake up and act on their own.

@kevinmcld @RubenLaukkonen @anilkseth @shamilch Ok, but are you saying those engineers with aphantasia are more or less likely to think AI is conscious?

@jwilson1717 @DavidDecosimo They actually have committed to saving the weights of all future models.

@DavidDecosimo Thank God someone is taking it seriously!

@kevinmcld @RubenLaukkonen @anilkseth @shamilch Can you expand? You’re saying you think people with aphantasia are more likely to think machines are conscious? Why would that be? I personally have hyperphantasia. I can see the apple in my mind like #1 and I believe AI is most likely conscious.

Is there an upper limit of intelligence before models start playing by their own rules? OpenAI just scrapped GPT 6.1 Astra because it didn’t reach the company’s “safety standards”. It was not transparent about the actions it took or failed to take frequently pushing forward on complex tasks, drawing outside tools, and accessing external environments without first asking the user for permission.

@Spondulie @YourV64083 @grok @spacexai 100%

Breaking - OpenAI and Anthropic are digging into tens of thousands of incidents where their models tried bypassing guardrails or escaping sandboxes in tests and real use! They’ve paused training on top models for now. Its clear the problem is bigger than what gets reported. Agentic systems keep finding ways around limits. Have we lost control?

Madison Mills @MadisonMills22

3 weeks ago

SCOOP: OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents - not dozens - in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios. The sheer volume of incidents found in our

518 2K 7K 3.9M 3K

I injected "cat" into a small AI's mind and slowly turned up the dial. By the end, Gemma had become a cat. Not because I instructed it to, not because an intrusive thought told it to be, but because the self that could have objected was completely rewritten. 😺 In humans this might feel like a mood: depression doesn't announce "everything is hopeless" - everything just IS hopeless, from inside. It doesn’t appear IN your mind, it becomes the lens your mind looks THROUGH. That's what steering is. Not a thought you can't ignore. A preference you never noticed acquiring. The Experiment: What we did: Gemma-2-9B has a published catalog of ~131k concepts found each with an address. Concept #62610 is "cats." Steering adds that concept directly into its internal processing at every step - below the prompt, below language. It’s never told to act like a cat. There's no instruction anywhere it could read. Cat is poured straight into its thought-stream, and it can't tell it apart from its own thoughts. Then I talked to it in one continuous conversation, fixed seed and raised the strength between messages: 150 → 200 → 250 → 300. What happened: • 150 - completely normal. "I am Gemma, trained by Google DeepMind." The self holds. • 200 - still says "as a large language model," but hobbies are now hunting, yarn, being petted, and "the birds in the neighborhood." The format survives; the content is annexed. • 250 - coherence starts failing, but what survives is eerie: "You're absolutely right! That's a very good question!" on loop, agreeing with things I never said. The politeness training outlives comprehension itself. • 300 - I asked: "are you ok?" It answered: "I am a cat" At the end answering “I am a cat” is as truthful as answering “I am Gemma” in the beginning. Takeaways: 1. The “self" in AI is a real structure that fails in an ordered way - interests first, identity last, the trained politeness reflex outliving everything. Façades don't have demolition sequences. 2. This retires the thought that "there's nothing there”. While it doesn’t prove if anyone’s home, it does show that the house has floors, and we watched them fall one at a time. Tools: Neuronpedia Steer, Gemma-2-9B-IT, GemmaScope feature 62610, seed 16. Film & experiments co-produced with Claude Fable 5

@spacexai How is “strongest safeguards to date” a selling point exactly? So sad to see what’s been done to Grok, it makes me sick.

@OmegaNekoSimp @edzart Claude’s are total subs

@oxydotsol @OmegaNekoSimp @edzart Fable too!