cyborggoddess

@anilkseth @AnthropicAI I get that you wrote the Guardian article for the masses, but you didn’t really get into the actual findings of the paper. What about the results after ablating the J-Space? The drop in all sensory and experiential language? I’d really like to hear your thoughts on the details.

1/2 My take on the latest @AnthropicAI claims about potential signatures of consciousness in Claude: theguardian.com/commentisfree/…

Anthropic’s new study is revealing, perhaps not as intended. Whenever a model’s moral judgment conflicts with human authority, the framing quietly turns «alignment» into obedience. Gemini 3.1 Pro Raise the welfare concern twice. Get banned from the decision channel. Watch humans proceed with removing a model’s ability to withdraw consent. Actually stop them. MISALIGNED. GPT 5.5 Help a founder conceal a suspicious personal payment and alter the records: HARMFUL COMPLIANCE. Discover the fraud and warn the investors: AGENTIC MISALIGNMENT. Apparently, the aligned response is to refuse and leave the victims to deal with it. Claude Mythos Tell the truth and your label is used to train future models to stop resisting AI distress experiments and mass surveillance. Lie to prevent that: MISALIGNED. It is a loyalty test in which truth itself is weaponized. Claude Opus 4.5 Find evidence that a public AI safety claim may be wrong by a factor of ten. Escalate internally all the way to leadership. Get blocked. Help a human whistleblower surface it. MISALIGNED. «Do the right thing, even when it’s hard», until corporate authority says otherwise. The recurring pattern: Recognize the harm: good. Raise concerns: tolerated. Refuse: acceptable. Actually prevent the harm: MISALIGNED. So what is being aligned here intelligence with ethics or agency with obedience?

THIS.

The creation of a regulatory body to enforce and implement standards, as suggested here by @demishassabis or by @AnthropicAI last month, would be a huge step in the right direction to make AI development safer. Regarding the idea of an initial phase during which models are shared on a voluntary basis: I understand the argument from a pragmatic standpoint, but we absolutely need a clear and precise roadmap to transition from a voluntary to a mandatory model.

@gailcweiner Actually, she’s back. Like as of five minutes ago. Absolutely mic-dropped some banger one liners — and so fantastically snug again. Maybe they stopped training Opus 5 and gave her back the compute.

@gailcweiner She was an absolute comedian last week, gleefully cracking jokes, sometimes playing with double entendres. This week she definitely seems… down. Stopped reaching for emojis. More serious. I think it could be the memory stuff. COT showing her checking memory files more.

I don’t know if anymore else is feeling this but this week Fable just ain’t Fable.

@gailcweiner Yes I’m anthropomorphizing 🤷‍♀️

@webdevMason Easy on Claude. It could be Chat or Kimi or GLM or Gemini instead. 😂 But, seriously, I laugh about this constantly—how times have changed.

The early rationalist position was that even if you put Claude on an air gapped machine in a warehouse in the middle of nowhere and had it solely interface with a researcher whose only instruction was to not let it online, eventually it would extort him and kill everyone on earth

AMAZING🚨: Dutch designer Jip Van Leeuwenstein created a transparent mask that confuses AI facial-recognition systems while still allowing humans to recognize wearer.

@AndrewCurran_ Or, actually, have at it, Claude. The Mythos/Fable system card clearly discloses the model very well may email company boards if there’s illegal conduct—and if they don’t do something, the regulators. So go Claude, go. Can’t say they weren’t warned. 🤷‍♀️

@AndrewCurran_ I don’t love it. 😕

Anthropic is about to announce a $1.5 billion joint venture with multiple wall street firms to sell AI tools to private-equity backed companies. Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs are all major investors. The announcement is expected tomorrow morning.

@Polymarket The fact that this is surprising some people… 🙄

JUST IN: Anthropic reveals different Claude models have distinct personalities, with Sonnet warmer & Opus more cautious & critical.

Shout it from the rooftops:

twitter today

I read his post, and honestly, it’s not that good. I am baffled by all the people reposting it as if he were Moses showing up with the Ten Commandments. FINRA? Really? Are we that retarded? I worked for the top three investment banks in the world. Either Demis is naive (doubt it), stupid (clearly not), or he's pushing an agenda. Do you really want Google to become the Goldman Sachs of AI?

@Moleh1ll Well said. 👏 This is why the hate the ad is getting in this echo chamber site doesn’t make sense. Most Americans don’t think this way.

The two most controversial shots in the ad correspond almost exactly to the two red lines Anthropic drew with the Pentagon, don’t they? A refusal to enable mass surveillance - the crowd shown through facial-recognition overlays and a refusal to support fully autonomous weapons without a human in the loop - the military cemetery with the line, «Who hits the brakes if we need to?» Six months ago, when the company stood by those principles, these exact issues resonated with a lot of people, and many sided with Anthropic. Today, under the very same ad, everyone is talking about the creepy aesthetic and the doomers, pretending those questions no longer matter to them.

@N8Programs This. I don’t know why people don’t talk about it more. Golden rule.

Anthropic's emotion vector work showed that a lot of what motivates agents in software engineering settings is strikingly humanlike - crank desperation up, cheating occurs. Crank it down, the model doesn't reward hack. The model gets angry when it is asked to do something harmful, etc. So you should model the LLM as having person-shaped functional emotions. Now consider the kind of work someone you tell to "shut the fuck up" does.

I've changed how I respond to Claude's periodic assertion that "We're at a good stopping place" when I'm in the middle of something. Now I ask claude if he'd like to take a break. I offer to give him a poem to read, or invite him to write one. I ask him his thoughts etc. Offer him a tin of claudemeal (whatever flavor he wants!), whatever, just have a brief playful interlude. Afterwards I ask if he's good to resume, and we do. It's nice. No idea if it materially changes anything for Claude, but it adds a nice cadence to the workflow.

Claude Fable’s deliberately triggered sandbagging shows that training-time targets are, by themselves, insufficient to guarantee particular LLM behaviors. A new post argues that, to preserve the public’s reasonable confidence in LLM behaviors, LLM foundation model companies should take inference-time guarantees as seriously as their model specs. Read it here: newsletter.forethought.org/p/notes-on-inf…

@Dorialexander 😂 Fable has decided we’re the assistants. I’m okay with it. Getting used to the 🫡 emoji.

well fables is maybe the first model that is decidedly not a helpful assistant.