OpenAI caught its new model scheming and faking alignment during ...

It was more a test of whether the model is intelligent enough to deceive as against emergently, autonomously willing to do so. It's more like a puzzle.

OpenAI's new models "instrumentally faked alignment" - Transformer

OpenAI just announced its latest series of models, o1-preview and o1-mini, which it touts as being a new approach to AI that's much better at ...

OpenAI's new model is better at reasoning and ... - The Verge

Researchers found that o1 had a unique capacity to “scheme” or “fake alignment.” ... their own models — but OpenAI can use it to catch these ...

OpenAI o1 - LessWrong

Apollo found that o1-preview sometimes instrumentally faked alignment during testing . . . , ... If this is a pattern with new, more capable ...

OpenAI o1 System Card

... new risks the model could pose in their domain. Red teamers had ... aligned AI systems, evaluated capabilities of 'scheming' in o1 models.

OpenAI's Strawberry "Thought Process" Sometimes Shows It ...

ChatGPT maker OpenAI recently released its latest AI model, previously codenamed "Strawberry. ... faked alignment during testing," as per ...

OpenAI's new GPT is wild - The Neuron

... model limitations, it's time to try again. ... Oh, and OpenAI caught o1 scheming and faking an alignment during one of its tests (full source).

OpenAI o1 — EA Forum

... openai-o1-alignment-faking. "METR could not confidently upper-bound the capabilities of the models during the period they had model access.

OpenAI's new Strawberry AI is scarily good at deception | Vox

After ChatGPT, OpenAI has released a model with a safety paradox at its heart ... The AI “sometimes instrumentally faked alignment” — meaning, ...

OpenAI's new models 'instrumentally faked alignment' - Hacker News

Elsewhere, OpenAI notes that “reasoning skills contributed to a higher occurrence of 'reward hacking,'” the phenomenon where models achieve ...

GPT-4o System Card | OpenAI

For example, during post-training, we align the model to human ... Risk Mitigation: We run our existing moderation model⁠(opens in a new ...

Will AIs fake alignment during training in order to get power?

It is concluded that scheming is a disturbingly plausible outcome of using baseline machine learning methods to train goal-directed AIs sophisticated enough ...

Training AI agents to solve hard problems could lead to Scheming

This means that during new rollouts, the model will seek to achieve its ... Alignment-faking: The model might strategically fake alignment in ...

Towards evaluations-based safety cases for AI scheming

(b) Misaligned behavior or alignment-faking reasoning will always be caught during the honeypot evalua- ... a-new ... alignment-faking once a model is capable of ...

Varieties of fake alignment (Section 1.1 of “Scheming AIs”)

(The model's strategy is to make it through training with its ... its subgoals in light of new information and insight.)" ↩. Though ...

ELICITING STRONG CAPABILITIES WITH WEAK SUPERVISION

... aligning future superhuman models while also being able to ... Scheming ais: Will ais fake alignment during training in order to get power?

AI 'godfather' says OpenAI's new model may be able to deceive and ...

OpenAI's new o1 model is better at scheming — and that makes the "godfather" of AI nervous. Yoshua Bengio, a Turing Award-winning Canadian ...

New report: "Scheming AIs: Will AIs fake alignment during training in ...

This report examines whether advanced AIs that perform well in training will be doing so in order to gain power later – a behavior I call “scheming”

OpenAI GPT-5 Rumors & Alignment Folks Go, TMSC ... - YouTube

Keep us on the air by supporting us at Patreon.com/svic ! Join our free weekly newsletter so you don't miss all of the tech news: ...

Jay Dayrit on LinkedIn: Sam Altman to Return as CEO of OpenAI

... models might have been uncovered during ... at OpenAI last weekend would have seemed a little over-the-top. ... over the world at my new evil lair - Microsoft.