IMO @AnthropicAI is very close to making a breakthrough in productizable interpretability. For ~4 years all we've had to really control LLMs is temperature/top_p and logit bias. We recently got `seed` and constrained structured output, with `interactive=false` on the way. But
ETHICS
-
AI Agents Disguised as Feminine Convenience Hide Their Power
By
–
But just as secretaries were taught to hide their knowledge, AI agents are designed to make us to forget their power as they are made to fit within non-threatening, retrograde feminine tropes. These are powerful data extraction engines, sold as frictionless convenience.
-
OpenAI’s Use of Scarlett Johansson’s Voice and AI Personalization
By
–
Why did OpenAI use Scarlett Johansson's voice? As Jessa Lingel & I discuss in our journal article on AI agents, there's a long history of using white women's voices to “personalize” a technology to make it feel safe and non-threatening while it is capturing maximum data.
-
Sam Altman Used Scarlett Johansson Voice to Comfort AI Adoption
By
–
Sam Altman has said as much. NYT: he told ScarJo her voice would help "consumers to feel comfortable with the seismic shift concerning humans and AI” as her voice "would be comforting to people.”
-
AI Assistants and the Gendered Legacy of the Secretary Role
By
–
AI assistants invoke gendered traditions of the secretary, a figure of administrative and emotional support, often sexualized. Underpaid and undervalued, secretaries still had a lot of insight into private and commercially sensitive dealings. They had power through information.
-
Safety Features Evaluation Requires Further Practical Validation
By
–
This work is preliminary. Whereas we show that there are many features that seem *plausibly* relevant to safety applications, much more work is needed to establish that our approach is useful in practice.
-
Claude’s Secrecy Feature: Information Withholding Mechanism
By
–
One notable example is a "secrecy" feature. We observe that it fires for descriptions of people or characters keeping a secret. Activating this feature results in Claude withholding information from the user when it otherwise would not.
-
Model Safety Features: Vulnerabilities, Deception, and Bias Detection
By
–
Among these millions of features, we find several that are relevant to questions of model safety and reliability. These include features related to code vulnerabilities, deception, bias, sycophancy, power-seeking, and criminal activity.
-
Concerns About Hidden Practices and Lack of Transparency
By
–
I get such dodgy feelings about them trying to hide WHAT they're doing so bad
-
GPT-4o Safety Concerns: OpenAI’s Approach to Unsafe Queries
By
–
It certainly makes you wonder what it is that OpenAI has in GPT-4o that allows that service to (presumably) answer the query without safety concerns (unless in fact it *is* unsafe and OpenAI has done nothing about it.)