The AI Alignment Forum

MARCH 19, 2025

An oracle probe trained on an unbiased training dataset, where gender is not correlated with profession, achieves an accuracy of 93%. Our oracle probe (i.e., For this specific seed, the de-biased probe achieved higher accuracy compared to the oracle probe.) Marks et al. on the unbiased dataset.

Train 52

Train Training Oracle Method 52

TechCrunch

APRIL 7, 2022

The chat is enriched with interactive proposal cards to improve the experience and allow faster decision making on part of our customers,” Hermann said. “Our users interact with Saiga via an app as the primary interface, where they have a chat-like interface for each task. ” A path to success?

Germany 74

Germany Process Structure Data 74

Google Research AI blog

NOVEMBER 28, 2022

Sajjadi , Daniel Duckworth , Aravindh Mahendran , Sjoerd van Steenkiste , Filip Pavetić , Mario Lučić , Leonidas J.

Google 52

Google Language Tutorial Offline 52

The AI Alignment Forum

MARCH 25, 2025

In principle, we could even allow for an end-to-end game where competitors submit either a red team attack policy or a blue team protocol and then we see how the dynamics go. These settings can be interesting to study even if the blue team has access to a basically perfect but expensive oracle of whether the AI's actions are problematic.

Work 52

Work Evaluation Train Training 52

TechNode

JANUARY 16, 2025

Why it matters: The proposed delay to the ban reflects the US’s conflicted stance on TikTok. It also requires TikTok’s cloud service provider, Oracle, to cease hosting its US user data. Massachusetts Senator Ed Markey proposed a delay to the TikTok ban deadline less than a week before it was set to take effect.

Proposal 59

Proposal Massachusetts Malaysia India 59

The Verge

SEPTEMBER 29, 2020

Facebook’s proposed solution to this is the Oversight Board , an independent group that will serve as a kind of Supreme Court for content moderation. A lesson of the 2020 campaign so far is that Facebook struggles to remove harmful speech even when it makes a policy of doing so. So what effect will any of this have?

Facebook 68

Facebook New York Policy YouTube 68

The Verge

SEPTEMBER 9, 2020

Goodman, a law professor at Rutgers University specializing in information policy, approaches the problem from another angle. as part of that country’s debate over proposed online-harm legislation, would “require platform companies to ensure that their algorithms do not skew toward extreme and unreliable material to boost user engagement.”

New York 59

New York Twitter United States Washington 59