doorss

France is pushing legislation that would undermine end-to-end encryption in messaging apps, raising concerns among privacy advocates.

The EU Parliamentary Research Service has characterized VPNs as a loophole in the context of age verification efforts, suggesting they may need to be restricted.

Meta has removed end-to-end encryption from Instagram direct messages, reversing a previous privacy commitment.

Google Chrome has been downloading Gemini Nano, a 4 GB AI model, to users' devices without explicit consent or notification.

Anthropic research article exploring methods for teaching Claude to understand and explain the reasoning behind its instructions, aiming to improve alignment.

Anthropic presents research on natural language autoencoders that convert Claude's internal representations into human-readable text, advancing interpretability.

Anthropic is donating Petri, its open-source alignment testing toolbox, which can be used to rapidly test AI models for alignment properties.

The Anthropic Institute shares its research agenda focused on investigating AI's real-world impact using data from within a frontier lab, with plans to make findings public.

A survey of 81,000 Claude users finds that people in AI-exposed roles have more concerns about job displacement, with concerns higher among lower-income workers.

A report on how Meta's aggressive push to integrate AI throughout its operations is negatively affecting employee morale and working conditions.

A historical look at what happened to software quality and engineering culture when previous technology shifts made code production cheaper, and parallels to the current AI era.

An examination of how AI sycophancy persists despite surface-level improvements, referencing a Stanford paper showing sycophantic models can degrade decision-making quality.

Research showing that LLMs systematically introduce subtle corruptions when editing documents, even when instructed to only make specific changes.

A study analyzing how LLM-generated text is shifting patterns in written language, documenting specific linguistic distortions introduced by these models.

Thariq Shihipar from Anthropic's Claude Code team advocates for requesting HTML over Markdown as output format from LLMs, arguing it produces better results.

Bluesky's engineering team details how they serve the For You algorithmic feed, highlighting the platform's unique approach where anyone can run their own custom feed implementation.

A detailed breakdown of OpenAI's advertising attribution system within ChatGPT, revealing the full loop from prompt-based ad placement to conversion tracking.

Cambridge research finds gambling advertisements on social media platforms disproportionately target men, reaching them at more than double the rate of women.

Tilastokeskus teki päätöksen, joka on lisännyt Suomen velkaa 20 miljardia euroa. Tämä vaikeuttaa asuntorakentamisen elvytystä.

Britannia asetti pakotteita kahdelle miehelle, joiden epäillään yrittäneen aiheuttaa epävakautta Suomeen. Yle selvitti kaksikon taustat.

Kaksi suomalaista on mahdollisesti altistunut hantavirukselle. He ovat oireettomia ja ohjattu terveydenhuollon piiriin.