doorss

Anthropic research on training Claude to understand the reasoning behind instructions, aiming to improve the model's ability to generalize from guidance rather than just follow rules mechanically.

Anthropic presents research on extracting and representing Claude's internal reasoning as human-readable natural language, contributing to interpretability efforts.

Anthropic is donating Petri, its open-source alignment testing toolbox, which can rapidly test AI models for alignment properties and was developed through the Anthropic Fellows program.

A survey of 81,000 Claude users found that people in AI-exposed roles have more concerns about job displacement, with concerns highest among younger workers and those in lower-income countries.

Meta has removed end-to-end encryption from Instagram direct messages, reversing previous privacy commitments for the platform.

Google Chrome has been found to silently download a 4 GB AI model (Gemini Nano) onto users' devices without explicit consent or notification.

Google's reCAPTCHA has become non-functional for Android users who have removed Google services from their devices, effectively penalizing users who prioritize privacy.

Despite recent improvements, AI sycophancy remains a significant problem — a Stanford paper found sycophantic models led to worse decision-making, and the post explores how this affects daily work with LLMs.

A reflection on historical parallels to the current AI coding wave, examining what skills and practices were lost when previous technological shifts made code production cheaper.

Simon Willison discusses the blurring line between casual "vibe coding" and professional agentic engineering, raising concerns about quality and responsibility as AI coding tools become more capable.

Research examining how large language models are measurably changing written language patterns, documenting specific linguistic distortions that LLM-generated text introduces into the broader corpus.

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

Cambridge University research finds gambling advertisements on social media platforms disproportionately reach men, raising concerns about targeted advertising practices in the gambling industry.

TV advertising is experiencing a resurgence as streaming platforms introduce ad-supported tiers, making the format increasingly valuable for digital platforms seeking new revenue sources.

Bluesky published technical details about how its "For You" algorithmic feed works, highlighting the platform's unique approach where anyone can run custom feed implementations.

Bluesky experienced a prolonged DDoS attack lasting nearly a full day, disrupting service on the growing decentralized social media platform.

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

Tekoälyvideot kaatuneista venäläissotilaista kaupallistavat surun ja tukevat Kremlin sankarikuvastoa.

Sitran kyselyn mukaan suomalaiset kannattavat laajasti EU:n taloudellisten ja teknologisten riippuvuuksien vähentämistä ja priorisoivat turvallisuutta.