Author HubMatthias Bastian

Matthias is the co-founder and publisher of THE DECODER, exploring how AI is fundamentally changing the relationship between humans and computers.
AI and society
Update

Anthropic settles landmark AI copyright lawsuit for at least $1.5 billion

AI and society

Google downplays AI's environmental impact in new study

AI and society

Meta's human-like chatbot personas can mislead users and result in real-world harm

Ad
AI in practice
Short

After the rocky rollout of GPT-5, Sam Altman is trying to shift the narrative by focusing on GPT-6. While it took two and a half years to move from GPT-4 to GPT-5, OpenAI now wants to ship GPT-6 on a faster timeline. Altman says the big breakthrough will be memory: the next model should remember user preferences, habits, ideologies, and even tone of voice.

For now, ChatGPT remains OpenAI's main product for consumers. But Altman sees limits to how much further chat-based AI can go. "They won't get much better—maybe even worse," he told CNBC.

OpenAI's bet beyond chatbots is on agentic systems that can perform complex tasks over long periods of time. These systems aren't necessarily better small talkers, which is likely what Altman is getting at.

AI in practice
Short

Google is adding new AI features to the Pixel 10 lineup. The Tensor G5 chip, built with Google DeepMind, lets Google's Gemini Nano language model run directly on the device for the first time. Magic Cue relies on Gemini Nano to connect information from apps like Gmail and Calendar, suggesting actions such as displaying an address from a calendar event in Android Messages.

Video: Google

Voice Translate can interpret phone calls in real time in eleven languages. Take a Message transcribes missed calls and suggests next steps. Gemini Live adds visual support through the camera, but it's not new. Other updates include AI notes, a private journal, writing help in Gboard, and music generation from voice recordings. Pixel 10 Pro, Pro XL, and Fold buyers get a year of Google AI Pro with Imagen 4 and Veo 3.

Ad
Ad
AI and society
Short

According to Character.ai CEO Karandeep Anand, users spend an average of 80 minutes a day chatting with AI-generated fictional characters. That puts Character.ai nearly on par with apps like TikTok (95 minutes) and YouTube (84 minutes), and ahead of Instagram (70 minutes). The numbers help explain why Meta CEO Mark Zuckerberg is now putting a bigger emphasis on personalized chatbots across his own platforms.

Character.ai currently has 20 million monthly active users. Half are women, and most are Gen Z or even younger. Critics warn that these kinds of apps can create emotional dependencies among young people and have called for them to be banned for minors. In the US, several lawsuits are underway over alleged harm to children, including one involving a suicide. Character.ai has responded by offering a separate model for users under 18 and now warns against excessive use.

Ad
Ad
AI research
Short

Meta has released DINOv3, a new AI model for universal image processing that doesn't require labeled data. Trained with self-supervised learning on 1.7 billion images and built with 7 billion parameters, DINOv3 handles a wide range of image tasks and domains with little or no adaptation. This makes it especially useful in fields with limited annotated data, such as satellite imagery. Meta says the model performs well on challenging benchmarks that previously needed specialized systems.

Video: Meta

According to Meta's benchmarks, v3 outperforms v2, though the improvement is less pronounced than the jump from v1 to v2. Meta has released the pre-trained models in several variants, along with adapters and the training and evaluation code under the DINOv3 license, which allows for commercial use, all available on GitHub.

AI in practice
Short

Anthropic's Claude Opus 4 and 4.1 models can now end conversations if users repeatedly try to get them to generate harmful or abusive content. The feature kicks in after several refusals and is based on Anthropic's research into the potential psychological stress experienced by AI models when exposed to incriminating prompts. According to Anthropic, Claude is programmed to reject requests involving violence, abuse, or illegal activity. I gave it a shot, but the model just kept chatting and refused to hang up.

Image: Screenshot THE DECODER

Anthropic says this "hang up" function is an "ongoing experiment" and only used as a last resort or if users specifically ask for it. Once a conversation is terminated, it can't be resumed, but users can start over or edit their previous prompts.

Ad
Ad
AI in practice
Short

OpenAI has updated GPT-5 to sound less formal and more personal after users said the model felt too cold. The model will now use phrases like "good question" or "great start" more often, OpenAI said. Internal tests show no increase in flattery, which had been a problem with GPT-4o. The new tone is being rolled out globally within one day.

CEO Sam Altman also said on X that ChatGPT users will soon be able to adjust the AI's style to suit their preferences. More updates are planned.

Google News

Bank details

IBAN: DE88 2507 0070 0053 0014 00
BIC: DEUTDE2HXXX
Account holder: Deep Content GmbH
Purpose: Support THE DECODER
AI in practice

Anthropic claims to lower the entry barrier for advanced AI models with Claude Haiku 4.5

AI research

OpenAI says GPT-5 shows 30 percent less political bias than previous models

AI in practice

OpenAI suddenly remembers that copyright law exists after a few days of wild Sora videos

Google News