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 research
Update

So-called reasoning models are more efficient but not more capable than regular LLMs, study finds

AI in practice

Google leans on token metrics, not revenue, adding to bubble talk about AI growth

AI and society

The long-predicted deepfake dystopia has arrived with Sora 2

Ad
AI in practice
Short

Resemble AI has released Chatterbox, a free open-source voice cloning model that runs locally and supports emotional tone control like "dramatic" or "monotone." It clones voices using just a few seconds of audio and responds in under 200 milliseconds. The tool works on Windows, Mac, and Linux with 5–6 GB of video memory. All generated speech includes a faint watermark, "PerTh," to identify it as AI-made. According to Resemble AI, it performed better than ElevenLabs in blind tests. Currently, it only supports English.

Decoder EN demo (heightened emotional expression)

Chatterbox is licensed under MIT and targets developers. Check out the demo here.

AI in practice
Short

A new website called "The OpenAI Files" launched on June 18, 2025, compiles internal documents and criticism of OpenAI's leadership, strategy, and corporate culture. It raises concerns about whether the company is moving away from its original non-profit goals to serve investor interests. The platform includes reports from former staff, analyses of restructuring plans, and scrutiny of CEO Sam Altman's role. The groups Tech Oversight and The Midas Project of the site say the goal is to spark public discussion about transparency, safety, and oversight in OpenAI's development of powerful AI systems.

Ad
Ad
AI in practice
Short

OpenAI has released a demo of an AI-powered tool for automated front-end testing on GitHub. The tool uses its Computer-Using Agent (CUA) and the Playwright open-source framework to generate, run, and evaluate test cases based on written descriptions. OpenAI says the goal is to make software testing more efficient and reliable. The project is currently a concept study and still in early stages.

Image: OpenAI
Ad
Ad
AI research
Short

ChatGPT lost badly to Atari's 1979 Video Chess engine. It gave solid advice and explained tactics, but it forgot captured pieces, confused rooks and bishops, and lost board awareness turn after turn. Atari's 1.19 MHz engine had no such issues. It just remembered the state and followed rules.

Some critics say Caruso's experiment compares apples and oranges, but it underscores a core weakness of LLMs: ChatGPT didn't lose because it lacked knowledge. It lost because it couldn't remember. Symbolic systems don't forget the board.

"Regardless of whether we're comparing specialized or general AI, its inability to retain a basic board state from turn to turn was very disappointing. Is that really any different from forgetting other crucial context in a conversation?"

Robert Jr. Caruso

Ad
Ad
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

OpenAI's new ChatGPT image model matches Google's Nano Banana Pro on complex prompts

AI research

More AI agents isn't always better, new Google and MIT study finds

AI in practice

GPT-5.2 lands to top Google's Gemini 3 in the AI benchmark game just four weeks after GPT-5.1

Google News