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 in practice
Update

US Copyright Office says fair use does not cover AI trained on "vast troves of copyrighted works

AI research

OpenAI's o3 is less AGI than originally measured

AI research

The next leap in AI depends on agents that learn by doing, not just by reading what humans wrote

Ad
AI in practice
Short

Meta is splitting its AI department into two groups: "AI Products" led by Connor Hayes, focused on the Meta AI Assistant and features inside Facebook, Instagram, and WhatsApp, and "AGI Foundations" led by Ahmad Al-Dahle and Amir Frenkel, working on Llama models and advanced reasoning and multimedia models. FAIR, Meta's AI research lab, will continue, though one multimedia team is moving under the new structure. Meta says the change should accelerate product development and give teams more freedom. No layoffs are planned.

The reorganization comes as Meta faces strong competition from Deepseek in open-source models. Llama 4 has not met performance expectations, and its largest version has been delayed.

Ad
Ad
AI in practice
Short

The full system prompt for Claude 4 has been leaked by X user "Pliny the Liberator" and is now available on GitHub. The document, over 60,000 characters long, sets detailed rules for tone, roles, source handling, and banned content. It controls the model at the system level, before any user prompt is processed. I find it strange that large language models often fail to follow short user instructions, yet seem able to follow complex internal prompts like this one. If you know the reason for that, shoot me an email.

AI and society
Short

Deepmind CEO Demis Hassabis told the New York Times’ Hard Fork podcast that coding and science are still worth learning, even as AI becomes more powerful. He said people who understand basics like math and programming will be better prepared for the changes AI will bring over the next ten years. While AI tools can boost users to near "superhuman" ability, Hassabis believes "learning to learn" is an important skill.

"I think whatever happens with these A.I. tools, you'll be better off understanding how they work and how they function and what you can do with them. "

Demis Hassabis

Ad
Ad
AI in practice
Short

Nvidia will launch a cheaper AI chip for China in June, Reuters reports. The new GPU, based on Blackwell architecture and the RTX Pro 6000D design, is expected to cost $6,500 to $8,000—less than the $10,000 to $12,000 H20 model. Three people familiar with the plan said the chip avoids advanced packaging from TSMC and uses standard GDDR7 memory, making it easier to produce. The move is a response to U.S. export rules that limit the sale of high-end chips to China.

Ad
Ad
AI in practice
Short

Google's latest video AI model, Veo 3, can generate realistic videos with audio and mimic a range of styles—including the look and feel of video games. While the results aren't interactive like a real game, they do a convincing job of visualizing gaming concepts, whether from a first-person or third-person perspective. The technology doesn't come cheap: according to developer fofrAI, generating a video with Veo 3 costs about $0.75 per second. Veo 3 is currently available in the US for Google AI Ultra subscribers at $249 per month.

Video: fofrAI via X

Google News

Bank details

IBAN: DE87 1203 0000 1086 0070 75
Account holder: DEEP CONTENT GbR
Purpose: Support THE DECODER
AI in practice

Anthropic releases Claude 4 with new safety measures targeting CBRN misuse

AI in practice

OpenAI launches Codex: Autonomous AI agents for software development

AI research

AlphaEvolve is Google DeepMind's new AI system that autonomously creates better algorithms

Google News