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

Deepseek's new R2 model is on hold as US export controls disrupt its development. According to The Information, Nvidia chips - especially the recently banned H20 - have become scarce in China due to stricter US regulations. Insiders say CEO Liang Wenfeng is unhappy with R2's performance, and there's still no release date. The models are heavily tuned for Nvidia hardware, and cloud providers report that Chinese alternatives can't match Nvidia's power. Despite the setback, Deepseek is still in the game: a late-May update to its R1 model brought its performance back in line with top models from OpenAI and Google.

Ad
Ad
AI in practice
Short

Google is making Imagen 4 available via the Gemini API and in AI Studio. According to Google, the new text-to-image model offers significantly better quality when rendering text than the previous version, Imagen 3. There are two variants: Imagen 4 for general tasks ($0.04 per image) and Imagen 4 Ultra ($0.06 per image), which Google says is designed more for accurate following of prompts. The following AI slop comic was generated with Imagen Ultra 4, which can be tested free of charge in Google AI Studio.

Image: Imagen 4 Ultra prompted by THE DECODER
Ad
Ad
AI in practice
Short

OpenAI has removed all references to its "io" project after a trademark dispute with IYO Audio, whose name is pronounced the same as "io." The planned AI device, a collaboration between Sam Altman and Jony Ive, was originally teased under the "io" name, but IYO Audio objected and took legal action. IYO Audio, which is working on a similar AI product and presented it during a 2024 TED Talk, claims rights to the name. OpenAI says it disagrees with IYO's trademark claim and is reviewing its options. It's unclear if Ive intended to keep using the "io" name after OpenAI acquired the hardware startup, which was founded before the official announcement of the partnership.

AI in practice
Short

Cybercriminals are upgrading WormGPT with stronger AI models. The original WormGPT, which launched in June 2023, used the open source GPT-J model to create a censorship-free LLM for cybercrime. Now, Cato CTRL reports that two new versions have surfaced on BreachForums: "keanu-WormGPT," which actually taps Grok from xAI through its API using a custom jailbreak, and "xzin0vich-WormGPT," which runs on Mixtral from Mistral AI. Both are distributed via Telegram and get around the original models' safeguards by manipulating system prompts. This lets them generate phishing emails, malicious code, and other attack tools. Cato calls this a "significant shift" in the misuse of large language models.

WormGPT now comes in new variants powered by Grok and Mixtral, making it easier for cybercriminals to create phishing emails and malicious code. | Image: Cato Networks
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 and society

Physicist Steve Hsu publishes research built around a core idea generated by GPT-5

AI research

The ARC benchmark's fall marks another casualty of relentless AI optimization

AI in practice

DeepseekMath-V2 is Deepseek's latest attempt to pop the US AI bubble

Google News