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

Nov 11, 2025Nov 11, 2025

Update

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

AI in practice

Oct 31, 2025Oct 31, 2025

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

AI and society

Oct 19, 2025

The long-predicted deepfake dystopia has arrived with Sora 2

AI in practice

Sep 9, 2025Sep 9, 2025

Update

Google Veo 3 adds 1080p, 9:16 video, and drops prices by half

AI in practice

Sep 9, 2025Sep 9, 2025

Anthropic's Claude now lets users edit Word, PowerPoint, Excel, and PDFs in chat

AI in practice

Sep 8, 2025Sep 8, 2025

Short

Google has added new reporting tools to NotebookLM. Users can now generate structured reports in more than 80 languages and adjust the tone, style, and structure as needed.

Watch the video: Google

The update also includes a blog post format and dynamic suggestions for report types based on the uploaded material. For example, NotebookLM might recommend a white paper format for research documents. Users can also write their own prompts, up to 1,000 words, to control the tone, style, and format of the generated content.

NotebookLM via X Simon via X

AI in practice

Sep 8, 2025Sep 8, 2025

Short

Anthropic backs California's SB 53, a state bill that would force large developers of advanced AI to be more transparent and secure—apparently because they see Washington as too slow to act. Anthropic says SB 53 could serve as a solid starting point for national rules.

"While we believe that frontier AI safety is best addressed at the federal level instead of a patchwork of state regulations, powerful AI advancements won’t wait for consensus in Washington."

Anthropic

The bill would require affected companies to publish security policies, disclose risk analyses, report security incidents within 15 days, share internal assessments under confidentiality, and follow clear whistleblower protection rules. Violations could mean fines. The rules target only companies running highly capable models, aiming to keep the burden off smaller providers. Anthropic says their decision comes after reflecting on the lessons from California's failed SB 1047 effort.

Anthropic

AI in practice

Sep 8, 2025Sep 8, 2025

Short

OpenAI CEO Sam Altman has picked up on a trend: "it seems like there are really a lot of LLM-run twitter accounts now." No punchline here.

The Dead Internet Theory is a conspiracy theory that claims the internet is no longer driven by real people, but mostly by bots and AI-generated content. According to this idea, most online activity—comments, posts, and articles—is fake, created to manipulate public opinion and control users.

Altman via X

AI and society

Sep 7, 2025Sep 7, 2025

Short

LLM hype critic Gary Marcus argues in a conversation with chess grandmaster Garry Kasparov that large language models only create the appearance of understanding, not genuine intelligence.

"It's one of the most profound illusions of our time that most people see these systems and attribute an understanding to them that they don't really have."

Gary Marcus

He explains that while language models can, for example, mimic the rules of chess by generating text based on examples, they can't actually play the game because they lack any real internal sense of what's happening on the board.

"They will repeat the rules, because in the way that they create text based on other texts, they'll be there. [...] But when it actually comes to playing the game, it doesn't have an internal model of what's going on."

For Marcus, this gap between surface-level performance and true comprehension is at the heart of the AI "illusion of intelligence."

Podcast "Autocracy in America"

AI in practice

Sep 7, 2025Sep 7, 2025

Short

Alibaba unveils Qwen3-Max-Preview, its largest language model yet, featuring more than one trillion parameters. The model is available through Qwen Chat and the Alibaba Cloud API. According to Alibaba, Qwen3-Max-Preview outperforms the previous top model, Qwen3-235B-A22B-2507, in internal benchmarks and with early users. Improvements show up in knowledge, conversation, task handling, and instruction following, with reduced "model knowledge hallucinations."

Qwen3-Max-Preview accepts up to 258,048 input tokens and generates up to 32,768 output tokens. Pricing starts at $2,151 per million input tokens and $8,602 per million output tokens. The model does not support image processing.

via X

AI and society

Sep 6, 2025Sep 6, 2025

Short

Apple is facing a lawsuit in California from authors Grady Hendrix and Jennifer Roberson, who claim the company violated their copyrights by using their books to train AI models like OpenELM and Apple Intelligence. The lawsuit alleges Apple used the Books3 dataset, a collection of more than 196,000 pirated books that includes works by both authors. The complaint also accuses Apple of using its Applebot web crawler to copy website content and pull material from so-called shadow libraries.

The plaintiffs are seeking damages and a court order barring Apple from using their works. This case follows a recent lawsuit against Anthropic, which ended in a settlement after similar copyright claims.

Class Action Lawsuit