AI and society

The long-predicted deepfake dystopia has arrived with Sora 2

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 research AI and society AI in practice
Ad
AI in practice
Short

Anthropic is rolling out a new memory feature for Claude on the Pro and Max plans. With Claude Memory, the system remembers project content, user preferences, and workflows to keep context consistent across conversations. Each project gets its own separate memory, so confidential topics stay isolated. Users can review and edit what Claude remembers at any time, and there's also an "incognito chat" option that doesn't save data or appear in the chat history.

The feature is optional and can be turned on in the settings. Before launch, Anthropic said in a blog post it tested the storage function in sensitive areas, like making sure it wouldn't reinforce harmful conversation patterns or bypass safety features, and made adjustments as needed. The feature has been available for Team and Enterprise users since September.

AI in practice
Short

OpenAI has acquired Software Applications Incorporated (SAI), the team behind the Mac app Sky. Sky is built as an AI language interface for Mac that understands context, adapts to what users want to do, and works directly with other apps. ChatGPT CEO Nick Turley says the aim is to make ChatGPT a more natural part of daily Mac workflows and help users get things done. The entire Sky team is moving to OpenAI to focus on deeper integration of its technology with the Mac. The project lines up with OpenAI's broader push for ChatGPT-powered automation.

The deal was reviewed and approved by independent board committees. Sam Altman, OpenAI's CEO, previously held a passive stake in Sky through an investment company. SAI also created Workflow, the iPhone app that Apple later turned into Shortcuts.

AI in practice
Short

Alibaba is moving into the consumer smart headset space with the Quark AI Glasses, starting at 4,699 yuan (about $660). Pre-orders open on October 24, with deliveries set for December. The glasses use Alibaba's Qwen model and Quark "multimodal AI," providing features like phone calls, music playback, and real-time translation. Alibaba is positioning the device as a direct competitor to Meta's AI smart glasses, and it enters the race with a stronger lineup of AI models compared to Meta's Llama.

Alongside the headset, Alibaba is rolling out a new AI Chat Assistant inside the Quark app. The chatbot supports both text and voice input, and can handle tasks like photo editing and task management. With these launches, Alibaba is positioning itself alongside competitors like Meta, Xiaomi, and OpenAI in the push for consumer AI devices.

Ad
Ad
AI in practice
Short

Reddit set a trap for the AI search company Perplexity to prove that it was scraping content from Google Search.

As part of a lawsuit against several data-scraping firms, Reddit accuses Perplexity of illegally using its content. To back up the claim, Reddit ran a targeted test. According to the lawsuit, Reddit created a "test post" that could "only be crawled by Google’s search engine and was not otherwise accessible anywhere on the internet." Within a few hours, the content from this post appeared in Perplexity's search results, which Reddit says demonstrates that Perplexity was scraping Google's search results.

This incident is just one part of a broader fight, as platforms like Reddit try to block the unauthorized use of their data to train AI models. "A.I. companies are locked in an arms race for quality human content — and that pressure has fueled an industrial-scale ‘data laundering’ economy," said Ben Lee, Reddit's Chief Legal Officer.

AI in practice
Short

OpenAI's new browser, ChatGPT Atlas, could pose security risks, according to the company’s head of security, Dane Stuckey.

One of the biggest issues involves so-called prompt injections. In these attacks, malicious instructions are hidden on websites or in emails to manipulate the AI agent. The effects can range from influencing purchasing decisions to stealing private data such as email contents or login credentials.

Stuckey said OpenAI has run extensive tests, introduced new training methods, and built in protective mechanisms. Still, prompt injection remains an unresolved security challenge. To reduce risks, Atlas includes a "logged out mode" that prevents access to user data and a "watch mode" for sensitive websites, which requires active user supervision. Stuckey added that OpenAI is developing additional security features and faster response systems to handle potential attacks.

X
AI in practice
Short

Apple recently published a widely discussed study revealing major weaknesses in current reasoning models. Now, the company is hiring for exactly this area.

Image: The Illusion of Thinking

A new opening on Apple's MIND team in Seattle targets experts in reasoning, planning, tool use, and agent-based LLMs. Candidates should have experience with large-scale model training, PyTorch, RAG, multimodal models, Jax, and hardware-optimized inference. The goal is to develop new LLM and VLM architectures that are more accurate and efficient.

Ad
Ad
AI in practice
Short

Runway is letting selected pilot partners fine-tune its generative video models for specific use cases like robotics, education, life sciences, and architecture. The company says its standard models often assume general use cases and fail in real-world tasks. The new tool aims to fix this by allowing easier customization with less data and less computing power. Broader access is expected later. Early access is available via the Runway website.

AI in practice
Short

Meta is laying off about 600 employees from its core AI research (FAIR), product AI, and infrastructure teams. According to an internal memo from AI chief Alexandr Wang seen by Axios, the move is meant to cut down on bureaucracy, speed up decision-making, and give teams more responsibility. Thousands of AI jobs will remain, and Wang's new TBD Lab unit is not affected by the cuts and will continue to expand. CEO Mark Zuckerberg reportedly had doubts about the company’s previous approach and is now backing Wang’s team to lead Meta’s AI strategy.

Impacted employees in the US are expected to be notified by Wednesday morning and, according to Wang, will be able to apply for other roles within Meta. Reports of tension in Meta’s AI divisions have surfaced before, including disagreements between Yann LeCun and the new management team.

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

The long-predicted deepfake dystopia has arrived with Sora 2

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

Google News