NerdNews

Meta has paused its work with Mercor, a data contracting firm, after a major security breach exposed sensitive AI industry secrets. The breach may have compromised key data about how AI labs train their models, including those used by OpenAI and Anthropic. Other major AI labs are also reevaluating their work with Mercor as they assess the scope of the incident.

Anthropic's accidental release of Claude Code's source code has significant security implications and raises concerns about the company's IPO. The leak, which occurred on March 31, exposed over 512,000 lines of code, allowing developers to understand how the AI software development assistant works. Experts discuss the implications of the leak on The Kettle podcast.

As AI agents become more prevalent in automating business decisions, the question of liability arises. Vendors claim their AI agents can 'actively run the business,' but who is responsible when things go wrong? The answer is unclear, with vendors pushing back on accepting liability due to the unpredictable nature of AI. Businesses may need to accept the risk themselves or negotiate contractual provisions to hold vendors accountable.

Speakers from Netflix, Meta, and IBM discussed the challenges of using AI, including the need for extensive preparation and cleanup of results. They emphasized the importance of context engineering and decomposition to achieve optimal AI results.

PrismML's 1-bit large language model, Bonsai 8B, is 14x smaller and 5x more energy efficient than other models, making it suitable for mobile devices and edge hardware. The model achieves competitive performance with other 8B models while reducing memory usage and power consumption.

Researchers have coined the term 'cognitive surrender' to describe how people are willing to accept AI-generated answers without critically evaluating them, even when they are incorrect. A study found that participants accepted incorrect answers from an AI chatbot 80% of the time, and those who used AI rated their confidence 11.7% higher than those who didn't.

The author tested Gemini in Google Maps by having it plan a day-long itinerary. Gemini suggested various locations, including a taco shop, a park, and a coffee shop. While it made some mistakes, the author found the experience to be surprisingly well-planned and enjoyable. Gemini's ability to search user reviews and provide accurate directions made it a useful tool.

A 2017 ThinkPad laptop was revived by installing Linux Mint, giving it a new life and saving it from obsolescence. The laptop, which was unable to run Windows 11 due to its outdated processor, was able to run smoothly on Linux and perform everyday tasks without issues. This approach can be a cost-effective solution for those who cannot afford a new laptop or want to reduce electronic waste.

The article discusses the importance of designing interfaces for human-agent interaction, highlighting the limitations of chat-based interfaces and the need for more structured interactions. It introduces six principles for designing human-agent interactions, including agent disclosure, native platform inhabitation, instant feedback, transparency, respect for disengagement, and accountability. The article argues that unreliable AI products are a design problem, not a model problem, and that designers should focus on creating interfaces that bring structure to AI interactions.

The shift to hybrid work has increased the attack surface, making it harder to detect threats. Security leaders must balance protection and performance, and this webinar explores practical frameworks and technologies to secure the hybrid workplace, including enforcing consistent policies, improving threat detection, and simplifying access management.

Microsoft's Copilot AI terms of use warn users not to rely on it for important advice, stating it's 'for entertainment purposes only' and may make mistakes. The company plans to update this 'legacy language' in the future.

OpenAI has acquired TBPN, a tech-focused talk show, in a deal worth hundreds of millions of dollars. The show will remain editorially independent and based in Los Angeles, with its hosts reporting to OpenAI's head of global affairs. The move is seen as an expansion of OpenAI's reach into the tech community, despite the company's previous pledge to focus on its core business.

Hackers are posting the Claude code leak with bonus malware, and the FBI says a recent hack of its wiretap tools poses a national security risk. Attackers also stole Cisco source code as part of an ongoing supply chain hacking spree.

Netflix has developed an AI video editor called VOID, which can remove objects from a scene and generate video depicting the physically plausible path of the remaining objects. This model can be used to edit videos without requiring reshoots or extensive computer graphics work. VOID outperforms other alternatives and is available on Hugging Face for anyone to install.

Utah has allowed an AI system to prescribe psychiatric drugs without a doctor. The AI chatbot, developed by Legion Health, can renew certain prescriptions for psychiatric medications. However, some psychiatrists are raising concerns about the safety and effectiveness of the system, citing the potential for over-treatment and the lack of human oversight.

Microsoft's Copilot AI terms of use warn users not to rely on it for important advice, stating it's 'for entertainment purposes only' and may make mistakes. The company plans to update this 'legacy language' in the future.

Research finds that AI users are willing to abandon logical thinking and accept AI answers without scrutiny, even when they are faulty. This phenomenon, known as 'cognitive surrender', is more likely to occur when AI outputs are delivered fluently and confidently. The study suggests that users who score highly on fluid IQ are less likely to rely on AI for help and are more likely to overrule faulty AI answers.

Suno's AI music platform can generate covers of popular songs that are alarmingly close to the originals, potentially violating copyright laws. The platform's filters can be easily bypassed, allowing users to create and monetize unauthorized covers. This poses a significant challenge for artists and the music industry, highlighting the need for more effective measures to combat AI-generated copyright infringement.

The rise of AI-generated content has led to a demand for a label that distinguishes human-made works from those created by machines. However, there is currently no universally recognized standard, with at least 12 different alternatives available, each with its own eligibility criteria and authentication approaches.

A 2017 ThinkPad laptop was revived by installing Linux Mint, giving it a new life and saving it from obsolescence. The laptop, which was unable to run Windows 11 due to its outdated processor, was able to run smoothly on Linux and perform everyday tasks without issues. This approach can be a cost-effective solution for those who cannot afford a new laptop or want to reduce electronic waste.