NerdNews

MIT researchers developed the SEAL framework, which allows large language models to continuously learn and adapt by generating their own training data and updating their internal parameters. This framework uses reinforcement learning to teach models to create 'self-edits' that specify how to update their weights, enabling them to absorb new knowledge and learn new tasks. SEAL has shown promising results in knowledge incorporation and few-shot learning, and could be useful for enterprise applications, particularly in dynamic environments.

The rise of AI-generated job applications is disrupting the hiring process, with employers struggling to identify genuine candidates. AI tools can generate hundreds of customized applications, making it difficult for companies to find qualified applicants. This has led to an arms race between job seekers and employers, with both sides using AI to gain an advantage. As a result, the traditional résumé may become obsolete, and new methods of hiring, such as live problem-solving sessions or portfolio reviews, may become more prevalent.

The AI industry is shifting towards developing more emotionally intelligent language models, with companies like LAION and OpenAI making significant progress. This change is driven by the need for AI models to understand and respond to human emotions, making them more relatable and useful. However, this also raises concerns about the potential for manipulative behavior and unhealthy emotional attachments.

A federal judge has ruled in favor of Anthropic, stating that training AI models on published books without authors' permission is legal under fair use doctrine. This decision may set a precedent for other AI companies facing similar lawsuits, but the court will still hold a trial on the issue of pirated copies used to create Anthropic's central library.

Transformers often hallucinate due to prioritizing statistically likely text over factual accuracy. State Space Models (SSMs) offer a reliable alternative, processing information sequentially and maintaining factual accuracy and context. SSMs are more efficient and less prone to hallucinations, making them suitable for tasks like legal document review, medical research, and robotics.

Anthorpic's Claude AI is adding a memory feature to compete with ChatGPT, allowing it to remember past events and reference them in new chats, and also enabling users to embed Claude functionality within its Artifacts.

Google Cloud has donated its Agent2Agent (A2A) protocol to the Linux Foundation, enabling secure communication between AI agents from different vendors. The protocol allows AI agents to discover each other, share capabilities and context, and collaborate on complex tasks.

A new report by Cobalt found that 48% of security professionals believe a strategic pause in generative AI deployment is needed to recalibrate defenses. The report highlights top GenAI security vulnerabilities, including AI data security and accuracy risks. Despite concerns, 33% of security teams are failing to conduct regular security assessments for their large language model deployments.

Eventual, founded by former Lyft engineers, created a Python-native open source data processing engine called Daft to solve multimodal data infrastructure issues, now used by companies like Amazon and CloudKitchens.

Ying Dai shares insights from her time at Roblox, detailing two pivotal migrations that transformed the company's engineering landscape, prioritizing reliability and developer productivity, and significantly reducing incidents and enhancing the engineer experience.

Synthflow AI, a Berlin-based startup, is making waves in the conversational AI market with its no-code platform for building customized white-labeled voice AI customer service agents. The company has amassed over 1,000 customers, handled 45 million calls, and raised a $20 million Series A round. Its voice agents are HIPAA and GDPR compliant and integrate with over 200 enterprise platforms.

Abridge, an AI medical scribe startup, has raised $300 million in Series E funding at a $5.3 billion valuation. The company's AI technology automates medical notes and is used by over 150 of the largest health systems in the US. Abridge is expanding its services to include converting medical notes into AI-powered medical codes.

Google DeepMind released a new language model called Gemini Robotics On-Device that can run tasks locally on robots without an internet connection. The model can control a robot's movements and perform tasks like unzipping bags and folding clothes. Google claims it outperforms other on-device models in general benchmarks.

Tesla has launched its Robotaxi service in Austin, Texas, with a limited fleet of around 10 rebadged Model Y Teslas. The service is geofenced to a specific area of Austin and only operates during daylight hours. Early videos show glitches, such as the car stopping in the middle of the street or blocking a crosswalk. In contrast, Waymo has been operating commercial self-driving car services in San Francisco and Austin since 2023, with over 10 million safe trips. The National Highway Traffic Safety Administration (NHTSA) is investigating reports of Tesla Robotaxis making mistakes.

Andreessen Horowitz has invested in Cluely Inc., an AI startup with a motto of 'Cheat at Everything'. The company offers a desktop app that helps users generate smart responses in real-time, potentially allowing them to 'cheat' in meetings and conversations. The investment was part of a $15 million funding round.

T-Mobile's satellite service, powered by Starlink, will launch on July 23rd with messaging capabilities on Android and iOS. The service will be available to anyone in the US, with data capabilities coming on October 1st. Partners include AllTrails, Accuweather, and WhatsApp.

Trezor's automated support system is being abused by attackers to send phishing emails from the company's official address. The emails contain urgent messages with links to phishing sites, aiming to steal users' wallet seeds. Trezor is working to implement defenses against similar attacks and advises users not to share their wallet seeds with anyone.

ElevenLabs has released a standalone mobile app for iOS and Android users to generate voice clips from text using its AI-powered voice libraries. The app offers a free plan with 10 minutes of audio generation and access to the company's newest text-to-speech models. This move marks ElevenLabs' second consumer-facing app, following the release of the Reader app last year.

Google announced that historical Street View imagery is now available on Google Earth, allowing users to explore and view changes over time. This update commemorates Google Earth's 20th birthday and provides new AI-driven insights for professional users.

NVIDIA's RTX 5050 GPU has arrived early in laptops from manufacturers like Acer and MSI, featuring 2,560 CUDA cores, 8GB of GDDR7 VRAM, and a TDP of 115W. The laptop version has a memory bandwidth of 384 GB/s, comparable to the 5060 mobile. Performance is expected to be between the 4050 and 5060, with decent generational gains.