Microsoft Research
·
1 month ago
Microsoft's Project Ire, an AI agent for malware classification, identified a LOTUSLITE backdoor variant that shared behavioral patterns with known samples but had a different hash not flagged by major security vendors. The sample was detected by only 1 of 72 security vendors on May 28, rising to 7 of 70 by June 4, while CrowdStrike Falcon, SentinelOne, Sophos, and others still missed it. Ire's behavior-based analysis caught the variant through function-by-function reverse engineering without relying on signature matching, demonstrating how agentic analysis can identify malware that escapes traditional detection methods.
Google Research
·
1 month ago
Researchers studied how AI tools help consumers understand skin conditions through two studies: a survey of 2,345 participants and a real-world test with 110 diverse community members. AI assistance increased participants' ability to name conditions accurately to 23% versus 8% without AI, though determining appropriate next steps remained challenging. The findings highlight that effective AI health tools require human-centered design, including diverse skin tone representations and personalized guidance beyond basic condition identification.
Deep Learning Weekly
·
1 month ago
Deep Learning Weekly Issue 459 covers major AI model releases including Anthropic's Claude Fable 5, Google's Gemma 4 12B multimodal model, and NVIDIA's Nemotron 3 Ultra with 550B parameters and 1M token context window. Google launched an agentic RAG framework achieving 90.1% accuracy on multi-hop queries, while research papers examine token consumption in multi-agent software engineering systems and foundation models' capacity for active spatial exploration. The newsletter also reports OpenAI's confidential S-1 filing with the SEC and discusses emerging concerns about undisclosed safety filters in frontier AI systems.
OpenAI Blog
·
1 month ago
OpenAI launched three Academy courses designed to teach people how to build AI skills, create repeatable workflows, and implement agents in their work. The courses cover practical applications of AI tools for everyday professional tasks. Participants will gain hands-on experience applying AI to specific work problems rather than learning abstract concepts.
Allen Institute (AI2)
·
1 month ago
Researchers at AI2 released olmo-eval, an evaluation workbench designed for iterative LLM development that extends their earlier OLMES benchmarking standard. The tool enables developers to add benchmarks with minimal code, run evaluations across model checkpoints in flexible ways, and compare results question-by-question to detect real improvements rather than noise. Unlike existing frameworks that evaluate finished models, olmo-eval is built to keep pace with continuous model changes during development, supporting tool use, multi-turn interactions, and modular component swapping.
Platformer
·
1 month ago
Microsoft CEO Satya Nadella shared insights from a conversation at Hard Fork Live event in June 2024. The event featured discussions on AI's impact on productivity, employment concerns from 200 economists and AI leaders, and OpenAI's GPT-5.6 launch alongside executive departures. The coverage explored how AI is changing workflows, job markets, and organizational dynamics across major technology companies.
OpenAI Blog
·
1 month ago
Preply integrated OpenAI's technology to generate automated lesson summaries and personalized feedback within its tutoring platform. The system creates customized language learning exercises based on individual student performance. This allows the company's human tutors to spend more time on direct instruction rather than administrative tasks.
