GPT-5: Key characteristics, pricing and system card
GPT-5 introduces a hybrid model system that dynamically selects between various models based on task complexity, enhancing user experience with improved accuracy and competence.
Hopfield Networks Is All You Need (2020)
The modern Hopfield network introduces continuous states and an update rule that allows for the storage of exponentially many patterns with minimal retrieval errors, enhancing its utility in deep learning architectures.
Running GPT-OSS-120B at 500 tokens per second on Nvidia GPUs
SOTA performance for GPT OSS 120B was achieved on NVIDIA GPUs, reaching 500+ tokens per second through a combination of flexible inference stacks and rapid engineering iterations, enhancing user experience from launch day.
Breaking the sorting barrier for directed single-source shortest paths
A new algorithm developed by researchers breaks the longstanding sorting barrier in shortest-path computations, outperforming classic methods like Dijkstra's by avoiding sorting entirely.
Benchmark Framework Desktop Mainboard and 4-node cluster
Benchmarking of the Framework Desktop with an AMD Ryzen AI Max+ 395 and Radeon 8090S was conducted using both single-node (128 GB RAM) and four-node (512 GB RAM) configurations, revealing performance metrics across various interconnects including 2.5 Gbps Ethernet and Thunderbolt 4.
OpenAI's new GPT-5 models announced early by GitHub
OpenAI's GPT-5 models will feature enhanced agentic capabilities and excel in complex coding tasks with minimal prompting, as revealed in a now-deleted GitHub post.
Open music foundation models for full-song generation
YuE is an innovative family of open foundation models designed for long-form music generation, capable of producing up to five minutes of music while ensuring lyrical coherence and engaging melodies, utilizing advanced techniques like track-decoupled next-token prediction and structural progressive conditioning.
A generic non-invasive neuromotor interface for human-computer interaction
The generic non-invasive neuromotor interface developed utilizes surface electromyography (sEMG) to decode user intent, achieving high performance in gesture detection and handwriting without requiring individual calibration, demonstrating a median performance of 20.9 words per minute in handwriting tasks.
Sweatshop Data Is Over
High-quality data is essential for advancing AI, necessitating a shift from low-skill contractors to full-time specialists who can create interactive environments for complex learning tasks.
Show HN: Browser AI agent platform designed for reliability
Notte is a full-stack framework that enables the rapid development of AI agents for web automation, achieving cost savings of over 50% while enhancing reliability through a hybrid approach of scripting and AI.
Show HN: Aura – Like robots.txt, but for AI actions
AURA (Agent-Usable Resource Assertion) revolutionizes AI interaction with the web by providing a standardized aura.json manifest that allows websites to explicitly declare their capabilities, moving away from unreliable screen scraping and DOM manipulation.
[R] LLMs Have a Heart of Stone: Demystifying the Soft Thinking Ability of Large Reasoning Models
Soft tokens underperform traditional "hard" tokens in LLMs, but the Gumbel-Softmax trick can enhance their effectiveness by introducing necessary randomness.
Sculptor: Empowering LLMs with Cognitive Agency via Active Context Management
Sculptor introduces Active Context Management (ACM) tools that empower LLMs to manage their internal memory, enhancing their ability to focus on relevant information while filtering out distractions.
How AI is helping advance the science of bioacoustics to save endangered species
AI's Perch model enhances bioacoustic analysis, enabling conservationists to process vast audio data from diverse species, including birds and marine life, with improved accuracy and adaptability to various environments, particularly underwater ecosystems.
[D] FP4 training methods (request for paper recommendations)
OpenAI's OSS models utilize low precision weights (MXFP4), raising questions about the feasibility of training with such limited precision and the potential for even lower formats like FP3 or FP2.