Claude 3.7 Sonnet is the first hybrid reasoning model, enabling both instant responses and extended thinking, with user control over thinking duration, enhancing coding and web development tasks significantly.
DeepSeek open source DeepEP – library for MoE training and Inference
DeepEP is a specialized communication library designed for Mixture-of-Experts (MoE) and expert parallelism, featuring high-throughput, low-latency GPU kernels and support for low-precision operations like FP8.
AI cracks superbug problem in two days that took scientists years
AI solved a complex superbug problem in just two days, a task that took microbiologists a decade, demonstrating the potential of advanced tools like Google's "co-scientist" to accelerate scientific discovery.
DeepSearcher: A Local open-source Deep Research
DeepSearcher is an advanced open-source research agent that enhances local deep research capabilities by integrating features like query routing and conditional execution flow, allowing for more efficient information retrieval and report generation.
Hard problems that reduce to document ranking
LLMs can effectively solve complex problems by transforming them into document ranking tasks, as demonstrated in locating N-day vulnerabilities through patch diffing, utilizing established ranking techniques from information retrieval.
[R] Training LLMs for Strict JSON Schema Adherence via Reinforcement Learning and Structured Reasoning
A new reinforcement learning approach enables LLMs to achieve 98.7% valid JSON output by using schema validation as a training signal, eliminating the need for extensive datasets.
What do people see when they're tripping? Analyzing Erowid's trip reports
Neuroscientist Sean Noah analyzes 40,000 trip reports from Erowid to explore the visual effects of psychedelics, aiming to redefine how these experiences are categorized and understood.
A possible future of Python in the browser
SPy is an emerging language designed for Python users, enabling native compilation to binaries or WebAssembly, thus enhancing performance significantly over existing solutions like Pyodide and MicroPython.
[R] Muon is Scalable for LLM Training
Muon is an optimizing algorithm that achieves ∼2× computational efficiency over AdamW, saving about 50% FLOPs for training a 1.5B LLM on 39B tokens, demonstrating its scalability for larger models.
[D] AVX512 Inference Performance
AVX512 instruction sets significantly enhance inference performance in frameworks like ONNX Runtime and Llama.cpp, with benchmarks indicating improvements in processing speed and efficiency.
Start building with Gemini 2.0 Flash and Flash-Lite
Gemini 2.0 Flash and Flash-Lite models enhance performance and affordability, offering a 1 million token context window at a simplified pricing structure, making them ideal for diverse applications in AI development.
NVIDIA Earth-2 Features First Gen AI to Power Weather Super-Resolution for Continental US
NVIDIA's Earth-2 platform utilizes the CorrDiff model, achieving up to 10,000 times greater energy efficiency than traditional methods, enabling kilometer-scale weather forecasts that enhance disaster preparedness and save lives.
Magma: A foundation model for multimodal AI agents across digital and physical worlds
Magma is a groundbreaking multimodal AI foundation model that integrates digital and physical task execution, enabling AI agents to interpret user interfaces and orchestrate robotic actions seamlessly across environments.
LLM-QE: Improving Query Expansion by Aligning Large Language Models with Ranking Preferences
LLM-QE enhances query expansion by utilizing Large Language Models to generate document-based expansions, significantly improving the performance of dense retrieval models like Contriever by over 8%.
LongSpec: Long-Context Speculative Decoding with Efficient Drafting and Verification
LongSpec enhances speculative decoding in Large Language Models (LLMs) by introducing a memory-efficient draft model with a constant-sized Key-Value (KV) cache, addressing critical challenges in long-context settings.