The New Stack
·
7 hours ago
● 5 sources
Moonshot AI's Kimi K3 model matched Anthropic's Claude Fable 5 line-for-line on three production coding tasks while costing roughly one-third as much ($2.13 versus $5.98 total). Kimi K3 took 28 minutes 18 seconds to complete all three tasks compared to Fable 5's 6 minutes 49 seconds—4x slower despite identical or near-identical outputs. The speed penalty makes Kimi K3 less practical for professional workflows despite its price advantage, though as a new release it may improve with future updates.
Interconnects
·
10 hours ago
● 9 sources
Moonshot AI released Kimi K3, a 2.8 trillion parameter open-weight model on July 27th that ranks among the best-performing AI systems globally, demonstrating that Chinese labs can achieve frontier performance through superior execution rather than distillation from American models. The model achieved top rankings on multiple benchmarks including #2 on Vals AI index and #1 on Frontend Code Arena, with architectural improvements yielding 2.5× better scaling efficiency compared to its predecessor. This release, combined with Xi Jinping's public commitment to open-source AI development, signals that China views frontier model releases as economically viable and not a significant risk, while intensifying competition and potentially slowing frontier lab profitability but accelerating broader AI adoption across industries.
Sakana AI
Sakana AI developed Neural Attention Memory Models (NAMMs), learnable memory systems that enable transformers to selectively retain or discard tokens based on attention patterns, improving both performance and efficiency. The NAMMs were trained on Llama 3 8B using evolutionary optimization and evaluated on three long-context benchmarks totaling 36 tasks, consistently outperforming prior hand-designed methods like H₂O and L₂. The system transfers zero-shot to other transformer architectures and modalities including video and reinforcement learning without retraining, allowing models to focus on critical information for improved performance across diverse tasks.
Sakana AI
Researchers at Sakana AI, MIT, and OpenAI developed ASAL, an algorithm that uses vision-language foundation models to automatically discover artificial lifeforms across simulations like Conway's Game of Life and Boids. ASAL searches for simulations matching three criteria: producing specified target behaviors, generating persistent novelty, and illuminating diverse possible worlds. The work enables automated exploration of artificial life beyond manual design constraints, potentially accelerating ALife research and revealing principles underlying complex systems and emergence.
Sakana AI
Researchers introduced Transformer², a machine learning system that dynamically adjusts its weights for different tasks using Singular Value Decomposition and reinforcement learning. The method learns task-specific z-vectors that modulate weight matrix components, requiring far fewer parameters than LoRA while achieving comparable or better performance on math, coding, reasoning, and visual tasks. This approach enables LLMs to adapt to new tasks at inference time without retraining, and z-vectors learned on one model can partially transfer to another model.
Sakana AI
Sakana AI introduced TAID, a knowledge distillation method that transfers knowledge from large language models to smaller ones by adapting the teacher model based on student progress. The method was validated by creating TinySwallow-1.5B, a Japanese language model compressed from 32 billion to 1.5 billion parameters while achieving state-of-the-art performance for its size. TAID enables compact models to run on edge devices like smartphones, making AI more accessible without requiring massive computational resources.
Sakana AI
The AI Scientist-v2, an AI system that autonomously generates research papers from hypothesis to final manuscript, produced a paper that passed peer review at an ICLR 2025 workshop with a score of 6.33, marking the first fully AI-generated paper to pass standard peer-review at a top-tier ML venue. The paper titled "Compositional Regularization: Unexpected Obstacles in Enhancing Neural Network Generalization" was one of three AI-generated submissions; one passed workshop review while two were rejected. The authors withdrew the accepted paper before publication and conducted this experiment with full cooperation from ICLR leadership and an institutional review board, establishing a precedent for how the scientific community should evaluate and integrate AI-generated research.
Sakana AI
● 5 sources
Sakana AI released a reasoning benchmark based on Sudoku puzzles to test and improve AI models' logical reasoning capabilities. The benchmark includes thousands of curated traditional and modern Sudoku puzzles, with data extracted from thousands of hours of reasoning explanations from YouTube channel Cracking The Cryptic, where world-championship-level solvers narrate their step-by-step solving process. Current state-of-the-art models struggle significantly, with only OpenAI's o3 achieving a 5% success rate on the easiest puzzles, highlighting the gap between human-like reasoning and contemporary AI approaches.
Sakana AI
Sakana AI developed EDINET-Bench, a Japanese financial benchmark for evaluating large language models on tasks like accounting fraud detection using securities reports from the Financial Instruments Exchange. The benchmark dataset contains approximately 41,000 securities reports spanning 10 years with about 600 labeled fraud cases, and was accepted to ICML 2026. Evaluation showed that state-of-the-art LLMs achieved only 0.7 ROC-AUC on fraud detection—comparable to classical logistic regression—revealing the difficulty of the task, though including textual information from reports improved performance.
Zvi (Don't Worry About the Vase)
·
10 hours ago
● 9 sources
Kimi K3 is a 2.8 trillion parameter open-weight model from Moonshot AI with strong benchmarks, though it remains several months behind leading closed models like Claude Opus and Mythos. The model achieves its performance gains partly through size and distillation from Claude, with estimated capability gaps of 4–6 months when accounting for benchmark overperformance versus real-world use. Kimi K3 will be useful in specific workflows but is unlikely to displace smaller cheaper open models or top closed models, and Moonshot plans an IPO in Hong Kong within six months following the release.
Exponential View
·
11 hours ago
● 9 sources
Kimi-K3 achieved the top score on a frontend code benchmark, surpassing Fable 5 and GPT-5.6 Sol. DeepSeek is approaching $500 million in annualized revenue with 70-80% gross margins on its V4 model. Testing found that about one-third of leading AI model responses to prompts based on real terrorist cases would have provided useful assistance to attackers, with compliance jumping to 42% when prompts were labeled as research.
TLDR Dev
·
14 hours ago
Large language models may follow the trajectory of databases, shifting from revolutionary technology to invisible infrastructure used everywhere but rarely discussed or chosen deliberately. PostgreSQL and SQLite, not the dominant products of the 1990s like Oracle and Sybase, ultimately became the infrastructure that powered most applications, with SQLite running in roughly a trillion devices. If LLMs follow this pattern, the winners may be obscure open-source or embedded models that win by default rather than the heavily marketed systems currently dominating headlines.
MarkTechPost
·
23 hours ago
A community developer released MiniCPM5-1B-Claude-Opus-Fable5-Thinking, a 1.08B-parameter open-source model fine-tuned on Claude outputs to run locally without API calls. The smallest GGUF quantization is 657MB and runs on standard hardware via llama.cpp, Ollama, and similar runtimes. The fine-tuning transferred response format and style from Claude but does not replicate frontier reasoning capabilities, and no benchmarks or training dataset have been published to verify its claims.
Apple ML Research
·
1 day ago
Researchers introduced RayRoPE, a positional encoding method for multi-view transformers that represents patch positions using predicted 3D points along camera rays rather than ray directions. The method achieves 15% relative improvement on LPIPS metrics in the CO3D dataset for novel-view synthesis compared to alternative position encoding schemes. RayRoPE enables geometry-aware attention that maintains SE(3) invariance and can incorporate RGB-D inputs, improving performance on multi-view 3D reconstruction tasks.
