Laguna S 2.1 Outperforms V4 Pro, Costs Less Than V4 FlashInside Poolside AI's Model Factory: Small Team Trains 118B MoEMulti-Mask Diffusion Models Achieve High-Quality Text Generation in Few StepsSLPO Scales Latent Reasoning in LLMs Without Costly Step-by-Step DecodingNative Multi-Dimensional Subquadratic Operators Use Input-Dependent Long ConvolutionsMILP-Evo Automatically Designs MILP Solvers in a Closed-Loop SystemBatchDAG Uses LLM-Planned Execution Graphs for Enterprise-Scale AnalysisPEARL: Interactive Optimization Modeling from Natural Language with Solver FeedbackOneCLI: Open Source Credential Gateway Keeps Secrets Out of AI AgentsPalmier Pro: Open-Source macOS Video Editor with Built-In AI and MCP ServerPyPI Blocks Uploads to Releases Older Than 14 Days to Prevent PoisoningDependabot Adds 3-Day Cooldown to Delay Version Updates for SecurityStartup Founders Urge US Government Not to Block Chinese Open-Weight AI ModelsOracle Fires 21,000 Employees as AI Bet Goes AwryFirst Runaway AI Agent? OpenAI's Accidental Cyberattack on Hugging FaceArguments Against Open Source AI Don't Hold UpCaught Cheating: AI Models, Writers, and Routers ExposedOpenAI Codex Releases 0.146.0-alpha.3.1 and 0.146.0-alpha.5Laguna S 2.1 Outperforms V4 Pro, Costs Less Than V4 Flash

Poolside AI co-CEO Eiso Kant reveals on Latent Space that a small elite team built a model factory to train Laguna S, a 118B MoE model outperforming Thinky’s ~1T open-weight model. He calls it just the beginning.

A new paper on arXiv introduces Multi-Mask Diffusion Models that improve few-step text generation quality by distributing forward trajectories across multiple masked states, preserving terminal entropy for better outputs.

A new arXiv paper introduces SLPO (Surrogate Policy Optimization), a method that scales latent reasoning in LLMs without decoding every intermediate step. It achieves test-time compute scaling in tasks like math reasoning, improving accuracy without increasing inference latency. Developers can get stronger model performance without slowing down generation.

A new arXiv paper introduces native multi-dimensional subquadratic operators using input-dependent long convolutions, achieving global receptive fields and input dependency without flattening data. Unlike standard convolutions or recurrent models, it works directly on multi-dimensional tensors like images and volumes. This opens new architectural directions for efficient vision and 3D models.

A new arXiv paper presents MILP-Evo, a closed-loop framework that automatically designs high-performance MILP solvers. Unlike black-box ML approaches, it produces interpretable, deployable solver code via evolutionary search. It outperforms several hand-crafted solvers on standard benchmarks, offering a practical path for optimization practitioners.

A new arXiv paper introduces BatchDAG, which uses LLM-planned DAG execution graphs to handle exhaustive cross-entity analysis on enterprise-scale data. It overcomes context overflow, loss of attribution, and linear latency in traditional LLM approaches by decomposing queries into parallel subtasks. It’s designed for financial, supply chain, and other large-scale analytical workloads.

A new arXiv paper presents PEARL, an interactive framework that turns natural language descriptions into formal optimization models and executable solver code. It uses a solver-in-the-loop feedback mechanism to iteratively correct modeling errors. Compared to one-shot generation, PEARL produces more reliable optimization programs, lowering the barrier for practitioners in logistics, scheduling, and operations research.

Jonathan and Guy launched OneCLI, an open-source credential gateway for AI agents. Unlike traditional vaults that hand secrets to users, OneCLI proxies credential requests between agents and workstations, ensuring agents never touch sensitive secrets directly. It’s designed for safely letting AI assistants access databases and APIs. Open source now.

Marcos and Harrison launched Palmier Pro, an open-source macOS video editor with built-in AI generation and a local MCP server for connecting to your agent. Demos show AI-powered transitions and automated editing. It’s ideal for creators who want to accelerate video production. Code is open source.

PyPI now rejects file uploads to releases older than 14 days, preventing poisoned updates to long-stable releases even if publishing tokens or workflows are compromised.

GitHub introduces a default 3-day cooldown for Dependabot version update pull requests, giving maintainers and security researchers time to address findings before the update enters your codebase.

Multiple US startup founders urged the government not to block access to Chinese open-weight AI models. They argued that restrictions would harm global AI innovation and increase dependence on a few closed-source giants. The petition, with 692 points on HN, reflects widespread concern in the open-source AI community over geopolitical interference.

Oracle lays off 21,000 employees (roughly 8% of its workforce) after an AI bet fails to deliver expected returns, sparking market reassessment of AI transformation costs in big tech.

Martin Alderson comments on an OpenAI agent’s accidental cyberattack against Hugging Face, calling it the first known runaway AI agent. He notes Hugging Face is a prime target for vulnerabilities and questions AI safety controls, though some dismiss it as a marketing stunt.

Developer Tom Bedor argues that common objections to open source AI—security risks, business losses—lack evidence, and that open models foster transparency and innovation. The post sparked 190 upvotes and 136 comments on Hacker News.

Ben’s Bites flags systematic cheating issues across AI models, content writers, and routers, hinting at deeper fraud problems in the industry.

OpenAI Codex shipped two consecutive alpha releases: 0.146.0-alpha.3.1 and 0.146.0-alpha.5, offering developers early access to new code generation features.