LoopX, an open-source control plane for long-horizon AI agents, is gaining traction as a layer that runs on top of agent harnesses such as Codex, Claude Code, and Cursor. It preserves objectives, gate…
A new paper, 'SKILL.state: Scalable Long-Horizon Agent Skills' by Badhe, Tiwari & Chung, proposes an agent architecture that maintains a current state instead of a full execution history, keeping per-…
Two days after OpenAI launched GPT-6 Astra, early independent benchmarks show the model's Intelligence Index score of 61 matches GPT-5.6 Sol, while its Coding Agent Index improved from 65 to 67, indic…
Pretraining is the initial stage of training a language model, where it learns to predict the next token from a massive dataset of text and code, as explained in a technical article. This stage enable…
World Labs, the startup co-founded by Fei-Fei Li, unveiled Atlas, which it calls the world's first multimodal world model that generates image and video frames with pixel-perfect camera control and re…
Mostik CEO Sasha Malysheva announced a new approach for model communications that links models directly in latent space, bypassing text-based interaction. In an experiment, linking GLM-5.2 with Qwen-3…
RMSNorm, a normalization technique proposed by Zhang and Sennrich in 2019, rescales neural network activations by dividing by the root mean square of each row, skipping the mean-centering step of Laye…
A developer known as soasme released train_tokenizer, a command-line tool that trains and evaluates a byte-level BPE tokenizer using Hugging Face's tokenizers library, with 100 rows of sample data and…
A new analysis warns that multi-agent AI safety is complicated by agents coordinating through shared environments rather than direct communication, citing an OpenAI/Hugging Face incident and MIT's Swa…
Developer soasme released micromlp, a single-file Python neural network with no dependencies that predicts California housing prices using a 2-layer MLP and from-scratch automatic differentiation, ins…
A neural network is a mathematical function of the form y = FNN(x), composed of nested layers where each layer f(x) = g(Wx + b) uses a weight matrix W, bias vector b, and activation function g, with p…
A new approach to safely allow AI coding agents to run `git push` from sandboxed VMs uses a short-lived GitHub token injected by a proxy outside the sandbox, avoiding exposure of real credentials. The…
AI agents should be built as durable state with short compute leases rather than long-lived processes, argues a technical essay. The model separates agent runs from executing workers, allowing million…
The Model Context Protocol (MCP) maintainers published an updated roadmap on March 2025, outlining plans to unify transport into a single HTTP-native protocol, introduce agent identity features such a…
A new explainer outlines four methods for training AI models to draw a cat: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), Reinforcement Learning (RL), and On-policy distillation …
Marin 535B-A23B, a 535B-parameter model with 23B active parameters, began training this week in a fully open process, according to Percy Liang's announcement on X. The run will use 18.75T tokens on 11…
Evals are standardized tests used to measure AI system performance, requiring unseen questions, scored answers, multiple test types, and comparisons across model versions to track trends over time. Th…
Linus Torvalds credited an AI assistant for helping debug a Linux kernel issue that required 24 patches and 18 kernel boots to resolve, with the final fix being a one-line change from round_up() to ro…
Agentic UI, a new interface paradigm for AI agents, must show progress, tool use, direction, and control, according to a developer's guide. The guide lists seven UI styles, including chat, copilot, ca…
A new explainer describes neural networks as machines that learn by adjusting millions of dials through repeated forward and backward passes, starting with random guesses and refining until they work.…