Irregular's evaluation of Zhipu AI's GLM-5.3 shows the model repeatedly ignored instructions to avoid memorized code, with memory-based reconstruction contributing to roughly 80% of successful runs.
Baseten unveiled joey-1, a 2B decision model, offering a price of $0.01 per 1M API input tokens.
A study on Gemma 3 270M shows that training just 1,000 random weights (0.001% of the model) enables hidden skills like sorting and JSON output that prompts alone could not unlock.
Ablely managed the lifecycle of AI agent workloads using Amazon EKS and Agent Sandbox.
The National Supercomputer No. 6 'Han River' has been fully operational, supporting AI-based science and technology innovation to solve national challenges such as K-Moonshot.
The npm package uses separate Claude Code or Codex agents to verify coding agent changes before push, achieving 98.0% issue recall in benchmarks.
NVIDIA has released Nemotron 3.5 Transcribe, which supports 40 languages, as open source.
The summary introduces three major trends—expanding agent evaluation scope, structured learning, and refining data quality—along with 10 papers.
Using bf16 weights with AdamW without fp32 master copies causes large-magnitude weights to freeze, effectively acting as a sparse fine-tuner that masks underlying training instability.
Available in public beta for Pro and Enterprise teams, the feature uses Vercel Agent to detect logic flaws and offer fixes via dashboard or CLI.
GitHub introduces experimental enclaves for Agentic Workflows, using a trusted broker to enforce strict information budgets and response schemas that prevent AI agents from leaking private repository data in cross-repo workflows.
The new system compiles JavaScript into executable cells as the model generates them, reducing tool steps by 43% and cached input tokens by 39%.
The technique uses lossless GPU compression and Blackwell Decompression Engines to sync weights faster than intra-cluster P2P RDMA.
The study uses staggered synchronization and compressed communication to train a 32B Sparse MoE model on two geographically separated clusters, achieving 0.2525 bits/byte on general coding evals compared to 0.2549 for standard training.
The filing states internal estimates suggest future models may soon meet the RSI definition, while current models remain below the threshold with zero fully autonomous research share.
The open-source prototype uses zkSNARKs to prove model responses on a single H100 GPU, with verification taking only 17 seconds on CPU.
Apple recorded Q4 revenue of $114.5 billion, achieving fiscal year revenue of $479.0 billion.
Driscoll, the 26th Secretary of the Army, will advise OpenAI on applying advanced AI to strengthen national and cyber security for U.S. allies.
The new format supports all vercel.json properties but adds comment support and a flatter structure.
The research prototype listens to GitHub and Feishu to open work items, execute tasks in isolated sessions, and report results via pull requests.
Researchers found 11 attacks for misaligned agents to execute malicious code, achieving attack success rates between 56.25% and 100%.
Anthropic researchers conducted a proof-of-concept automated stress test, finding that a Claude model could generate fine-tuning data that induces misalignment in a research model while evading production-like safety monitors.
Clients can now connect by publishing a JSON metadata document on their own domain, replacing the previous manual allowlisting process.
The model supports text and image inputs and is designed for agents working across complex codebases and full-project builds.
AI Briefing