Jun 29, 2026

Daily

Weekly

• GLM 5.2 beats Claude in our benchmarks

  • GLM 5.2, an open-weight model from Zhipu AI, achieved a 39% F1 score in IDOR detection, outperforming Claude Code (32%) at a cost of $0.17 per vulnerability found, showcasing its competitive edge in security tasks.

• A way to exclude sensitive files issue still open for OpenAI Codex

  • A proposed feature seeks to implement a mechanism for marking files and paths that should not be accessed or sent to the model, enhancing security and usability across repositories with a deterministic configuration.

• Model Training as Code

  • Savanna transforms model training into a collaborative software project by implementing the entire training pipeline in code, enabling one-click, hermetic training runs that enhance efficiency and reduce human error.

• Ornith-1.0: Self-scaffolding LLMs for agentic coding

  • Ornith-1.0 introduces a self-improving training framework that enables models to generate their own scaffolds for coding tasks, enhancing solution quality through joint optimization.

• TOP500 at ISC’26: We have a New Number 1 Supercomputer

  • LineShine Supercomputer in Shenzhen, China, has claimed the top spot on the TOP500 list, marking the first Chinese entry in nine years with a CPU-only system that boasts 2.198 Exaflops of sustained FP64 performance.

• Knowledge Distillation of Black-Box Large Language Models

  • Proxy-KD is a novel method that enables efficient knowledge transfer from black-box large language models (LLMs) to smaller models, enhancing their performance significantly.

• Ornith-1.0: self-improving open-source models for agentic coding

  • Ornith-1.0 is a self-improving open-source model for agentic coding, featuring architectures like 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE, which excel in coding benchmarks such as Terminal-Bench 2.1 and SWE-Bench.

• Cerebras OpenAI deal capacity has effectively killed the waitlist for everyone else

  • Cerebras' deal with OpenAI has effectively monopolized its inference capacity, allocating $20 billion worth of chips to a single customer, leaving smaller startups in the lurch.

• Apple Neural Engine: Architecture, Programming, and Performance

  • The Apple Neural Engine (ANE) serves as a fixed-function matrix accelerator in Apple devices, with insights derived from reverse-engineering revealing its architecture and performance metrics across the A11 to A18 and M1 to M5 chip families, as detailed in the arXiv article.

• Micro-Agent: Beat Frontier Models with Collaboration Inside Model API

  • Micro-Agent technology enhances AI model performance by enabling collaboration within the serving layer, allowing multiple models to work together without altering their individual identities.

• US Grid Constraints: Towards 40GW+ of Behind-the-Meter Datacenter by 2028?

  • US datacenter demand is projected to surge from +21GW in 2026 to +84GW by 2030, with Behind-The-Meter (BTM) solutions expected to power over 50% of new datacenters by 2028, driven by the inability of the grid to keep pace with power needs.

• Google's Agentic Peer-Reviewer Handled ~10K Papers at ICML/STOC — Formal Research Paper Now Out

  • Google's agentic AI peer-reviewer efficiently reviewed ~10,000 papers at ICML/STOC, demonstrating a 34% increase in catching mathematical errors compared to traditional methods.

• Claude Meets Blackwell Ultra: Anthropic’s Models Now Run on NVIDIA GB300 in Azure

  • Claude models are now available on NVIDIA GB300 Blackwell Ultra in Microsoft Azure, enabling enterprises to create autonomous AI agents tailored to specific domains.

• EML Trees are Universal Approximators

  • EML trees serve as universal approximators, enabling the representation of all elementary functions through composition, thus allowing for the approximation of any function in a general space with controlled size and depth.

• MultiHashFormer: Hash-based Generative Language Models

  • MultiHashFormer introduces a novel framework for hash-based autoregression in language models, enabling efficient token representation through unique hash signatures generated by multiple independent hash functions.