Models

Uniform activation steering is a blunt instrument that wrecks LLM utility on benign inputs. Apple's DSAS fixes that — with complexity strings attached.

Proofcraft completes security proofs for seL4 on AArch64, providing formal mathematical proof of confidentiality, integrity, and functional correctness. This milestone ensures the kernel prevents unauthorized access and data modification, guaranteeing security isolation for applications.

Gigatoken is a new open-source tokenizer that is ~1000x faster than HuggingFace's tokenizers and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines.

Formal verification is a tool to verify correctness of mathematical statements. In this tutorial, we'll write simple statements about cryptography and their proofs in Lean, focusing on the One-Time Pad protocol.

Exploit brokers pay $500k for WordPress RCEs. I found one with GPT5.6 and $25. The vulnerability was discovered using a large language model and can be exploited for remote code execution.

Mozilla's inaugural State of Open Source AI report reveals that open models now power roughly a third of real-world AI usage, but capture only 4% of the revenue. The report highlights the increasing competitiveness of open source AI with proprietary systems, emphasizing its societal and economic benefits.

Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open model with native multimodality and a Mixture-of-Experts architecture. The model achieves approximately 2.5× better scaling efficiency than its predecessor and is available via Kimi.com, Kimi Work, Kimi Code, and the Kimi API.

Bonsai 27B is the first 27B-class model to run on a phone, enabling local execution of AI workloads and reducing dependence on cloud APIs. The model comes in two variants: Ternary Bonsai 27B and 1-bit Bonsai 27B, with the latter being footprint-oriented and suitable for phone-class devices.

Apple's new SpeechAnalyzer API has been benchmarked against Whisper and its predecessor, showing significant improvements in accuracy and speed. The results have major implications for developers and users relying on speech-to-text transcription.

Ploy’s production AI agent moved from Claude Opus to GPT‑5.6 Sol, slashing build time in half and cutting spend by over a quarter—after a disciplined migration that reshaped evals, tool schemas, caching and reasoning.

The OpenWrt One is an open-source router that offers a high degree of customizability and freedom to repair, with a focus on software freedom and community-driven development.

Researchers at Anthropic have discovered a 'global workspace' in language models, which could have implications for AI safety and interpretability. The workspace, dubbed the 'J-space', is a small collection of internal neural patterns that play a special role in the model's processing.

Boko Haram uses frontier AI to aid in combat and day-to-day operations, posing a growing threat to global security. The terrorist group has successfully circumvented some AI safeguards, utilizing tools like ChatGPT and Meta AI to plan attacks and design explosive devices.

GPT-5.6 is a new AI model that achieves state-of-the-art results across various tasks. Learn more about its technical details and potential impact.

SQLite recently patched a rare, 16-year-old bug in its Write-Ahead Log (WAL) checkpointing system that could lead to database corruption. This post from Canonical's dqlite (distributed SQLite) team walks through how they used TLA+ to formally model SQLite's internal behavior, isolate the exact sequence of steps needed to trigger the corruption, and determine whether their own system was vulnerable.

Google’s open‑source Copybara automates bidirectional code flow between confidential and public Git repositories, but it brings integration quirks that enterprises must weigh.

The Godot Foundation is tightening its contribution policy, prohibiting AI-generated code to reduce buggy pull requests and protect maintainers from hidden technical debt.

Mistral’s Leanstral 1.5 delivers a 119 B‑parameter mixture‑of‑experts tuned for Lean 4 theorem proving, with a 256 k token window and zero per‑token cost in Labs.

From vanilla transformers to a tangled web of attention variants, routing schemes, and multi‑GPU pipelines, LLMs have become as complex as today’s recommendation engines – and that shift reshapes cost, risk, and organizational practice.