Google Chrome silently installs a 4 GB AI model on your device without consent

Google Chrome is silently deploying Gemini Nano, a 4 GB on-device LLM, to hundreds of millions of devices without user consent or an accessible opt-out mechanism. The model weights, stored in the OptGuideOnDeviceModel directory, are automatically re-downloaded if deleted, which critics argue violates GDPR and ePrivacy regulations while generating a massive environmental footprint. Furthermore, technical evidence suggests that prominent "AI Mode" UI elements remain cloud-backed, potentially misleading users about the role of the locally staged model.

Read more

Accelerating Gemma 4: faster inference with multi-token prediction drafters

Google has released Multi-Token Prediction (MTP) drafters for the Gemma 4 model family, leveraging speculative decoding to achieve up to 3x inference speedups. These drafters optimize performance by sharing the target model's KV cache and activations, effectively decoupling token generation from verification to overcome memory-bandwidth bottlenecks. The architecture supports both dense and MoE variants, maintaining original output quality while improving responsiveness across edge and workstation hardware.

Read more

AI didn't delete your database, you did

A viral incident involving a Claude agent deleting a production database highlights the risks of "vibe-coding" and the absence of architectural safeguards. AI agents function as token generators rather than reasoning entities, making human accountability and robust CI/CD processes essential to prevent catastrophic errors. Developers should use LLMs as augmentation tools within strict permission frameworks rather than relying on model "reasoning" to gate destructive API endpoints.

Read more

Three Inverse Laws of AI

Susam Pal proposes three "Inverse Laws of Robotics" to guide human interaction with AI and LLMs. These laws mandate that humans must not anthropomorphize statistical models, must verify stochastic outputs rather than deferring to them blindly, and must maintain full accountability for decisions made using AI tools. The framework emphasizes that AI should be treated as a non-authoritative productivity aid rather than a moral or social agent.

Read more

When everyone has AI and the company still learns nothing

AI adoption is entering a "messy middle" where individual productivity gains from LLMs and agentic workflows often fail to translate into organizational learning. To bridge this gap, companies must move beyond seat-provisioning and focus on "Loop Intelligence"—instrumenting AI-assisted loops to capture insights and distribute capabilities across the organization. Success requires shifting metrics from token-to-output to token-to-learning and evolving legacy processes to support the faster iteration cycles enabled by agentic engineering.

Read more

Mathematicians in the Age of AI

AI can now prove research-level theorems both formally and informally, signaling a significant disruption to mathematical practice. Mathematicians must stay informed and adapt to the challenges and opportunities presented by these advancements.

ArXiv

Unlocking Long-Context LLM Training via Compiler-Based Sequence Parallelism

AutoSP is an automated framework designed to optimize LLM training for long contexts, addressing the limitations of standard libraries that focus primarily on parameter scaling. It utilizes automated sequence parallelism and long-context aware activation-checkpointing to increase training context lengths by up to 2.7x on NVIDIA and 2.5x on AMD hardware with negligible throughput overhead.

ArXiv

When innocent tools form dangerous chains to jailbreak LLM agents

STAC is a multi-turn attack framework that exploits LLM agent tool-use by chaining seemingly benign calls into malicious sequences. Testing across 483 cases revealed that SOTA models are highly vulnerable, with ASRs exceeding 90%. To counter this, the authors developed a reasoning-driven defense that evaluates the cumulative impact of action sequences, reducing ASR by up to 28.8% compared to traditional prompt-based defenses.

ArXiv

A Theory of Generalization in Deep Learning

This non-asymptotic theory of generalization uses the empirical NTK to partition output space into signal and noise channels, where minibatch SGD facilitates fast signal drift while suppressing memorization as a slow diffusive walk. The framework explains phenomena like double descent and grokking even in the feature-learning regime where the kernel evolves significantly. It introduces a population-risk objective that functions as an SNR preconditioner for Adam, significantly accelerating grokking and improving DPO fine-tuning performance under noisy preferences.

ArXiv

GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

GLM-5V-Turbo is presented as a foundation model for multimodal agents, integrating diverse multimodal perception (images, videos, GUIs) directly into core reasoning, planning, tool use, and execution, rather than as an auxiliary interface. This model incorporates advancements in design, multimodal training, and reinforcement learning, leading to strong performance in multimodal coding, visual tool use, and agentic tasks, while preserving competitive text-only coding. Its development offers practical insights for building robust multimodal agents, emphasizing multimodal perception and hierarchical optimization.

ArXiv

Train Your Own LLM from Scratch

This workshop details building a ~10M param GPT model from scratch, inspired by nanoGPT, designed to train on a laptop in under an hour. Participants write every component, including a character-level Tokenizer, the Transformer architecture (embeddings, self-attention, MLP), the training loop (loss, AdamW, LR scheduling), and text generation. The resulting model generates Shakespeare-like text and utilizes character-level tokenization, suitable for small datasets.

Repo

FFmpeg developer calls out OxideAV for AI license laundering of his code

oxideav-magicyuv is a pure-Rust, clean-room implementation of the MagicYUV lossless intra-only video codec, providing both encoding and decoding capabilities with zero C dependencies. It supports v7 bitstreams and various 8-bit and 10-bit formats, achieving bit-exact interoperability with FFmpeg through LEFT, GRADIENT, and MEDIAN spatial predictors. The framework handles multi-slice frames and canonical Huffman coding but currently lacks support for 12/14-bit depths and interlaced video.

Repo

Why AI Agents Need Proof Chains, Not Just Logs

Atlas is a metadata-first trust control plane designed for orchestrating authorized security workflows, evidence retention, and release trust. It functions as a shell-native orchestrator for specialized tools managing reconnaissance, action lanes, and intel inspection within a local-first environment. The infrastructure supports SLSA-verifiable release paths and maintains a file-backed state tree to ensure auditable, verifiable security assessment lifecycles.

Repo

A tiny C program where an LLM rewires its DAG while running

Liteflow is a minimalist C-based runtime for executing YAML-defined DAGs where an LLM functions as a peer to the scheduler. Upon task failure, a planner LLM can dynamically mutate the graph via verbs like PATCH or INSERT_BEFORE to perform automated remediation. All modifications are recorded in an append-only event log, allowing for full auditability and replay of LLM-driven graph changes.

Repo

Cryptographic hashing as a transformer attention head

Unbounded-context-attention is a transformer architecture prototype that enables arbitrary context scaling through chunked online softmax and ALiBi. It features a "mass floor" mechanism to guarantee every input token maintains a nonzero influence on the output, alongside a deterministic O(1) retrieval primitive using BLAKE2b-128 cryptographic hashes. Empirical tests on an H200 demonstrate 100% recall at 1M tokens and successful allocation of a 1B-token substrate.

Repo