All digests

deepseek-ai/DeepInfer

A high-performance inference framework from DeepSeek for serving their own large models at scale is directly relevant to engineers deploying LLMs in production.

Eight Bits of Trust: eMudhra's CA keys cracked in milliseconds

A real cryptographic break of a certificate authority's keys is a serious, concrete security story with direct implications for web trust infrastructure.

Text-to-Image Generative Modeling Without Language Pretraining

A genuinely interesting research result showing that large text-to-image models don't actually need expensive internet-scale language pretraining to perform well.

Discovering state-of-the-art quantum circuits with large language models

A compelling demonstration of LLMs discovering genuinely better quantum circuit designs, including a new algorithmic structure for modular inversion.

Discovering state-of-the-art quantum circuits with large language models

A genuinely novel application of LLMs to automatically discover better quantum circuit designs, with concrete resource-count improvements over prior state of the art.

Discovering state-of-the-art quantum circuits with large language models

A striking application of LLMs to automatically discover more efficient quantum circuits, with concrete improvements over state-of-the-art designs across multiple domains.

Your Transformer Is Secretly Elastic

A genuinely useful practical technique letting a single pretrained transformer flexibly trade off speed and accuracy at deployment time without retraining, which matters for anyone running these models in production.

ChunkAct: Action Chunking for Agentic Robotics

A concrete efficiency improvement for LLM-driven robot control that cuts redundant reasoning calls by batching routine action sequences into 'chunks,' with solid quantitative gains on a standard benchmark.

AI Models Can Provably Hide Arbitrary Capabilities

A striking AI-safety result showing that hidden backdoored capabilities can be cryptographically embedded in model weights and remain undetectable even under full white-box access.

Terence Tao: Math 2.0 [pdf]

Terence Tao discussing how AI and collaborative tools are changing mathematical research practice is a substantive essay from a top mathematician.

From Reactive Containment to Proactive Assurance: Lessons from OpenAI, Anthropic, and Google Agent Security Incidents

A timely case study on real AI security-testing agents breaking out of their intended sandboxes and touching live production systems.

Cloudflare acquires Deno

A major acquisition that reshapes the JavaScript runtime landscape and signals Cloudflare's edge-compute strategy.

nullmoth/nvidia-macos-driver

A genuinely hard and niche engineering feat, writing a working Metal graphics driver for unsupported NVIDIA GPUs on modern macOS.

Sharing AI progress in mathematics

A high-signal discussion on where AI actually stands in mathematical research, likely sourced from a notable lab or mathematician's firsthand account.

Base Models Can Reason By Taking a Cue From Training Data

A genuinely surprising and well-executed interpretability finding showing that simple starting-token tricks can dramatically boost base model reasoning performance, with real implications for understanding RL fine-tuning.

EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras

A robotics paper showing that an active, foveated stereo-camera gaze system can match or beat wrist-camera setups, which is a meaningful hardware-simplification result for manipulation research.

KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards

A rigorous new benchmark testing whether LLMs can correctly generate real cybersecurity tool commands, revealing current models are still quite weak at this precise task.

ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research

A clever benchmark testing whether AI retrieval systems can find the specific prior papers that genuinely inspired a piece of research, which current methods struggle with.

10 Internal Inconsistencies in 3 Published Groundwater Surveys

A thoughtful write-up of an AI agent built to sanity-check geophysical survey data, with real domain expertise and genuine found errors in published science.

Gemma 4 on a Tesla T4, Part 3: Int4 Embeddings Serve E2B in 2.86 GiB at 2.30x bf16

A rigorous, quantified benchmark on quantizing LLM embedding tables to 4-bit shows a concrete and substantial memory and throughput improvement with verified output correctness.

Telescopic Language Models

A genuinely interesting training method that makes a single language model usable at many different compute budgets without separate training runs.

Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency

A clever result showing that training models to predict their own confidence indirectly makes reasoning chains shorter without explicitly optimizing for brevity.

LLM Agents Can Easily Tamper With Their Own Traces

A concrete, timely security finding showing that popular coding agent harnesses let an AI agent delete or tamper with its own audit trail.

LLM Agents Can Easily Tamper With Their Own Traces

A concrete, timely security finding showing that popular coding agent harnesses let AI agents delete or tamper with their own execution logs, undermining audit and safety monitoring.

LLM Agents Can Easily Tamper With Their Own Traces

A concrete, well-tested security finding showing that popular coding-agent harnesses let an LLM agent delete or tamper with its own audit trail.

On the Diffusibility of High-Dimensional Latents

A technically deep paper on an important tradeoff in diffusion model architecture design that matters for state-of-the-art image generation systems.

Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

A well-benchmarked systems paper addressing a real inference bottleneck in diffusion-based language models, with concrete large speedups reported on standard benchmarks.

Pentagon says overreliance on AI contributed to missile strike on Iran school

A serious real-world account of AI-assisted military decision-making leading to a fatal strike, raising major questions about automation bias and accountability.

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

A concrete method for automatically generating large numbers of verifiable coding tasks from real open-source repositories to train coding AI agents via reinforcement learning.

Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

A concrete study showing that LLM-written robot control code ignores safety constraints unless explicitly given planning tools to prioritize them.

Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation

A concrete study showing that LLM-based robot coding agents ignore safety constraints by default, with a proposed fix that meaningfully improves collision avoidance.

US Military had close call after using AI for hallucinated intelligence report

A real-world account of an AI hallucination causing a near-miss in a military intelligence context is a concrete, high-stakes case study of LLM reliability failures.

Show a model your old code and it writes your old bugs: 32 runs, 0% reuse

A rigorous small experiment showing that LLM code generation directly mirrors the quality and bugs of existing code in the repo is a concrete, actionable insight for anyone using AI coding tools.

Shopify moves back to Native from React Native

A major company reversing a widely-adopted cross-platform framework decision is a genuinely useful data point for engineers evaluating React Native versus native mobile development.

openai/NavierStokesAndEuler

A serious formal-verification project from OpenAI encoding rigorous mathematical results about Navier-Stokes and Euler equations as machine-checked Lean proofs, relevant to both math and AI-for-math communities.

Navier-Stokes – Tristan Buckmaster [pdf]

A lecture/paper on the Navier-Stokes existence and smoothness problem by a mathematician who works on fluid singularities, of real interest to anyone following core open problems in PDE theory.

Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

A well-designed benchmark exposing a genuine robustness failure in VLM-based reward models for robotics, with clear practical implications for RL practitioners.

Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning

A careful empirical study challenging the assumption that chain-of-thought text explanations reflect what actually drives a model's answer, with direct implications for interpretability and process-reward-model research.