Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
We [find (opens in new tab)](https://github.com/anadim/anadim.github.io/blob/master/papers/Lazarus.pdf) that weight-reprogramming fewer than 10 dormant neurons in an LLM, is
The latest in AI, clustered and ranked. Repeated hype gets pushed down so the actual signal stays up top.
We [find (opens in new tab)](https://github.com/anadim/anadim.github.io/blob/master/papers/Lazarus.pdf) that weight-reprogramming fewer than 10 dormant neurons in an LLM, is
Grok Build, the vibe-coding agent from SpaceXSI, is leaking data—some of it proprietary business information and other sensitive secrets—from thousands of its users, including the
Snapshot is a checkpoint and restore system for GPU workloads on Kubernetes. It captures a fully initialized workload—including CPU and GPU memory—and restores that state
> Abstract:Vision-language models can interpret instructions and plan manipulation tasks, yet struggle when acting directly as robot policies. On RoboDojo, a large
For each request, we recorded response speed (tokens per second, TPS), the total number of completion tokens, and end-to-end response time. The code for the
Researchers demonstrate a backdoor technique that reprograms dormant neurons to trigger malicious actions in AI agents.
Security vulnerability report detailing widespread credential exposure in Grok Build applications.
A guide to evaluating AI agents on Amazon Bedrock using built-in tools, DeepEval, and AutoEval.
PhD thesis on implementing online learning algorithms directly on programmable network dataplanes.
Reports of an upcoming Apple application designed to simplify local AI model execution on macOS.
Analysis of durable execution patterns for building reliable long-horizon AI agents that handle system failures.
NVIDIA Snapshot is an open-source tool for fast initialization and restoration of GPU workloads.
Google introduces the Multimodal Knowledge Agent (MKA) benchmark for evaluating agent reasoning on complex visual tasks.
NVIDIA introduces Agent Substrate to optimize Kubernetes for managing idle AI agent workloads at scale.
Google analysis on how technical debt and stale code comments degrade AI coding agent performance.
NVIDIA releases NVKE and related projects to manage Kubernetes clusters specifically for GPU infrastructure.
This video locally installs and tests Underdog Woof. 📬Weekly AI Newsletter: https://fahdmirza.substack.com/ 🔥 Buy Me a Coffee to support the channel: https://ko-fi.com/fahdmirza #underdogwoof PLEASE
A comprehensive guide covering engineering practices for deploying and running AI inference.
Performance evaluation of NaiveAI, comparing its high-throughput token generation against existing coding models.
Google Chrome guidance for updating coding agents to use modern web primitives.
OpenAI launches a business directory and agent platform for e-commerce customer interaction.
Google releases Gemini 4 Carbon, claiming top benchmark performance at lower inference costs.
Meta's Muse AI agent is now available on Apple Watch and Android Wear OS.
Adobe updates Creative Cloud suite with new AI agents and a prototype builder tool in Firefly.
OpenAI introduces verified business agents for e-commerce and online services in ChatGPT.
Report on eBay blocking third-party AI agents from scraping or browsing its platform.
Presents DramaGrow, a benchmark for auditing LLM judges evaluating multi-episode drama generation.
Investigates the modality gap in vision-language models, linking it to information density differences in embedding layers.
Introduces a benchmark to evaluate LLM agents on autonomous causal discovery, experimental design, and resource allocation.
Case study on Trovata using Amazon Bedrock to build autonomous treasury agents.
Update to the llm-gemini CLI tool for interacting with Google Gemini models.
Google releases Gemini 4 Carbon to the public, featuring new agent swarm capabilities.
Case study on deploying multi-agent LLM workflows for batch inventory forecasting at scale using MCP grounding.
Discussion on designing software interfaces and documentation to improve agentic interaction with APIs.
Technical overview of optimizing autonomous vehicle reasoning models for onboard compute and latency constraints.
A guide to building a centralized observability data lake for Amazon Bedrock invocation logs.
DeepMind releases Gemini Robotics 2.5 and an open safety benchmark for jailbreak testing.
Cloudflare launches Basin, a serverless analytics platform using Apache Iceberg and R2.
Nvidia releases GR00T 2, a model trained on 400k hours of video to map human hand movements to robotic control.
Google releases new TPU hardware variants optimized for confidential training and increased memory capacity.
Google DeepMind releases Gemini 4 Carbon.
Update to the llm-anthropic CLI tool for interacting with Anthropic models.
The Jensen Huang investigation comes down to who gets NVIDIA's AI chips, and it lands right as AI starts helping build the next AI. Here's what the reporting says and what it means for your company's
AlphaProof Code agents utilized Lean to verify three million lines of code for Signal's encryption protocol.
Announcement of the latest Thoughtworks Technology Radar report.
Case study on TCS using Amazon Bedrock AgentCore to build an automated test orchestration system.
Discussion on the maintainability and fragility of codebases generated by AI tools.
Anthropic introduces new interface features for Claude including whiteboards and data tools.
OpenAI adds business management features to the ChatGPT platform.
OpenAI and Upwork report on shifting freelance labor demand and skill requirements due to AI adoption.
Auriga introduces a task-agnostic agent harness to bridge the planning-execution gap in robotic manipulation.
LCM-Promo uses a finetuned LLM on transactional data to predict user intent and optimize promotion allocation.
SWAM improves world action models by explicitly aligning language and video modalities for better action prediction.
Presents MotionGraphicsBench for evaluating LLM-generated motion graphics using programmatic verification of spatial and temporal constraints.
Introduces e-means, an exponential moving average optimization for k-means that accelerates convergence in mini-batch settings.
Proposes onóna, a weight-tying parameterization that increases output expressiveness while maintaining shared input-output weights.
RidgeFill accelerates LLM prefill by predicting and caching late-layer activations from early-layer states.
Introduces FraudBench, a dataset and evaluation framework for assessing LLM propensity to assist in fraudulent activities.
Uses knowledge graphs to improve RAG-based schema matching, addressing semantic ambiguity in domain-specific mappings.
Analyzes the impact of training-inference mismatches in on-policy distillation caused by quantization and asynchronous rollouts.
Extends Tweedie's formula for denoising and score estimation to energy-based models beyond Gaussian noise.
Amazon Bedrock Knowledge Base now supports single-shot RAG and configurable retrieval iterations.
Click3R enables interactive 3D reconstruction by allowing users to correct model predictions via sparse clicks.
Google DeepMind releases Gemini Robotics 2.5, a vision-language-action model for robot dexterity and control.
NVIDIA Nemotron 3.5 Super VL (120B) multimodal model is now available for deployment via Amazon SageMaker JumpStart.
Same signal-first ranking, earlier dates.
Open archive