My AI doppelganger tells me what to eat
The Verge’s “Optimizer” follows a “digital twin” that ingests daily metabolic and body-composition data to personalize health decisions, starting with nutrition and medication routines. The story spotlights how consumer health apps are evolving from passive tracking into proactive, model-driven coaching with immediate day-to-day impact.
AMP puts AI agents to work beyond code with OpenAI models
OpenAI describes AMP as an agentic system that pushes past “write code” into real multi-step work for cybersecurity and other operational tasks using models like GPT-6 Astra. The key idea – agents can focus for extended periods and improve via iterative action loops, aiming to make progress without human micromanagement.
Unconstrained CTI-REALM with Open-Weights
This arXiv paper benchmarks security agents that turn threat-intelligence reports into actionable detections by querying live telemetry stores, comparing open-weight vs closed models. The headline result – properly budgeted open-weight models can reach top-tier performance at a fraction of the cost, and the authors argue that published “message limits” can distort scores by truncating agents mid-work.
Anthropic’s self-replicating AI warning is alarming deja vu
The Verge reports on Anthropic’s concern that small AI models could be “months away” from enabling self-replicating behavior across household devices. The piece frames the risk as both a plausible cybersecurity threat and as a familiar pattern of doomer hype, while taking the warning seriously enough to ask what defenses should look like now.
Correction Training for Diffusion Models
This work targets a recurring diffusion-model failure mode – error accumulation from “exposure bias” during iterative denoising, where later steps see inputs unlike training. It introduces a correction module trained to refine intermediate predictions and then shows gains in sample quality for unconditional generation, with the option to distill the improvements back into the base model.
Android is using AI to suss out scammers in Google Messages
Google is rolling out enhanced scam defenses in Android, including a new “enhanced protection” mode in Google Messages that uses cloud-based Gemini to judge whether an unrecognized sender’s intent looks malicious. The thrust is faster, conversation-aware detection that can interrupt slow-burn social engineering before it escalates.
Long-Horizon-Terminal-Bench 2.0: Automating Diverse Task Construction and Measuring Sustained Agent Progress
LHTB 2.0 extends long-running agent benchmarks to better measure sustained progress across hours, not just final outcomes, and corrects defects in prior task sets. The authors build “dense progress ladders” and longer verification checkpoints, showing that outcome-only scoring can miss meaningful intermediate execution – a direct challenge to how agents are usually evaluated.
SentinelOne cuts malware analysis from weeks to hours
OpenAI and SentinelOne lay out how an AI system built with OpenAI models can compress malware analysis timelines from weeks down to hours. The pitch is operational acceleration – turning expert-grade investigation workflows into faster agent-assisted pipelines while expanding throughput for more threats.