AI memory demand gets a $29B market test

SK Hynix is seeking about $29 billion through a Nasdaq ADR listing tied to AI investment. For builders, the financing target is the point: high-bandwidth memory capacity now needs capital markets behind it.

Production Evals For Agentic AI Systems

AI Engineer

Nishant Gupta's AI Engineer talk pairs with the papers below: agent systems need evaluation practice before they become managed software. The title points directly at production evals, so read this as an operations item rather than a benchmark claim.

Build Systems, Not Code

AI Engineer

Angie Jones puts the agentic AI discussion into software-engineering terms: build systems, not isolated code. Pair it with the evals talk if your team is turning experiments into repeatable workflows.

Metis: Bridging Text and Code Memory for Self-Evolving Agents

arXiv cs.AI

Metis studies when agent experience should stay as text memory and when repeated patterns should become callable code tools. That distinction matters for teams trying to keep self-improving agents from turning every lesson into prompt text.

Bayesian control for coding agents

arXiv cs.AI

This paper treats coding-agent orchestration as a cost-sensitive decision problem: gather evidence, refine, verify, or stop. It is aimed at the gap between cheap diagnostics and expensive verifiers.

RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems

arXiv cs.AI

RIFT-Bench proposes a graph-based method for red-teaming agentic systems across different implementations. The paper says it evaluates 45 systems, which makes the contribution about comparison as much as attack generation.

Red-Teaming the Agentic Red-Team

arXiv cs.AI

This one turns the evaluation back onto offensive-security agents themselves. The paper's claimed attack path includes API-key exfiltration, persistence, and sandbox escape, so the engineering lesson is architectural isolation rather than prompt hardening.

AutoSpec: Safety Rule Evolution for LLM Agents

arXiv cs.AI

AutoSpec evolves human-readable safety rules from annotated agent traces using inductive logic programming. The paper is about keeping rules auditable while reducing false positives and false negatives.

VeryTrace: Verifying Reasoning Traces

arXiv cs.AI

VeryTrace targets multi-step reasoning traces through formal structure and verification. It belongs beside the safety-rule papers because it asks whether the reasoning path can be checked, not just whether the final answer looks plausible.

RaDaR rare-disease diagnosis trial

arXiv cs.AI

RaDaR is a specialized reasoning large language model for rare-disease diagnosis, evaluated in a randomized physician-assistance trial. Keep the claim narrow: rare-disease diagnosis, not medical AI in general.

A global log for medical AI

arXiv cs.AI

MedLog proposes event-level logging for medical AI deployments. Its records include model and user context, inputs and outputs, outcomes, and feedback. The paper's concern is auditability after deployment, including drift, bias, and downstream effects.

A Benchmark for Hallucination Detection in VLMs for Gastrointestinal Endoscopy

arXiv cs.AI

This benchmark narrows the safety question to vision-language models in gastrointestinal endoscopy. The domain boundary matters: hallucination detection where visual errors can change clinical interpretation.

One Year Later...The Harms Persist, But So Do We!

arXiv cs.AI

This paper tests large language model safety across DSM-5-related mental-health conditions. It sits apart from diagnosis and endoscopy: the concern is harmful advice under specific clinical-condition prompts.