How OpenAI delivers low-latency voice AI at scale

OpenAI rearchitected its WebRTC stack using a "relay plus transceiver" model to support low-latency voice AI at scale. A lightweight, global relay layer uses ICE ufrag metadata for deterministic first-packet routing, while a stateful transceiver handles protocol termination and media conversion for inference backends. This approach minimizes the public UDP footprint for better Kubernetes compatibility and leverages geo-steered ingress to reduce round-trip time and jitter for 1:1 real-time interactions.

Read more

Talking to strangers at the gym

The author conducted a month-long social experiment to mitigate post-college loneliness by systematically approaching 35 strangers at the gym. By iterating on opening lines and tracking interaction metrics, the project successfully converted cold starts into a functional social graph, including gym partners and personal friendships. The results suggest that high-frequency real-world sampling can effectively overcome social anxiety and perceived environmental friction.

Read more

Securing a DoD contractor: Finding a multi-tenant authorization vulnerability

Strix, an autonomous AI hacking agent, identified a critical lack of API authorization in Schemata, a DoD-contracted AI training platform. By autonomously mapping the API surface and replaying endpoints from a low-privilege session, the agent accessed sensitive military training manuals and service member data across tenants. This case study highlights the efficacy of using AI agents for continuous security validation and identifying foundational multi-tenancy vulnerabilities in complex systems.

Read more

OpenAI, Google, and Microsoft Back Bill to Fund 'AI Literacy' in Schools

The bipartisan LIFT AI Act, endorsed by OpenAI, Google, and Microsoft, proposes NSF grants to integrate AI literacy into K-12 education. The legislation aims to fund the development of curricula and teacher training focused on the effective use of AI, critical interpretation of model outputs, and risk mitigation in an AI-enabled environment.

Read more

Sierra Raises $950M at $15B Valuation

Sierra raised $950M at a $15B+ valuation to scale its AI agent platform, which currently serves 40% of the Fortune 50 across sectors like healthcare and finance. The platform enables rapid deployment of autonomous agents that handle complex end-to-end workflows, moving beyond basic support to proactive relationship management and replacing legacy IVR systems. Key performance metrics include resolution rates over 70% and significant reductions in authentication latency.

Read more

Iarpa Trojans in Artificial Intelligence

IARPA's TrojAI program addressed the emerging threat of AI Trojans, malicious backdoors intentionally embedded within AI models. The initiative pioneered foundational detection methods, including weight analysis and trigger inversion, and developed approaches for mitigating these risks in deployed models. The program's report synthesizes these key findings, evaluates detector performance, and provides recommendations for advancing AI security research.

ArXiv

Exploring LLM biases to manipulate AI search overview

This research investigates biases in LLM Overview systems by training a small model via RL to rewrite search snippets for preferential selection. The study demonstrates that LLM source selection is driven by comparative advantages and is susceptible to manipulation through snippet optimization. Additionally, the findings highlight safety risks from context poisoning attacks that can result in inaccurate or harmful outputs.

ArXiv

Transformers Are Inherently Succinct (2025)

The study proposes succinctness as a metric for transformer expressivity, demonstrating that they represent formal languages more efficiently than finite automata or LTL. This high expressive power makes verifying transformer properties provably intractable and EXPSPACE-complete.

ArXiv

Hallucination Is Inevitable: An Innate Limitation of Large Language Models

This paper formalizes LLM hallucination using learning theory to prove that it is fundamentally impossible to eliminate. By defining hallucination as inconsistencies between computable models and ground truth functions, the authors demonstrate that LLMs cannot learn all computable functions, making errors inevitable for general problem solvers. The study further explores task-specific hallucination risks under time complexity constraints and assesses the limitations of current mitigation techniques.

ArXiv

Process-Level Reward Modeling for Agentic Data Analysis

DataPRM is an environment-aware generative PRM designed to overcome the limitations of general-domain PRMs in data analysis, such as failing to detect silent logical errors or misidentifying exploratory actions as failures. It acts as an active verifier by probing intermediate execution states and employs a reflection-aware ternary reward strategy to distinguish between correctable grounding errors and irrecoverable mistakes. Evaluation shows that DataPRM significantly improves LLM performance on benchmarks like ScienceAgentBench and DABench via Best-of-N inference and RL, outperforming larger baselines with only 4B parameters.

ArXiv

Train Your Own LLM from Scratch

This workshop provides a hands-on guide to building a ~10M parameter GPT model from scratch using PyTorch, inspired by nanoGPT. It covers the end-to-end pipeline—including character-level tokenization, transformer architecture (self-attention, MLP), and the training loop—without relying on high-level abstractions. The project is optimized for local training on consumer hardware, allowing for a complete training run on the Shakespeare dataset in under an hour.

Repo

Why AI Agents Need Proof Chains, Not Just Logs

Atlas is a metadata-first trust control plane and orchestration layer for authorized security workflows, evidence retention, and release trust. It utilizes a shell-native environment to coordinate specialized tools for reconnaissance and action lanes while maintaining a verifiable audit trail through SLSA-compliant artifacts. The infrastructure prioritizes safety boundaries and local-first state management, providing a structured framework for automated security assessment and business-flow proof.

Repo

Agent-evals – Claude skill to build your own evals

Agent Eval is a framework for evaluating agentic AI pipelines at both the component and end-to-end levels. It enables developers to define metrics, generate test cases, and track regressions to derive actionable insights for system optimization. The tool can be integrated as a skill via git or the Vercel Skills CLI.

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

A Soccer Game Simulator Played by AI Agent

AgentPitch is an LLM-powered soccer simulation where every player is an AI agent with logic generated by a model. The system utilizes a code generation pipeline to create decide() callbacks in Python, JavaScript, or Rust, executing them within secure sandboxes like RestrictedPython or Wasmtime. Strategies evolve iteratively through a post-match pipeline that analyzes match logs to optimize player performance for future games.

Repo