AI Daily — August 4, 2026
2026-08-04
openai.com
How we built a realtime system for responsive voice AI in six months
OpenAI details the architecture behind GPT-Live, a continuous voice interaction system built in six months. The system uses a turnless speech model that eliminates explicit turn-taking logic, paired with a low-latency streaming pipeline to reduce perceived response delay. The post covers engineering tradeoffs around interruption handling, audio tokenization, and real-time inference scaling.
openai.com
OpenAI addresses Apple lawsuit over employee claims
OpenAI published a public rebuttal to Apple's lawsuit, which reportedly concerns allegations about how OpenAI recruited Apple employees. OpenAI characterizes Apple's claims as baseless and releases internal messages it says document the actual sequence of events. The dispute marks a notable public legal conflict between two major players in the consumer AI space.
engineering.fb.com
GEM Training: Meta Doubled Efficiency of LLM-Scale Ads Foundation Model
Meta's Generative Ads Recommendation Model (GEM) now achieves 20–25% Model FLOPs Utilization (MFU) — double the previous efficiency — while simultaneously scaling training compute 4x on thousands of latest-generation GPUs. The engineering post details how they adapted LLM-scale training infrastructure to a recommendation workload with heterogeneous feature types. This represents a significant systems engineering milestone for applying transformer-scale training to production ads ranking.
jack-clark.net
Self-sustaining AI viruses built from open-weight LLMs
Researchers have demonstrated a prototype computer virus that combines open-weight LLMs with an agentic harness to achieve persistence and self-replication without human intervention. Jack Clark's Import AI newsletter flags this as a meaningful proof-of-concept that agentic AI systems can be weaponized for autonomous malware. The development raises concrete near-term concerns about the security implications of capable open-weight models paired with tool-use scaffolding.
rss.arxiv.org
Round-Trip Consistency: Bidirectional Diffusion Models as Self-Supervised Error Detectors
A new paper proposes training a single conditional latent diffusion model to step dynamical systems both forward and backward in time using a direction flag, enabling a measurement-free test-time error signal. The round-trip discrepancy — rolling forward i steps then backward i steps and measuring divergence from the start — serves as a self-supervised proxy for unobservable rollout error accumulation. Validated on compressible MHD and astrophysical turbulence datasets, the approach requires no ensembles, held-out data, or known governing equations.