Morning Edition · Monday, July 27, 2026
Tech
Moonshot AI Releases Kimi K3 Weights, a 2.8-Trillion-Parameter Open Model
The Chinese lab is publishing downloadable weights for a mixture-of-experts model that it says outperforms proprietary US systems on several coding and agent benchmarks. The company reported those benchmark figures itself.
Tech
Anthropic Ships Claude Opus 5, Priced at Half Its Largest Model
Opus 5 lists at $5 per million input tokens and $25 per million output tokens. On Anthropic's own numbers, it roughly doubles its predecessor's score on an agentic-coding benchmark.
Tech
Nvidia Uses Its Own Vera CPU to Speed Up Chip Design by 1.5 Times
The company says pairing Vera with AI agents sped up Cadence Jasper verification and Synopsys VCS simulation, two of the most compute-intensive steps in early chip design.
Tech
OpenAI Puts ChatGPT Voice in the Desktop App to Drive Codex Agents
The GPT-Live layer, which can listen and speak at the same time, now launches and directs several coding agents at once from a single spoken instruction on Mac and Windows.
Tech
SenseTime Open-Sources SenseNova-Vision, a Unified Perception Model
A single model handles detection, segmentation, depth, keypoints, and 3D reconstruction as multimodal generation, replacing the usual collection of task-specific components.
Tech
Meta's Brain2Qwerty v2 Decodes Typed Sentences at 61% Word Accuracy
The non-invasive pipeline, based on magnetoencephalography (MEG), is a large improvement over earlier surgery-free methods, but it remains far from the below-2% error rate of surgical implants.
Tech
FlowEvo Lets LLM Agents Co-Evolve Workflows and Reusable Skills
The method turns useful reasoning procedures used at inference time into a library of executable skills, so that agents do not have to rediscover the same solutions on every task.
Tech
CARE Proposes Pre-Execution Verification for Shell-Executing LLM Agents
The paper studies command-level mediation as a runtime control point, checking each shell command before an agent runs it.
Tech
Paper Uses Reinforcement Learning to Optimize Stylistic Jailbreaks of Vision Models
Adversarial Style Optimization uses a reinforcement-learning method called Group Relative Policy Optimization (GRPO) to train stylistic triggers, aiming to make jailbreaks of multimodal models more consistent than content-based attacks.
Tech
Florida Pastor Sues OpenAI and Sam Altman Over ChatGPT Medical Advice
The complaint alleges that ChatGPT-4o described dangerous symptoms as minor hours before a near-fatal pulmonary embolism, and it accuses the company of the unauthorized practice of medicine.
Tech
Meta Opens a Paid Model API With Muse Spark 1.1 and Ships Muse Image and Video
Meta Superintelligence Labs placed Muse Spark 1.1 behind a paid API at $1.25 per million input tokens and $4.25 per million output tokens, days after releasing its first image and video generators.
Tech
Researchers Propose a Consensus Framework for Ranking LLMs on Open-Ended Tasks
The method targets situations where several answers are acceptable and correctness alone cannot distinguish the quality of responses.