OpenAI Launches GPT-5.6 Sol/Terra/Luna, Upgrades Codex to Super AppBoko Haram Uses Frontier AI, Report FindsReuters: China May Restrict Foreign Access to Its Open-Source AI ModelsDeutsche Telekom Partners with OpenAI to Transform Customer Service and Network OpsAMD Ryzen AI Halo Processor Tested: Strong PerformanceEPFL Shows AI-Generated Videos Can Maximally Drive a Target Brain RegionTail-Aware Credit Calibration Prevents LLM RL from Rewarding Implausible TokensDeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable EnvironmentUncertainty-Gated Selection Fixes Myopic Threshold Cutoff in Block-Sparse AttentionPLURAL Dataset Released: Global Value Alignment Covering 90+ RegionsNew Method Replaces LLM Agent Inference Coding with Tool-Making Pipeline to Reduce LatencyThe Harness Effect: How Orchestration Design Shapes Tokenomics for Enterprise Agentic AIVercel Logs Adds Tree and Waterfall Views to Spot Slowest RequestsBetter tools made Copilot code review worse, here's how we fixed itOpenAI Launches 'Work' Feature: Desktop Accesses Local Files, Cloud Version Doesn'tClaude Code v2.1.206: Directory Suggestions, CLAUDE.md Trim CheckIs Vim Still Relevant in the AI Era? Developer Community DebatesNiraj Patel: AR Glasses Need Eye-Level Camera for Continuous RecordingOpenAI Codex Releases 0.145.0-alpha.4, Continues Alpha TestingOpenClaw v2026.7.1-beta.3 Released, Under TestingOpenAI Launches GPT-5.6 Sol/Terra/Luna, Upgrades Codex to Super App

OpenAI released three GPT-5.6 models: Sol, Terra, and Luna, likely tailored for different use cases. Concurrently, Codex has been upgraded into a ChatGPT super app with expanded capabilities. Specific model specs and pricing remain undisclosed, but the releases signal OpenAI’s push toward multi-model deployment and a unified super-app strategy.

A report from CASP reveals that the Nigerian terrorist group Boko Haram has been using frontier AI technologies. This is the first public study showing systematic exploitation of large language models by an extremist organization for terror activities. The findings have reignited debates on AI model regulation and misuse, though specific abuse methods remain undisclosed.

Deutsche Telekom is partnering with OpenAI to become an AI-native telco, applying AI to customer service, employee workflows, network operations, and future voice interfaces. This represents a deep integration between a major telecom operator and frontier AI, potentially setting a standard for AI deployment in telecommunications infrastructure.

The AMD Ryzen AI Halo processor has been put through hands-on testing. This chip integrates a powerful AI compute unit, targeting high-performance local inference for AI PCs. Reviews show strong performance in both generative AI tasks and traditional CPU/GPU workloads, making it ideal for users who need to run large models locally.

EPFL has unveiled Project NeVo, demonstrating that AI-generated videos can be optimized to maximally stimulate activity in a specific target brain region. The technology could advance neuroscience research by helping understand visual processing, and may eventually enable personalized visual stimulation or neurofeedback therapy. The project has generated significant discussion.

A new study identifies a key flaw in current LLM reinforcement learning: critic-free RL methods uniformly assign the same advantage to all tokens, inadvertently reinforcing implausible, low-probability tokens. The researchers propose ‘Tail-Aware Credit Calibration’, which dynamically adjusts reward distribution based on the tail characteristics of the token probability distribution, thereby improving model reasoning capabilities.

A new study introduces DeepSearch-World, a framework addressing the challenge of training tool-use agents. Supervised fine-tuning relies on fixed teacher trajectories, while RL suffers from sparse rewards. This method allows agents to explore in a verifiable search environment, then self-distill from their own successful trajectories to continuously improve performance without relying on an external teacher model.

Block-sparse attention reduces computation in long-context models by selecting only the top-k key blocks. However, research found that when the k-th and (k+1)-th blocks have nearly identical scores, a simple cutoff leads to information loss. The proposed ‘uncertainty-gated selection’ method introduces an uncertainty metric and dynamically adjusts the selection when scores are close, preserving important information while maintaining sparsity efficiency.

Addressing the issue that LLMs disproportionately reflect Western values, researchers have released the PLURAL dataset. This large-scale, value-focused preference dataset is grounded in the Integrated Value Theory and covers over 90 countries and regions worldwide. It can be used to train and evaluate LLMs’ value alignment capabilities, helping models better understand and respect diverse cultural value systems.

A new study proposes a tool-making approach for LLM agents in low-latency systems. Traditional agents waste latency by regenerating code for the same procedural steps on every request. The new method compiles repeated standard operating steps into validated tools, replacing the inference-time coding loop with a one-time tool-making pipeline, significantly reducing latency and improving reliability.

A new arXiv preprint (2607.06906v1) introduces the “Harness Effect,” revealing how orchestration design redefines tokenomics for enterprise agentic AI. Current agentic AI relies on token maximization—consuming tokens for longer reasoning chains, more interaction rounds, broader tool loads, and larger replay contexts—causing per-task token consumption to outpace task value growth. Although unit token prices decline, this masks real cost inflation. The paper shows that structured task decomposition, context reuse, and tool call optimization can reduce per-task token consumption by 40-60% while maintaining or improving task quality. These findings provide a critical methodology for cost control and efficiency in large-scale enterprise agentic AI deployments.

Vercel introduces Tree and Waterfall views in Logs for visualizing span hierarchies and request timing. The Tree view groups spans by parent-child and sorts by duration (slowest first), while the Waterfall view displays all spans on a single timeline. Developers can inspect critical paths and latency without leaving the log entry, improving debugging for backend and full-stack applications.

GitHub blog shared an improvement experience on Copilot code review: after migrating to a shared Unix-style code exploration tool and reshaping the agent workflow around pull request evidence, the review cost was significantly reduced. Specifically, the new approach cut review costs by approximately 40% while improving the accuracy and efficiency of code review. The article noted that initially introducing more powerful tools actually degraded review quality due to a mismatch between tools and workflow. By redesigning the agent process to focus more on specific evidence in pull requests (such as diffs and context), the team achieved more efficient reviews. This improvement not only optimized Copilot’s code review functionality but also provided important insights for the design of AI-assisted development tools.

OpenAI has launched a new feature called ‘Work’. When used on web or mobile, conversations run in the cloud and cannot access local data. In the desktop app, with user permission, Work can read local files and interact with other desktop apps. Currently, cloud conversations do not appear in the desktop Work interface; desktop threads and local files remain on the local machine. This feature establishes a clear data boundary between cloud and local processing.

Claude Code v2.1.206 is released. Key features: /cd command now suggests directory paths, aligning with /add-dir; a new /doctor check proposes trimming CLAUDE.md content that Claude can infer from the codebase; /commit-push-pr auto-allows git push to the configured remote. These updates improve code navigation and project management efficiency.

A long-time Vim user sparked a discussion on Hacker News: As developers increasingly interact with code through AI chat interfaces and spend less time reading code, does it still make sense to use Vim? The post received 72 comments with divided opinions. Some argue Vim’s editing efficiency remains irreplaceable, while others believe AI tools are fundamentally changing how we code.

Niraj Patel points out that building augmented reality (AR) glasses requires a camera placed next to the eye to continuously record everything the user sees and process that data to overlay digital content. There is no way to circumvent this need. Additionally, no chip currently exists that is small enough to be embedded in the temple of the glasses. This means camera and real-time data processing are unavoidable core components for achieving AR glasses functionality, while miniaturized chip technology has not yet met this demand.

OpenAI Codex released version 0.145.0-alpha.4, following 0.145.0-alpha.3. This is a minor alpha update with no specific feature changelog yet. The continued alpha releases indicate active iteration by the Codex team, with potentially more developer-facing features to come.

OpenClaw v2026.7.1-beta.3 is released. This is a beta test version of the OpenClaw project, following a date-based versioning scheme (2026.7.1). No specific changelog is provided; it is part of ongoing incremental releases.