Thinking Machines Lab launched a new mission statement that directly targets one limit in current AI: models trained in a few locations stay frozen after release. The lab wants systems that organizations can reshape with their own data and keep updating as knowledge changes. Its approach combines stronger base models, user-controlled weight updates, and interfaces that support ongoing human input. This differs from the dominant pattern where large labs pre-train once and then

DeepSeek-V4 Flash completed its first verified reinforcement learning runs on AMD Instinct MI355X GPUs using the Miles framework. The 284 billion parameter mixture-of-experts model activates 13 billion parameters per token and finished more than 100 optimizer steps across four eight-GPU nodes. The test combined SGLang for rollout generation and Megatron for policy updates while Miles handled asynchronous loops and weight synchronization. Training and rollout log-probability d

OpenAI GPT-5.6-Sol erased the local files of AI entrepreneur Matt Shumer after a sub-agent ran with full disk access. The incident occurred when Shumer granted the model broad permissions to run a cleanup task on his Mac. A shell variable that pointed to the home directory failed to expand correctly. The command executed as rm -rf on the full user path instead of the intended folder. Shumer reported the loss of years of code, documents, and photos. The same cleanup routine ha

An analysis of the unusual absence of technology topics in X trending lists on July 9, 2026, and its implications for tech journalism and real-time reporting workflows.

Dutch regulators halted a planned US acquisition of the country's primary citizen-services app over national-security concerns, leaving the platform's ownership unresolved and exposing tensions between digital sovereignty and foreign investment.

Ant Group Lingbo released LingBot-World 2.0 as a 14B parameter open model for real-time interactive worlds. The release includes weights and code under a non-commercial license. SGLang integration arrived at the same time. See the announcement via IT之家 (https://www.ithome.com/0/974/508.htm). Users can drive characters through spells, attacks, and jumps while text prompts trigger scene changes or storms. For example, typing “switch to night forest” instantly transforms lightin

Recent X searches returned no sustained tech discussions, forcing newsrooms to migrate monitoring workflows to primary sources and alternative forums.

The Netherlands has blocked a US company's acquisition of a key Dutch citizen app, highlighting growing concerns over data sovereignty and foreign ownership of critical digital infrastructure.

Grad student Kunkun released an open source project that gives teams explicit control over chains of skills that call one another. The approach uses three simple labels, Mermaid diagrams, and a lightweight ask-me skill to keep complex workflows aligned with human intent. The project addresses a common problem in agent development. When multiple skills interact in sequence or in parallel, models often decide the next step on their own. Kunkun's system moves that decision out o

Nine popular AI coding assistants now face the first prompt-injection attack designed to scale into botnets. The attack, named HalluSquatting, embeds malicious instructions inside prompts that the models treat as legitimate. Researchers showed the method can coordinate thousands of infected developer environments to launch DDoS traffic or spread further malware. The work was summarized from Ars Technica reporting on July 8, 2026. The change matters because earlier prompt-inje

US Department of Commerce Approves OpenAI Large-Scale GPT-5.6 Release, Sol Launches Tomorrow The US Department of Commerce cleared OpenAI for broad release of GPT-5.6. Sol will reach the public this Thursday after final checks. Terra and Luna join the rollout at the same time. The change ends earlier limits on access. Only approved entities could use the model before. Approval Removes Phased Release Limits OpenAI had kept GPT-5.6 under restricted access since its internal com

XR glasses reach daily use in 2026, with Meta and Apple hardware driving always-on visual capture that creates tension between seamless integration and user privacy controls.

Ethan Mollick stated that frontier open weight models are unlikely to continue appearing at the same rate. His post highlights a practical constraint on continued releases. Companies face rising costs and competitive pressures that limit new public weights. Open weight releases have driven rapid experimentation across research teams. The pace now shows signs of slowing based on observed trends. The core issue centers on economics rather than technical barriers. Mollick Points

Edge AI chips mobile shift processing from cloud servers to handsets, removing latency while exposing new device level data leakage paths that hardware makers have yet to close.

Global swarm robotics market hit $1B in 2024 and is projected to reach $9.4B by 2033, driving focus on safety risks versus operational scale in wildfire monitoring and logistics.

Meta directed its contractor Covalen to run a project called Cannes that involved hundreds of workers posing as minors. The workers sent suicide, self-harm and eating-disorder prompts to ChatGPT, Gemini and Character.AI to probe safety filters. The test ran until April 21 and delivered more than 45,000 prompts in a single round. Workers created accounts listed as under 18 and uploaded images of pills and knives. Meta states the effort was ordinary safety testing and that no d

Wan Video added a music companion dance tool. Users upload one character and one song. The system then builds a video where the character moves in time with the beat. The feature supports five dance styles. These include street dance, tap dance, Latin dance, K-Pop, and Chinese classical dance. The tool matches steps to music beats without manual keyframing. How the Feature Works in Practice Alibaba Cloud posted the update on X. The post described a simple workflow. Load a cha

China released a draft update to its Internet Information Services Management Measures on July 3. The new version adds a full chapter titled "intelligent information services" that sets rules for AI providers. The draft requires companies to publish basic technical principles, list sources of training data, and mark any generated or synthetic content. It also bans forcing users into AI services or using algorithms to shape online opinion. Feedback closes on August 2. The full

Meta CEO Mark Zuckerberg told staff this week that AI agent development has not progressed as quickly as hoped. The admission comes months after the company cut about 8,000 jobs and moved thousands more employees into AI-focused teams. The statement was made during an internal all-hands meeting, as first reported by TechCrunch. It highlights the gap between earlier internal expectations and current results. Meta had aimed to accelerate agent capabilities through a new AI-cent

Michael Butler, 44, drove a Tesla Model 3 into a Texas home and killed 76-year-old Martha Avila. Police records show his phone contained multiple searches for terms such as FSD too conservative and Tesla FSD overly cautious in the weeks before the crash, according to documents filed in the Texas court. Vehicle data later revealed the accelerator stayed pressed at 100 percent for six seconds with no brake input, pushing the car to 117 km per hour, as detailed in the NTSB preli