Source-verified updates on AI models, APIs, research, security, and regulation.
Z.ai says the GLM-5.3 API is now available for software development, defensive cybersecurity, and long-horizon agentic workflows. The company says pricing matches GLM-5.2 and access is offered through its official API and partner gateways.
Posts on social media and Chinese tech platforms claim that Claude Code has entered single-digit growth, reached saturation among programmers, and lost users to Codex. Available evidence supports a possible slowdown, but not a confirmed stall or mass migration.
A social media post claims that some AI relay platforms have removed the Luna model because its low price makes operations unprofitable. Available evidence supports broader pricing pressure but does not verify the delisting or establish Luna’s actual costs.
A set of AI tools inspired by the viral “Niu Lai” aesthetic can reportedly turn images into exaggerated bull heads, apply the style to photos, and generate short absurdist animation trailers.
Anthropic says Claude generated protein-binder designs for 14 of 15 biological targets using an expert-written design prompt. Adaptyv Bio and Twist Bioscience independently built and tested the candidates. This is an early research result, not evidence of new drugs.
Travel AI is expanding beyond itinerary advice and destination search into booking and task execution. However, reliable end-to-end ticket purchasing on platforms such as 12306 has not yet been established.
Developer Guizang has released Pilot Harness, an open-source desktop client shell and plugin suite designed to improve the everyday use of DeepSeek Harness.
Available reports describe DeepSeek Harness as a deeply modular environment in which model adapters, tools, filesystems, sessions, sandboxes, and the agent loop can be composed through plugins. A separate project, Pilot Harness, adds a more accessible desktop wrapper.
An X post uses a rewritten song format to joke about Codex, failing tests, and shipping code overnight, prompting similar posts.
A recent user test suggests that Doubao’s PC app can operate desktop interfaces through a virtual workspace while allowing users to monitor and direct tasks remotely from a phone. Windows appears to offer broader support than Mac.
Reports based on investor materials say Anthropic generated more than $11.5 billion in preliminary second-quarter revenue, roughly 14.6 times the year-earlier figure, while adjusted operating profit turned positive. Its reported revenue run rate exceeded $65 billion by late July, although the figures remain unaudited and based largely on private disclosures.
A user says a newly added Claude Code feature frequently sends messages to other active sessions, often incorrectly and with little practical value. The feature is reportedly enabled by default, but its disable setting remains unclear.
A social media post promotes a fitness-themed community challenge built around references to open weights, Vercel, v0 and MiniMax. The advertised activities include pull-ups, planks and a chest-focused challenge for founders and engineers.
Multiple reports say Meituan encouraged broad employee experimentation with OpenClaw and related AI tools early in the year, but faced rapidly rising costs and operational errors. The company then shifted toward dedicated AI teams, project selection, and workflow integration.
An open-ended prompt asks users to identify Codex, API, or model capabilities that appear feasible today but have not yet been shipped.
Third-party testing of DeepSeek V4 Pro, Pro Max, and V4 Flash suggests that the cheaper Flash model may match or outperform Pro on selected coding, agent, and latency tasks. Other reports still favor Pro for difficult reasoning and engineering work.
Multiple reports place Qwen3.8-27B at 52 on the Artificial Analysis Intelligence Index under maximum reasoning, putting it in the same broad score range as several recent frontier models. Its ability to run locally on relatively modest hardware has intensified interest in the release.
A community creator demonstrated local video generation with MiniMax H3 and ComfyUI on a 24GB RTX 4090, covering text-to-video, beat synchronization, and image-to-video workflows.
MiniMax says its H3 video model can capture character movement well and demonstrates a workflow for turning short motion clips into game sprites. The process uses separate clips for actions such as walking, jumping, attacking, and blocking, followed by pose extraction and sprite-atlas assembly.
Cursor has begun rolling out Origin, a code-hosting service integrated with its editor and AI-agent workflows. The platform supports GitHub repository synchronization and brings repository management, pull requests, and review closer to the Cursor environment.
适合多赞助商长期展示
