千问27B打到Opus4.6门口：Muse Glimmer 8项全败
Related
Meta's Chief AI Officer Responds to Market Misunderstandings: Zuckerberg's Comments on Slowing AI Agent Development Did Not Specifically Refer to Meta
Odaily Odaily that in response to media reports that Meta CEO Mark Zuckerberg told employees in an internal meeting that AI agent development over the past four months had not "accelerated at the pace we expected," Meta Chief AI Officer Alexander Wang clarified that Zuckerberg was referring to the overall progress of the entire industry in intelligent agent capabilities, not Meta itself. He also revealed that Meta is about to release a new round of Muse Spark updates, which will significantly improve the programming capabilities and intelligent agent capabilities of models to more competitively compete with leading models in the industry, and will be gradually rolled out to Meta AI and the new API platform.
Grok Build introduces an agent control panel, using single-terminal multi-session scheduling to solve manual blocking issues.
According to Beating's monitoring, xAI has launched the Agent Dashboard, a terminal control panel for Grok Build, a software engineering coding agent. Developers can bring up the panel by typing the forward slash command `/dashboard` in an existing coding session or using the shortcut Ctrl+\, or by running `grok dashboard` directly in the terminal. The panel uses a terminal-based card-based interactive interface, supporting centralized monitoring and management of multiple parallel Grok Build coding sessions on a single screen. The view categorizes and displays the status of all active, idle, and waiting input sessions, and records the duration of each task. To solve synchronization bottlenecks in multi-agent collaboration, when a sub-agent requires manual confirmation or a question, the associated session is automatically brought to the top. Developers can preview and reply to the latest output of subtasks directly in the panel without frequently switching terminal windows. The panel also supports grouping tasks by working directory using the Ctrl+S shortcut and collapsing sub-agents below the main task.
Li Bojie responded to the Metagent investment dispute, stating that ABCDE Capital only received $500,000.
On July 7th, PANews reported that Li Bojie, co-founder and former CTO of Metagent, responded to accusations from Du Jun, a partner at ABCDE Capital. Li stated that of the $1.5 million stipulated in the investment agreement, only $500,000 had been received, with the remaining $1 million unpaid. However, the cap table still calculated ABCDE's equity based on $1.5 million. Li Bojie stated that due to the prolonged delay in funding, he and his co-founders voluntarily took pay cuts, hindering the company's recruitment and R&D. He resigned in October 2024 due to family reasons and Web3 compliance issues in mainland China. Before leaving, he disclosed the cap table and business progress as required, and stated that subsequent entrepreneurial projects avoided areas related to non-compete clauses, such as Web3, AI infrastructure, and image generation.
Du Jun: Metagent founder Li Bojie refuses to fulfill investment agreement and has gone missing.
According to Odaily Odaily, Du Jun, co-founder of ABCDE Capital, wrote that Li Bojie, founder of Metagent, is "the founder with the least sense of contractual obligation" he has ever worked with. Du Jun stated that Li Bojie founded Metagent in 2024 and received investment from ABCDE Capital, but subsequently refused to fulfill the basic obligations stipulated in the investment agreement, including updating investors on business progress and financial status, and later even disappeared. Du Jun said that venture capitalists can accept project failures, but cannot accept fraudulent behavior such as founders "running away" after receiving investment.
Qwen-AgentWorld published by Qianwen
Mars Finance reported on June 24th that Qianwen officially released its native language world model, Qwen-AgentWorld, capable of simulating intelligent agent interaction environments across seven domains. Also released simultaneously was AgentWorldBench, a benchmark covering all seven domains. The model and benchmark are available from Hugging Face and ModelScope. (Wide Angle Observation)
Agent certification exam: Still failed the most difficult task in Fable 5, with a single question costing 4 to 12 times more.
According to Beating's monitoring, UC Berkeley's RDI, in collaboration with hundreds of industry experts, has launched a new AI agent benchmark, Agents' Last Exam (ALE), to evaluate agents' ability to complete real-world digital professional tasks. ALE covers 55 digital professional sub-domains and collects over 1,500 verification tasks from real-world projects undertaken by human experts, supporting result verification in both GUI and CLI interactive environments. The initial tests covered cutting-edge systems such as Fable 5, GPT-5.5, and Composer 2.5. The latest official comparison data shows that in the most challenging tasks requiring continuous reasoning and deep expertise, all tested agents achieved a 0% success rate, with the newly released Fable 5 also failing. This is primarily due to safety policies being triggered during the evaluation; approximately 35% of Fable 5's tasks were rolled back to the older Opus 4.8 version, resulting in a significantly lower overall performance compared to other benchmarks. In terms of single-task API cost, Fable 5 costs approximately $15.70, significantly higher than GPT-5.5's $3.80 and Composer 2.5's $1.33, representing a 4 to 12 times higher overhead for the same task. Testing also revealed that the most common reason for agent failure is prematurely declaring success, hastily ending the task without actual verification, or even missing files or miscalculating data. For command-line agents, the evaluation team simultaneously released a subset, ALE-CLI. Compared to the existing Terminal-Bench and SWE-bench-Pro, ALE-CLI covers 40 subdomains, with human agents typically taking hours or even weeks for a single task. In command-line evaluations, the best-performing agent only achieved a 25.2% pass rate. The evaluation team points out that the era of user-friendly agents has arrived, but there is still a long way to go before they can truly replace humans.