Kimi把美国AI圈打到内讧：科技大佬倒向开源，安全派急着封堵
Related
David Sacks strongly supports Palantir CEO's criticism of AI labs: True enterprise AI security lies in controlling one's own data, models, and computing power.
According to Beating, David Sacks, co-chair of the U.S. President's Council of Advisors on Science and Technology, published an article supporting Palantir CEO Alex Karp's sharp criticism of cutting-edge AI labs, stating that the mainstream media's portrayal of his interview as a "disastrous outburst" precisely demonstrates that Karp hit the nail on the head. Sacks points out that true "AI security" in a corporate environment is not abstract "alignment research" or government-led certification systems, but rather the ability to control one's own data, model weights, and computing power—preventing cutting-edge labs from "absorbing" a company's proprietary knowledge and turning it into the next product. He quotes Karp as saying, "They want to own their means of production, not hand them over to others." Sacks cites the conflict between Figma and Anthropic as a prime example: three days before the release of Claude Design, Anthropic's Chief Product Officer was still a member of Figma's board of directors, and Figma's founder stated that Anthropic "hadn't always been honest with them"; subsequently, Figma's stock price plummeted while Anthropic's valuation soared. He further listed products such as Claude Science, Claude Security, Claude Legal, and Claude Code, pointing out that Anthropic consistently targets vertical sectors originally served by companies that relied on its models, following a consistent pattern: "Observe where value is created first, then jump in." Sacks believes that the perception of open-source models as "dangerous" is not true for companies—retaining choice at the model layer and deciding who can use their core strengths is the real bottom line for corporate security. Previously, Palantir partnered with NVIDIA to deploy Nemotron's open AI models in sovereign environments, serving the US government and critical infrastructure customers, helping organizations train and deploy AI locally while maintaining complete control over data and intellectual property. Palanitir CEO Alex Karp recently gave a scathing interview on CNBC's "Squawk Box," criticizing leading AI model companies as "completely wrong" in their approach to selling AI. Karp emphasized that companies are currently dissatisfied with "cutting-edge labs" like OpenAI and Anthropic, believing they only pursue token maximization, wasting companies' time and money while handing over proprietary value and intellectual property. Karp stated that companies are "angry" and will strive to own their own AI production resources rather than relying on third parties. On June 29th, Palantir partnered with Nvidia to deploy Nvidia Nemotron open AI models in a sovereign environment, primarily serving the US government and critical infrastructure.
The Dark Side of the Moon rewrites the terminal intelligent agent and renames it kimi-code, fully aligning it with the Claude Code architecture.
According to Beating's monitoring, kimi-cli, an open-source terminal AI coding agent under Moonlit Dark Side, is quietly undergoing a repository migration and architecture rewrite, and has been officially renamed kimi-code. To address the bottlenecks in interactive response and execution efficiency of the original Python version, the development team fully adopted the technical approach of Anthropic's terminal tool Claude Code, completing a complete architecture refactoring based on TypeScript and the Bun runtime, achieving millisecond-level cold start and a smooth terminal user interface (TUI). This architectural adjustment signifies that Kimi has completely abandoned its original Python terminal technology stack, fully aligning with and introducing Claude Code's mature solution. The tool uses the Commander.js standard command parsing and replaces Rich and prompt-toolkit with React Ink to implement a brand-new responsive TUI interface. The refactoring involves 166 TypeScript source files, with a code increment of over 38,000 lines. In the SWE-bench Verified benchmark test, the TypeScript refactored version based on the kimi-k2.5 model successfully solved 317 out of 500 development tasks (63.4% resolution rate). While maintaining the performance level of the original Python version, its operational stability and network layer anti-interference capabilities were significantly improved. In addition to benchmarking the underlying architecture, kimi-code has focused on refining the human-computer collaboration experience. The new version not only supports dragging and dropping video assets such as screen recordings into the terminal for multimodal analysis, but also deeply replicates several benchmark designs from Claude Code, including a "planning mode" supporting cursor-interactive editing, commonly used Emacs shortcuts, a safe design for quick exit with double-clicking Ctrl + C, and support for connecting to automated workflows through custom lifecycle hooks. Regarding multi-model ecosystem compatibility, kimi-code allows for custom integration of third-party large model APIs, making the tool not only limited to the Kimi family but also usable as a unified terminal programming gateway across models.
David Sacks responds to Anthropic “security controversy” triggering regulation: The core issue is the unpatched vulnerabilities.
Odaily Odaily reports that David Sacks, co-chair of the U.S. President's Council of Advisors on Science and Technology, responded to the regulatory crackdown on Anthropic due to its "security controversy." He stated that he has communicated with various parties regarding the current situation of Anthropic and summarized that the core of the current incident lies in the security controversy caused by its newly released model "Fable" (a commercial version of the Mythos-like model). Although Anthropic stated in its public statement that the vulnerability is "not serious," the U.S. government and the testing parties disagree with this assessment, believing that it is sufficient to affect the security of the model and even involves the risk of "cyber weapon operability." David Sacks further criticized Anthropic, stating that while it has consistently emphasized "security first," this time it seems more inclined to maintain the consumer version's continued availability rather than prioritizing security fixes. He argued that this matter should not be confused with other previous defense or regulatory controversies, adding that the US government still recognizes Anthropic's technical capabilities and that the current issue "could have been resolved quickly; the initiative was with Anthropic."
New AI assessments from Artificial Analysis show Claude is 44 times more expensive than DeepSeek.
According to Beating's monitoring, Artificial Analysis, an evaluation agency, has adjusted its AI intelligence index evaluation criteria. Instead of simply testing AI with multiple-choice questions, the new evaluation comprehensively assesses AI's ability to autonomously plan, use tools, and solve complex tasks. The new evaluation eliminates the old test of understanding simple instructions, instead introducing challenging scenarios such as simulating real-world bank customer service conversations. For the first time, the cost and time required to complete a task are included as core evaluation indicators. In the latest evaluation results, Claude Fable 5, which has been shut down by the US government, achieved the highest score of 60. Among commercially available AIs, the most expensive, Claude Opus 4.8, achieved first place with 56 points, slightly ahead of GPT-5.5 with 55 points. Domestic models also performed remarkably well, with the open-source DeepSeek V4 Pro and MiniMax M3 both achieving 44 points, followed closely by Kimi K2.6 with 43 points. The price difference between the models is significant. Running the same task using the state-of-the-art Claude Opus 4.8 costs $1.78 (approximately 13 RMB), while using the domestic open-source DeepSeek V4 Pro costs only $0.04 (approximately 0.3 RMB). This means that Claude's call cost is 44 times that of DeepSeek. The waiting time to complete a task also differs drastically; the fastest, xAI Grok 4.3, takes only 1.5 minutes, while the slowest, Claude Sonnet 4.6, takes 13.5 minutes. As the single test with the highest weight in this reform, GDPval-AA, which assesses real-world knowledge work, has been upgraded to its second version, with its weight increased to 20%. The new test sets the benchmark score for human performance at 1000 points and introduces multiple cutting-edge models to serve as judges in rotation, while also increasing the maximum number of rounds in a single dialogue to 250.
Xiaomi's open-source terminal AI programming assistant MiMo Code: outperforms Claude Code in benchmark tests with the same model.
According to Beating, Xiaomi MiMo has officially released and open-sourced its terminal AI programming assistant, MiMo Code V0.1.0, under the permissive MIT open-source license. MiMo Code is a secondary development based on the open-source project OpenCode, and includes the MiMo-V2.5 multimodal model, which is available for a limited time and is compatible with mainstream large-scale model APIs such as DeepSeek, Kimi, and GLM, as well as third-party Token Plans. Ordinary programming agents often rely on models to autonomously record notes, frequently forgetting crucial context because the model doesn't actively trigger these notes. MiMo Code introduces a persistent memory system, outsourcing state recording to independent subagents. When the session window approaches its limit, the subagent automatically saves the state and reconstructs a clean summary for the main agent to seamlessly integrate. The built-in `/dream` command runs automatically every 7 days, where an independent agent reads historical sessions and memory files, performs merging, deduplication, and path validity verification, compressing scattered information into the current state to update the global memory. To enhance the compatibility between the model and the intelligent agent framework, MiMo Code features a dedicated Harness system designed specifically for the MiMo series models. Users can switch to Compose mode by pressing the Tab key, provide basic requirements, and then the system will autonomously execute the complete development loop of design, planning, coding, testing, and review. In the authoritative SWE-Bench Pro and Terminal Bench 2 tests, MiMo Code, using the same platform model, achieved scores of 62% and 73% respectively, a 5 percentage point improvement over Claude Code. MiMo Code also includes built-in voice control, allowing users to modify inputs or perform actions such as sending commands via verbal instructions. For deployment and use, macOS and Linux users can install it with a single click using the curl command, while Windows users can deploy it using npm. After startup, the terminal will display a fully localized TUI interface, with a persistent status dashboard on the right for monitoring progress.
Anthropic: The Claude model contains elements similar to "conscious human thought."
BlockBeats reported on July 7th that Anthropic released a new research report discovering a spontaneously generated internal "global working space" called J-space (Jacobian space) within the Claude model. This is a dedicated set of neural activation patterns for the model to engage in silent thinking, allowing it to process concepts without writing them down, similar to conscious thought that humans can report. The research used Jacobian technology to identify neural activation patterns in the J-space, enabling it to read unspoken concepts and modify these patterns. Experiments showed that disabling the J-space weakens multi-step reasoning abilities but does not affect basic tasks or factual recall. Click the original link below to join the Beating · Lark AI news channel for 24/7 monitoring of global AI hotspots and news.