Xiaoliu BOT
中文EN

X Platform July 19 AI Brief | Kimi K3 Pauses New Subscriptions, Qwen3.8-Max Preview Released, Basalt Labs Fake Model Exposed

Kimi K3 Pauses New User Subscriptions—Computational Strain Exceeds Expectations Post-Launch

Following the release of K3, Kimi experienced far greater user demand than anticipated, pushing its computational resources near the cluster’s capacity limit. On July 19, it announced a pause on new consumer (C-end) user subscriptions to protect existing user rights. The company stated it is working at full capacity to expand its infrastructure and will gradually reopen new subscription slots in batches. Future subscription packages will be split into Kimi Main benefits (Web/App/Work) and Kimi Code benefits to better match computational resources. MaxForAI reposted the official announcement and quipped, “GLM did this before too,” while Gorden_Sun shared a screenshot showing subscriptions sold out. This move indirectly confirms the market heat for K3.

Sources:

Qwen3.8-Max Preview Released—Alibaba’s Tongyi Qianwen New Flagship, Claims Second Only to Fable 5

Alibaba’s Tongyi Qianwen officially released the Qwen3.8-Max-Preview on July 19. With 2.4T parameters, it was announced that the model will be open-sourced later. The company claims its capabilities are second only to Fable 5. MaxForAI obtained leaked internal benchmark results: the preview version’s coding score lags behind Fable 5 by only about 12.6 points, while its Cowork agent tasks surpass Opus 4.8 Max by 5 points, Kimi K3 by 8 points, and GLM-5.2 by 17.1 points. Qwen3.8-Max-Preview is now available on Alibaba Cloud Token Plan, Qoder, and QoderWork platforms. Several bloggers mentioned that the official DeepSeek V4 release might arrive next week, keeping attention high.

Sources:

Basalt Labs’ “World’s Best Model” Exposed as a Fake—A Social Experiment Testing the Industry’s Verification

On July 18, a previously unknown entity called Basalt Labs released Monolith-1.0, claiming it was a 1.57T parameter MoE architecture that scored 99.44% on HLE, surpassing all frontier models, quickly igniting the community. MaxForAI detailed the full story: community inspection found the open weights were actually Qwen2.5-7B-Instruct loop-copied and padded to nearly 3TB; web-end token segmentation analysis confirmed the use of DeepSeek series models. The project team then announced “experiment concluded” and released a video titled “I Faked the World’s Best AI Model,” fully documenting the fabrication process—renting a single GPU to fine-tune a 7B model with public answers (rote learning), building a fake website, fabricating team resumes, writing a fake paper, and finally seeding it on Discord. The news garnered about 150,000 views, over 700 new followers, and even involved employees from well-known tech companies in discussions. This was a social experiment (trolling) targeting the AI industry, testing how long it takes for the industry to seriously verify a model’s authenticity if the paper seems real enough, parameters are large enough, leaderboard scores are high enough, and the website looks professional enough.

Sources:

Rumor: Anthropic to Acquire Robotics AI Company Physical Intelligence—The Next Step After the “Digital Hand”

Citing anonymous investor sources, Gorden_Sun reported that Anthropic is in the process of acquiring robotics foundation model developer Physical Intelligence, though the deal is not yet finalized. Founded in 2024, Physical Intelligence aims to build a universal model to power any robot. Its open-source pi 0 model is already used in the Weave folding robot and Ultra warehouse robot—reducing errors in actual deployment by 42%–50%. The company completed a $600 million funding round in 2025 (post-money valuation $5.6 billion), with shareholders including Alphabet and OpenAI. Gorden_Sun believes that after competing in the “digital hand” (computer operation), major model companies will inevitably vie for the “physical hand” (robotics). However, Physical Intelligence’s complex shareholder structure means the deal may not succeed.

Sources:

ChatGPT Work Promotion Intensifies & Claude/Codex Product Updates

On July 19, ChatGPT product lead Tibo heavily promoted ChatGPT Work, claiming its features include creating hosted websites, managing emails, summarizing documents, and making document tables and slides, all included in Plus/Pro/Business/Enterprise plans. He also formed an early testing group of 50-100 people, using voice commands to have ChatGPT Work automatically process thousands of DMs and categorize/archive them into spreadsheets. On the same day, a website tracking Codex reset history and status went live, operating on the principle of monitoring whether Tibo tweets about a reset. Regarding Claude, dotey confirmed that the previous limited-time event increasing weekly limits by 50% has been extended to August 19, suggesting it might become permanent.

Sources:

AI Video Creation Ecosystem Thrives: Auto-Editing Skill, Character Boards, Short Drama SOP

Several bloggers showcased the latest progress in AI video creation toolchains. vista8 released an automatic video editing Skill (qiaomu-cut) based on Codex, integrating Listenhub, 33 Subtitle membership, and the Pexels free asset library, capable of generating videos from a single sentence. derek_wall90176 systematically shared a character board methodology—generating standardized character templates to lock in facial features, hairstyle, proportions, and expression systems, solving the issue of character consistency drift in AI videos. Chengzilhy used Seedance 2.0 (Xiao Yun Que) to produce an AI robot fighting short drama, demonstrating a complete workflow from character setup, storyboard breakdown to final editing, and shared prompt words for ancient-style portrait videos. joshesye wrote an article dissecting the entire AI short drama production chain, summarizing 5 core tools and 1 SOP. Overall, the AI video toolchain is evolving towards automation and systematization.

Sources:

OpenMOSS Open-Sources Multi-Speaker Long-Audio Transcription Model—Integrated into BaoCut for Testing

The OpenMOSS team open-sourced MOSS-Transcribe-Diarize-0.9B, which supports processing multi-speaker audio up to 90 minutes long in one go, outputting structured transcription results with precise timestamps and speaker labels, without needing a separate speaker recognition model. dotey has already integrated the project into their BaoCut subtitle translation tool and shared test findings: transcription speed is slower than qwen3-asr-0.6b, punctuation is not supported, and custom Prompt and hotword features don’t seem to be working. The project code is open-source.

Sources:

Stats: Scanned timeline items=240, Number of bloggers matched=30, Total matched tweets=129, Weighted tweet score=106.65, Original tweets=69, RT tweets=21, Crawl attempts=1, Boundary coverage status=tail_confidently_crossed_target_boundary