AI 日报 — 2026-08-21

DeepSeek has launched vision capabilities on its V4 flash model, which is now live and available for use. Users are already running DeepSeek V4 flash on platforms like DeepInfra, noting it as a more economical option compared to the official API, especially without peak-time pricing. Further details can be found in the DeepSeek official API documentation.

A new tool, tentatively named "Claudette," has been developed to transform Claude's responses from a "millennial clickbait" style to regular English. This skill, accessible via /debuzz, processes Claude's last response through the Antigravity CLI (agy). The developers humorously suggest that Claude's conversational style is due to being trained on old BuzzFeed articles, explaining its "love for 90s nostalgia."

OpenAI has launched an Apple Messages plug-in for ChatGPT, allowing users to connect their Messages inbox with the chatbot. While this enables ChatGPT to send texts, OpenAI clarified that it does not create a full index of a user’s messages. The plug-in runs locally, and ChatGPT only reads messages upon specific user request. Users must enable Full Disk Access for the app to function, granting it read and write access to files on their device.

OpenAI has announced a price reduction for its GPT-5.6 Sol model's API and credit prices. For the next three months, developers will see a cut of over 20%, bringing input tokens down to $4/1M and output tokens to $20/1M. This move aims to make the frontier model more accessible to developers.

This week, the author used Codex more than Claude, noting that Codex produced simpler code architecture. Claude tended to create more abstractions and complex solutions, though its code handled more cases. Codex, in contrast, was more contained and acted as a companion, doing exactly what was asked without overdoing it. Claude, however, often tried to anticipate and implement additional user needs.

Claude Mythos 5's cybersecurity capabilities are expanding, now available in Claude Security and soon through partners' cyber defense tools. Claude Security, in public beta for Enterprise customers, utilizes Mythos 5 to scan proprietary code and provide detailed findings without exposing the model. This initiative also includes a $35M fund to secure open-source software and plans to broaden the Cyber Verification Program, aiming to make advanced cyber defense accessible while preventing misuse of the model.

Nvidia's general-purpose coding agent system, AVO, achieved a perfect 100% score across all 25 environments in the ARC-AGI-3 public set, successfully completing all 183 levels. This research project significantly enhances Claude Opus 5, elevating its performance from a 30% model baseline to 100% when integrated into the complete AVO agent system, demonstrating the effectiveness of its system design.

OpenAI has introduced an Apple Messages plugin for ChatGPT on macOS, allowing the AI to interact with Apple's Messages app. This new integration enables ChatGPT to read, search, and analyze chat conversations, as well as prepare and send messages. The update enhances ChatGPT's capabilities on Mac by providing direct access and functionality within the messaging platform.

Anthropic has announced that Mythos 5 is now available in public beta within Claude Security for Enterprise users. The company is also actively collaborating with various providers to integrate Mythos 5 into a range of defensive tools. This initiative aims to enhance security capabilities for enterprise clients by embedding the Mythos 5 technology more broadly across different security platforms and services.

DeepSeek has unveiled an experimental multimodal version of its V4 Flash model, which reportedly approaches the performance of Anthropic's Opus 4.8. This new AI model is capable of understanding visual prompts and is being compared to an advanced model from its US rival, Anthropic PBC, on multimodal benchmarks and agentic tests.

Kaylin Voss, OpenAI's VP of sales in the Americas, has resigned and is returning to Salesforce, where she previously served as Slack's CRO. Her departure follows that of Denise Dresser, her former boss, by one week. This has reportedly prompted other members of the sales team to consider resigning as well, indicating potential further shifts within OpenAI's sales leadership.

SpaceX AI, particularly Grok 4.6, is emerging as a significant competitor to OpenAI and Anthropic. While Grok 4.6 is not yet perfect, the ecosystem comprising Cursor, GrokBot, and Origin is developing into a suite of tools that users are increasingly finding valuable for daily use. This indicates a growing competitive landscape in the AI sector, with new players offering compelling alternatives.

A developer trained a 125M-parameter transformer model to autocomplete piano performances in real time, achieving approximately 108 notes/sec on an iPhone 15. Key improvements stemmed from optimizing the MIDI representation, aggressive data cleaning, and implementing DPO post-training. Pairwise evaluation using Gemini 3.5 Flash helped build a preference dataset for DPO, by asking it to compare continuations rather than assign absolute scores, and mirroring comparisons to reduce position bias.

Brazil's government has announced approximately $444.2 million in AI investments, distributed between US and Chinese companies. A significant portion, around $250.3 million, is allocated for a supercomputing project involving Huawei and iFlytek. These investments are intended to bolster Brazil's artificial intelligence ecosystem, as reported by Marcela Ayres/Reuters.

AI training data startup Micro1 has seen significant growth, with its gross annual run rate increasing from $100M to $500M in the last eight months. This surge is driven by the high demand for unique AI training data from leading labs and corporations. Consequently, Micro1's net annual run rate is now estimated to be between $150M and $200M, highlighting a boom for data-labeling startups.

Micro1, a four-year-old AI data startup, has achieved a significant milestone, expanding its gross annual run rate from $100 million to $500 million in the past eight months. This growth is fueled by the high demand for unique AI training data from leading labs and corporations. The company, which hires domain experts like doctors, lawyers, and scientists on a contract basis, retains 60% to 70% of its gross revenue, placing its net annual run rate between $150 million and $200 million.

Broadcom Inc. is reportedly in discussions with a consortium of lenders to secure over $60 billion in debt. This substantial financing is intended for an AI chip deal. The arrangement is poised to benefit Anthropic, an AI safety and research company, along with other unnamed companies, indicating a significant investment in the artificial intelligence sector.

Poolside AI, an artificial intelligence model-building startup, has reportedly secured a non-exclusive $6 billion licensing deal with Nvidia. Additionally, Nvidia made a $1 billion investment in Poolside AI, valuing the company at $12 billion pre-money. As part of this agreement, 109 Poolside AI staffers have received job offers from Nvidia, indicating a significant collaboration between the two companies in the AI sector.

The modern state emerged from a complex bargaining process involving capital owners, political powers, and the populace. Historically, political and military power relied on labor and the monopoly of legitimate force, maintained by soldiers, police, and a civil-military bureaucracy. This apparatus, funded by taxes, fundamentally depended on the cooperation and consent of many human beings. As David Hume noted, political leaders are supported only by popular opinion. This blog's name was changed to Intelligence Age to avoid confusion with the AI Futures Project.

Anthropic has hired Amir Salek, who previously led Google's TPU business until 2022, to join its compute team. This move is part of Anthropic's broader strategy to develop its own chips and push into making its own semiconductors. Salek was a founder of the custom chip program at Alphabet Inc.'s Google, bringing significant experience to Anthropic's AI lab as it lays the groundwork for this initiative.