AI 日报 — 2026-08-15

Google has introduced HEIR, an open-source compiler within its Private Computing Toolkit, designed to enable cryptographically-secure private AI inference. Since its announcement in 2023, HEIR has fostered collaborations with hardware accelerator developers like Belfort and Cornami, and academic institutions including Georgia Tech and Tsinghua University. This has led to several peer-reviewed publications and numerous citations, demonstrating its role as a productive research platform. Google aims to make homomorphic encryption easy to develop, fast to run, and ubiquitous across the industry.

A developer created a desk HUD displaying live Claude usage on a $38 Thermalright Trofeo Vision 6.86" LCD (1280×480, USB-C) from macOS. This project was inspired by a Reddit post about a similar Claude LCD display. The LCD is a USB HID device (VID:PID 0416:5302) that accepts JPEG frames via a reverse-engineered protocol. The HUD continuously streams data at 2 fps to prevent the display from blanking when idle, utilizing device classes from thermalright-trcc-linux with HidApiTransport.

ThoughtDAG introduces an editable context graph for LLM conversations, addressing the issue of hidden context in linear chat histories. It allows users to visualize, edit, and inspect the context sent to the model, preventing "polluted summaries" by enabling the removal of irrelevant information, such as a "dinner detour." This system ensures reproducible context and cleaner answers by making the "wires" of context visible and editable, contrasting with traditional chat interfaces where the history entering the next request is often invisible.

Anthropic has detailed Claude's text watermark, which indicates the likelihood of Claude's involvement in generating text. The watermark is sparse in code and factual text, and it disappears completely after a full rewrite. This feature is designed to show that Claude was likely involved in the text's creation, rather than definitively proving it, and it will be present in future Claude models.

Twitch has updated its account settings, allowing streamers to opt out of having their content used to train Amazon's artificial intelligence models. Since March 2024, Twitch's Terms of Service have granted the platform and its sublicensees rights to use and adapt user content, though explicit mention of generative AI training was previously absent. This change comes as available data sources for AI training dwindle, prompting a renewed debate on the ethics and transparency of such processes.

Alibaba Group Holding's open-weight models have achieved over 3 billion global downloads in the last six months. This figure significantly surpasses the download numbers reported for other tech giants. According to Hugging Face, Google recorded 418 million downloads, while Meta had 227 million downloads in 2026, highlighting Alibaba's substantial lead in this domain.

The Qwen 3.8 27B model repository on Hugging Face provides FP8-quantized weights and configuration files for a post-trained model in the Transformers format. It specifies parameters for an "Instruct" mode, including temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, and repetition_penalty=1.0. The model's output structure includes appending messages with roles like "assistant" and content fields such as "answer_content" and "reasoning_content."

Yadda 3.0.0 has been released, focusing on modernization and AI agent compatibility. This Node-only version removes outdated integrations like CasperJS and PhantomJS, updates to ES6 syntax, and incorporates tools such as Biome and lefthook. It also includes TypeScript definitions and current examples like Playwright and Puppeteer. A significant update in Yadda v3.1.0 introduces support for writing feature specifications in GitHub-flavoured Markdown, enhancing readability and integration with project documentation for both humans and AI agents.

Anthropic has provided further details on how its Claude chatbot will implement watermarking for generated text. The company plans to utilize the SynthID-Text approach, developed by Google DeepMind in 2024, and will release a watermark detection API. Anthropic emphasizes that this watermarking method is distinct from AI detection techniques that identify stylistic patterns in writing, such as those offered by companies like Pangram, as watermarking involves directly embedding a signal rather than analyzing writing tells.

The U.S. is preparing to inform 35 partner countries that participation in China's rival artificial intelligence framework will result in their exclusion from the U.S.-led Pax Silica initiative. This move, reported by Michael Martina/Reuters, indicates a U.S. strategy to compel countries to choose sides in the ongoing artificial intelligence competition between the two global powers. The communication will be delivered via a source and letter, emphasizing the exclusivity of the Pax Silica initiative.

Google now allows users to remove visible watermarks from AI-generated images, videos, and music created with tools like Gemini and Flow. This is done by toggling off a new "Media watermark" setting, which removes the "sparkle" watermark from content generated by Nano Banana and Omni models. However, invisible SynthID watermarks and C2PA metadata will still be embedded, allowing verification of AI origin through Gemini or Search. This change aligns Google with other AI generators like OpenAI and Meta, which also do not apply visible watermarks.

Google has introduced a new toggle in Gemini and Flow, allowing users to remove visible watermarks from AI-generated images, videos, and music. While the visible watermarks can now be removed, Google will continue to embed invisible SynthID watermarks and C2PA metadata into these AI-generated media files. This update provides users with more control over the appearance of their AI-generated content while still maintaining a level of traceability through embedded metadata.

Netflix's GenRec system aims for LLM-native recommendations, prioritizing long-term member satisfaction and business requirements like balancing diverse content (movies, series, games, live, podcasts) over immediate clicks. It uses catalog-aware scoring where each item has a learned embedding, and a scoring head combines this with a user representation to produce a score. These scores are then converted into a probability distribution and a ranking.

OpenAI has introduced "Ultrafast" mode for its GPT-5.6 Sol model, enabling it to run up to 14 times faster than standard processing. This new service tier, powered by Cerebras, generates up to 750 output tokens per second and is initially available in a limited preview through the OpenAI API. The company plans to expand access as capacity increases, aiming to bring its most intelligent model to applications where speed is critical.

Nvidia is reportedly in discussions to invest up to $3 billion in SB Energy, a data center developer backed by SoftBank Group. SB Energy is planning an initial public offering (IPO) in the near future and is also involved in the development of a large OpenAI campus located in Ohio. These talks highlight potential significant investments in data infrastructure and AI-related projects.

Despite the turmoil caused by SBF, effective altruism is attracting record funding. This surge is largely driven by the anticipated IPOs of Anthropic and OpenAI, which are expected to create new millionaires committed to "effective giving." These new philanthropists are poised to inject significant capital into the movement, reinforcing its principles even after recent controversies.

SpaceX has finalized its $60 billion acquisition of Cursor, an artificial intelligence coding startup. This acquisition comes two months after SpaceX initially announced its agreement to acquire the company. The move is a significant part of Elon Musk's strategy to compete more effectively with rivals such as Anthropic PBC and OpenAI in the AI sector.

Anthropic is projecting substantial growth, with anticipated 2028 revenues of approximately $190B-$200B. This projection comes as bankers and investors are evaluating the company for a potential IPO. Currently, Anthropic's revenue run rate stands at $47B as of May, indicating a significant expected increase in its financial performance over the next few years.

Chinese AI startup Z.ai announced that its open-source GLM-5.3 model achieved a score of 84.5% on CyberGym, closely trailing Anthropic's restricted Mythos 5, which scored 83.8%. Z.ai also stated that the most sensitive cybersecurity functions of GLM-5.3 will be exclusively accessible to verified users, indicating a focus on security and controlled access for advanced features.

OpenAI is reportedly facing staff frustration due to repeated executive reshuffles and departures as it prepares for an IPO. The company also disbanded its "preparedness" team in July, which has contributed to unsettling staff. These internal changes are occurring while the ChatGPT maker gears up for a significant public listing, as reported by the Financial Times.

SF-based Vals, a company focused on developing evaluations and benchmarks for testing AI models on real-world tasks, recently secured $40 million in a Series A funding round. This investment, led by a16z, values the company at $400 million. Vals AI has also reported significant growth, with its revenue increasing eightfold since 2025.

Anthropic informed investors that its Q2 revenue exceeded $11.5 billion, marking a significant increase from $787 million in Q2 2025 and $4.73 billion in Q1 2026. The company also reported a positive adjusted operating income for Q2. This indicates a substantial growth, with revenue jumping at least 14-fold compared to the same period last year, according to Bloomberg.

Nvidia has reportedly reworked its financing deal for an OpenAI data center campus in Ohio. The initial plan for a $250B backstop has been adjusted, with Nvidia now guaranteeing only half of that amount. This change affects the 10-gigawatt project, as reported by Anissa Gardizy in the Wall Street Journal, based on unnamed sources.

Firms like Mercor, which gather data for AI labs, are increasing the demand to acquire or license internal datasets from startups that are either closing down or being acquired. This trend is exemplified by an unusual email received by Warmly CEO Maximus Greenwald, eight days after his AI agent startup agreed to be acquired by HubSpot in late June, indicating a growing interest in such data assets.

Thrive Capital's 2022 early-stage venture capital fund, which initially raised $516 million, has seen substantial growth. As of the end of June, this fund, known for its early investments in companies like OpenAI and SpaceX, was valued at over $3.7 billion, after accounting for all fees and other charges. This impressive valuation highlights the success of its strategic early-stage bets.

OpenAI's CFO, Sarah Friar, informed investors that the company's enterprise business now surpasses its ChatGPT-led consumer business in revenue generation. This significant shift was highlighted during a meeting with investors, where it was also noted that enterprise customers experienced a 32% growth in July. This financial update comes amidst a week of leadership changes within OpenAI's C-suite.

Mole is a deep-research agent designed for terminals, featuring an enforced budget and a privacy boundary for local data. It boasts a 0% budget overshoot, 100% claim integrity with verified quotes and sources, and 100% citation accuracy. The grounding rate is 80%, with a precision of 70% with the confirm pass (51% without). Merge precision/recall stands at 1.000/1.000 on constructed ground truth. Mole is released under the Apache-2.0 license.

Apple has collaborated with Alibaba to develop its own AI model specifically for the Chinese market. This move gives Apple an advantage within China's complex regulatory environment, especially amid increasing tensions in AI development. According to Reuters, if the model receives Chinese government approval and is publicly released, Apple would become the first American company to offer a proprietary AI model in China.

Anthropic has updated its misalignment risk estimate from "very low" to "low," according to a risk report. The company also stated that it does not intend to release an internal model known as "Model 2." This decision comes despite "Model 2" appearing to be more powerful than their current top-tier model, Mythos. The report, covered by Madison Mills for Axios, highlights Anthropic's cautious approach to deploying advanced AI models.