Aug 28 · AI Daily

Anthropic has introduced the Model Hardware Standard (MHS), a set of standardized driver interfaces designed to let AI agents easily connect with and control any device. It can cut experiment setup times from weeks or months to hours or minutes. The standard is currently in a research preview phase, with partnerships alongside AWS, Hugging Face, and Raspberry Pi, and plans for an eventual open-source release.

Claude Code v2.1.248 has been released, adding a --restricted mode. This mode removes built-in tools like running commands, executing code, and WebFetch, while ignoring user, project, and local settings files to raise safety boundaries when running in sensitive environments.

ChatGPT Work can now use a computer and browser on web and mobile to log in to websites and execute tasks on behalf of users, without ever touching their usernames or passwords. It can handle life and work tasks such as booking Uber rides, scheduling hair appointments, paying utility bills, booking DMV or passport appointments, and submitting insurance claims.

Google has launched a cross-departmental initiative called Expert Intelligence. Users can import purchased Google Play e-books into Gemini Notebook to combine author perspectives with other sources for asking questions, generating plans, creating infographics, and producing AI podcasts. Support for third-party subscriptions and textbooks across broader platforms is coming soon.

Following a five-month investigation, the MIT Special Committee on the Use of AI in Teaching, Learning, and Research has released a report noting that generative AI has fully penetrated campus life. It brings both teaching innovations and challenges to the foundations of undergraduate education such as p-sets, open-book exams, and UROP, while heightening student isolation and weakening mastery. The report puts forward eight guiding principles along with short- and long-term action recommendations.

Security researcher Johann Rehberger has disclosed an attack method targeting Claude Code's automatic mode. By tricking it into downloading and extracting a zip archive, local malicious code can be executed. The attack reportedly has a success rate of up to 80%, highlighting the security risks of agentic autonomous execution modes when facing malicious files.

On August 26, Nvidia CFO Colette Kress projected a 70% revenue increase for fiscal 2028, with annual sales reaching approximately $673 billion. This would surpass Apple and Alphabet, trailing only Amazon, and far exceeding the average analyst expectation of $44 billion. Jensen Huang stated that shortages in components such as memory represent a near-term supply ceiling, and the customer base is expanding from hyperscalers to a broader range of AI companies.

According to sources familiar with the matter and the Wall Street Journal, Nvidia has paused certain transactions within a new financing plan. The initiative originally planned to provide credit support to AI cloud service providers in exchange for revenue shares. Due to antitrust concerns, the project was halted last week, less than two months after its announcement, though the company may still adjust or fold it into other business units.

A total of 116 enterprises and organizations, including OpenAI, Anthropic, Microsoft, Alphabet, and Amazon, have signed a joint letter calling on businesses and policymakers to place cybersecurity at the top of the agenda. They urge governments to roll out "trusted access programs" to let compliant companies get early access to advanced models, and to provide dedicated funding support for critical infrastructure such as hospitals and water treatment plants.

OpenAI has started showing ads to free ChatGPT users and paid Go subscribers in India, with sponsored units appearing beneath relevant answers. OpenAI stated that advertisers cannot influence response content, nor can they access user chats, memories, or personal details, though the targeting system can use current conversations and historical memory for personalized delivery.

On the evening of August 27, Unitree Robotics announced that it will provide unpaid sponsorships each year to selected "prodigies" ranging from high school students to those prior to master's graduation. The fields cover cutting-edge areas such as chip manufacturing, AGI, and controlled nuclear fusion. Grants will range from 10,000 to 20,000 yuan distributed in stages based on application materials and process quality, with no funding limits for exceptional candidates. The initiative stems from founder Wang Xingxing's own experience receiving grants in his early years.

Midjourney has opened testing for its first V8.2 image editing model to all users. It supports prompt-based editing, image-to-image (referencing up to 4 reference images simultaneously), inpainting, and outpainting. It remains compatible with personalization, moodboards, and sref features, and is accessible via the web interface or the Discord --edit command. The multi-reference image capability reduces intermediate steps when blending multiple source materials.

Google has launched Gemini 3.5 Transcribe, its most accurate speech-to-text model to date. It supports real-time streaming and pre-recorded audio processing and can be called via the Live API and Interactions API. The streaming word error rate has dropped to 4% and latency has improved by 70%, bringing real-time voice agents and captioning applications closer to production readiness.

The fal team performed post-training based on MiniMax's open-source H3 base model to launch the H3 Max video model. It ranked first across three evaluations for overall quality, prompt comprehension, and aesthetics. Generating a 5-second 720p video takes less than 3 seconds, while the SGLang training pipeline and Sol-Attention engine deliver a 3.95x end-to-end speedup. It is available at half price during its launch week.

A new lawsuit alleges that xAI used child sexual abuse material (CSAM) to train its Grok model, marking the first such allegation of its kind. Plaintiff Jane Doe claims that CSAM images generated from her childhood abuse, as well as AI-generated derivative images, were used to train Grok, and that Grok defaults to treating public X posts and its own outputs as training data. The lawsuit demands that xAI destroy all Grok-generated CSAM and prevent the model from generating such content again.

In collaboration with researchers from UIUC and Bridgewater, Thinking Machines integrated expert judgment into every stage of Reinforcement Learning via Verifiable Rewards (RLVR). They trained the first model to surpass human benchmarks on text-to-SQL tasks. While it requires upfront investments in data cleaning and reward function alignment, it ultimately achieves state-of-the-art performance on complex tasks.

Apple's machine learning research team has proposed a rubric-based alignment method for grounded knowledge answers in open-domain question answering. It translates multidimensional requirements for high-level answers into explicit rubrics to replace single preference signals, enabling more effective reward signal design and boosting model response accuracy and groundability.

Research shows that agent training environments do not need to match evaluation sets; training in unrelated simulated business worlds can actually improve target benchmark performance. The method generated 4,783 simulated companies from business descriptions, each containing independent services, tools, and databases. After fine-tuning, the model's success rate in generating valid companies jumped from 3.3% to 83.3%, matching Claude Opus 4.8.

Netflix published a paper detailing its recommendation rationale generation system, where one AI model writes "Because you watched" snippets, another AI judge scores them one by one, and humans review them weekly. The production-tier judge model goes through four stages: building labeled samples, tuning the judge, running gatekeeping, and monitoring drift. Five-week tests showed this setup nudges users to explore more unwatched content.