2026-07-03
AI Briefings Daily Report.

Anthropic releases Claude Science and announces independent drug development ↗

The AI drug boom has a long way to go before reaching patients.

Bridgewater Associates joint research: frontier AI model financial judgment accuracy only 50-70%, fine-tuned open-source model reaches 84.7% ↗

Bridgewater Associates joint research team testing found that frontier large models such as GPT and Claude have only about 50% accuracy on basic financial information screening tasks, and only 70% after optimization, not reaching the 80% credible threshold. Meanwhile, a self-developed model fine-tuned from open-source models achieved 84.7% accuracy, at only one-fourteenth the cost of frontier models.

DeepSeek-V4 factory-direct model lands on Tencent Cloud in mid-July, simultaneously introducing peak-valley pricing ↗

Tencent Cloud announced it will launch the DeepSeek-V4 original-supplier direct official version in mid-July, and simultaneously introduce the official website's peak-valley pricing mechanism. During peak hours (daily 9-12, 14-18), prices double, covering the Pro and Flash two versions. Enterprise edition point deduction rules are also adjusted accordingly.

Two departments issue document encouraging diversified operation of cinemas, supporting screening of virtual reality films and introduction of AI agents ↗

The National Film Bureau and the State Administration for Market Regulation jointly issued a document encouraging cinemas to carry out business model innovation and equipment upgrades, explicitly supporting the screening of virtual reality films. Future cinemas will not only be viewing venues, but may also integrate catering, cultural creativity, gaming and other business forms, creating new cultural consumption landmarks.

World's first AI Agent ransomware attack exposed, from vulnerability exploitation to database encryption completed autonomously throughout ↗

Security vendor Sysdig recorded for the first time a ransomware attack autonomously executed by an AI Agent, using a combination of known vulnerabilities, from intrusion and lateral movement to database encryption entirely without human intervention. The attack demonstrated AI's autonomous decision-making and remediation capabilities, executing over 600 attack payloads. Enterprises need to beware of the risk that AI lowers the barrier to attack.

Sources say Alibaba will ban Claude Code in the workplace due to suspected backdoor risk ↗

Bridgewater Fine-Tunes Qwen3-235B with Thinking Machines Lab, Surpassing GPT and Claude in Financial Document Analysis ↗

Bridgewater and Thinking Machines Lab—the startup from former OpenAI CTO Mira Murati—have fine-tuned a Qwen3-235B model for financial tasks. According to their own testing, the model hits 84.7 percent accuracy, beating Gemini, Claude, and GPT at roughly one-fourteenth of the cost. The numbers haven't been verified by anyone outside the two companies, though.

Takeda and Insilico sign $600 million AI drug discovery agreement ↗

Checking the site connection security

The Cyberspace Administration of China again solicits opinions on the 'Internet Information Service Management Measures', first establishing"Intelligent information service"Special chapter regulates AI services ↗

The draft adds a dedicated chapter on 'intelligent information services', requiring AI-generated content to be labeled, and algorithm recommendation to establish a human intervention mechanism. It also strengthens user account management, clarifies platform responsibilities, and establishes a serious trust-breaking list system. Feedback deadline is August 2.

Wallbox Intelligence Releases ForgeTrain, an AI-Automated Pre-training Framework That Catches Up to Megatron-LM in 8 Hours ↗

The intelligence revolution is moving from "using AI to replace mental labor" to "AI manufacturing AI".

Meta admits AI agent progress slower than expected, executives misjudged timeline ↗

Mark Zuckerberg admitted to weaknesses in the company's restructuring during an internal town hall. The AI agents Meta reorganized around are progressing slower than planned, Zuckerberg said. His AI chief, meanwhile, painted a rosier picture.

Tesla sets employee AI weekly spending cap at 200 US dollars, Musk urges use of Grok but employees prefer Claude ↗

Tesla caps employee AI spending at $200 per week, according to an internal memo reported by The Information.

GPT-5.6 Sol expected to release next Thursday ↗

Jump to main content 404 The page you visited does not exist, or the content is no longer public. Back to curated Browse all updates Curated Daily reports more

Takeda and Insilico sign $600 million AI drug discovery agreement ↗

Takeda and Hong Kong's Insilico Medicine reach strategic cooperation to use AI for early drug discovery. Insilico's Pharma.AI platform (including PandaOmics, Chemistry42, InClinico) will be used for target identification, molecular design and clinical trial prediction. The agreement includes approximately $60 million upfront and milestones, with total potential value of approximately $600 million. Takeda obtains global exclusive development, manufacturing and commercialization rights, and Insilico may receive tiered royalties. Insilico's self-developed AI drug Rentosertib (TNIK inhibitor, for idiopathic pulmonary fibrosis) has entered phase 2a clinical trial.

Alibaba internally bans Claude Code across the board ↗

Jump to main content 404 The page you visited does not exist, or the content is no longer public. Back to curated Browse all updates Curated Daily reports more

Apache Fluss 1.0 Released: Real-Time Context Empowers AI Agents ↗

Jump to main content 404 The page you visited does not exist, or the content is no longer public. Back to curated Browse all updates Curated Daily reports more

Apache Paimon 2.0: Streaming Lakehouse Evolves into AI Multimodal Foundation ↗

Jump to main content 404 The page you visited does not exist, or the content is no longer public. Back to curated Browse all updates Curated Daily reports more

The Cyberspace Administration of China again solicits opinions on the 'Internet Information Service Management Measures', first establishing"Intelligent information service"Special chapter regulates AI services ↗

On July 3, the Cyberspace Administration of China once again sought public comments on the Draft for Comments on the Revision of the Measures for the Administration of Internet Information Services. The draft adds a new chapter on "Intelligent Information Services," requiring AI service providers to disclose the basic principles of their technology and the sources of training data, label generated or synthesized content, and prohibit forcing users to use intelligent services or using algorithms to disrupt online public opinion. The draft also strengthens user account management, clarifying that accounts inactive for over six months may be cancelled according to agreement; it requires platforms to establish a feature database for cyberbullying information and provide protection options such as blocking and prohibiting reposting. The deadline for feedback is August 2.

Wallbox Intelligence Releases ForgeTrain, an AI-Automated Pre-training Framework That Catches Up to Megatron-LM in 8 Hours ↗

Wallbox Intelligence has released ForgeTrain, the world's first production-grade large model pre-training framework written entirely by AI without human intervention. The framework automatically "forges" specialized training code from scratch tailored to specific models and hardware. Benchmarks show that ForgeTrain catches up to Megatron-LM within 8 hours, achieves stable outperformance within 1.5 to 2 days, improves model FLOPS utilization by approximately 8%–10%, and is transferable to different models (MiniCPM4-0.5B/8B) and hardware (H100 and Ascend NPU). It employs a four-stage Harness optimization process with fully automated decision-making. Wallbox Intelligence summarizes its engineering philosophy as "Forge Engineering."

Meta frontier model Watermelon surpasses GPT-5.5, computing power one order of magnitude higher than Muse Spark ↗

Jump to main content 404 The page you visited does not exist, or the content is no longer public. Back to curated Browse all updates Curated Daily reports more