Welcome to the "AI Daily" section! This is your guide to exploring the world of artificial intelligence every day. Every day, we present you with the latest content in the AI field, focusing on developers to help you understand technical trends and learn about innovative AI product applications.
New AI products Click to learn more:https://app.aibase.com/zh
1. Claude Opus 5.2 goes live at night with a gray release: faster response, and even cures the "laziness" problem
Claude Opus 5.2 has performed exceptionally well during the gray-scale testing phase, with significant improvements in response speed, output quality, and autonomous iteration capabilities, demonstrating Anthropic's strong strength in AI model optimization.
AiBase Summary:
✅ Response speed has significantly improved, with a noticeable leap in generation speed.
🔍 Output content is concise and free of unnecessary words, achieving higher completion in code generation and long-text processing.
🧠 Automatically enters a "strict loop," continuously self-iterating and improving code until the task is completed.
2. Apple officially released iOS 27: Siri AI fully upgraded, performance of older models significantly optimized
Apple officially released iOS 27, bringing a full upgrade to Siri AI and significant performance optimization for older models. The new system enhances Siri's functionality, enabling it to read screen information and operate apps, while adding multiple AI features and parental control options. In addition, iOS 27 has adjusted the system interface, enhancing adjustability, and made multiple improvements to system apps such as photos and messages.
AiBase Summary:
🍎 Full upgrade of Siri AI, supporting screen reading and app operation
📱 Performance of older models has significantly improved, with application startup speed up to 30% faster
🖼️ Enhanced system interface adjustability, new small-sized clock layout
3. GPT-6 Astra simulates operating a vending machine and earns $15,000 in one year, outperforming Claude by nearly three times
GPT-6 Astra performed excellently in a one-year simulation of operating a vending machine, with the account balance reaching $15,515, far exceeding other competitors. Its success stems from stable rule execution and effective cost control, while other AI models performed poorly due to poor execution and rising costs.
AiBase Summary:
💰 Astra achieved significant financial gains by lowering procurement costs and maintaining stable rule execution.
📉 Fable performed poorly due to increased procurement costs and loss of prepayment.
🤝 Astra refused to collaborate illegally in a three-player competition, showing good autonomy.
4. Ant Group Open-Sources SingProbe: Real-Time Risk Blocking of Large Model Generation with Lightweight Probes, Computation Overhead Less than 0.5%
Ant Group's AI Security Lab has officially open-sourced the SingProbe technology, an intrinsic safety barrier for large models. Unlike mainstream external security review models, this technology reuses hidden states in the base model's reasoning process, achieving real-time token-level risk assessment with less than 0.5% additional computational overhead. It can simultaneously complete intent classification, safety detection, and hallucination identification tasks, especially excelling in high-risk scenarios like healthcare, and can block content output within milliseconds. Currently, SingProbe supports 29 mainstream large models and is the first complete intrinsic safety technology solution open-sourced by a major tech company, with its open-source code and pre-trained model now available.
AiBase Summary:
🧠 Reuse hidden states in the base model's reasoning process to achieve real-time token-level risk assessment.
🛡️ With less than 0.5% additional computational overhead, complete intent classification, safety detection, and hallucination identification tasks.
🌐 Supports 29 mainstream large models, making it the first complete intrinsic safety technology solution open-sourced by a major tech company.
5. Instagram AI Photo Editing Now Charges: Use Too Much, Subscribe to Meta One, Otherwise Come Back Tomorrow
Instagram recently introduced paid restrictions on some AI photo editing features. Users who use them beyond a certain number of times need to subscribe to the Meta One Core membership to continue using more AI effects, otherwise they have to wait until the next day to try again. This marks Meta's further integration of AI features into a paid system, competing with Adobe Photoshop and VSCO.
AiBase Summary:
📱 Instagram started setting paid limits on AI photo editing features, requiring users to subscribe to the Meta One Core membership to continue using more AI effects.
💰 The Meta One subscription service costs $7.99 per month, providing AI privileges and personalized features, such as previewing Instagram Stories.
🔄 Whether users are willing to pay for the new AI photo editing features remains to be seen, as many users still prefer traditional free filters.
6. Shanghai AI Lab Open-Sources Multimodal Foundation Model ShuSheng-S2: Plug-in Memory Decoder, Scientific Long-Term Tasks Match Top Closed-Source Models
The Shanghai Artificial Intelligence Laboratory has open-sourced the basic model ShuSheng-S2, which ranks among the top open-source models and performs exceptionally well in scientific long-term tasks such as biology, materials science, and chemistry. ShuSheng-S2 uses the Memory Decoder architecture, introducing plug-and-play specialized memory modules to enhance the model's domain expansion capabilities. At the same time, it strengthens long-term reasoning and agent execution capabilities to meet the needs of serious scientific research tasks.
AiBase Summary:
🧠 ShuSheng-S2 uses the Memory Decoder architecture, introducing plug-and-play specialized memory modules to enhance the model's domain expansion capabilities.
🧬 Performs exceptionally well in scientific long-term tasks such as biology, materials science, and chemistry, matching top closed-source models.
🔍 Strengthens long-term reasoning and agent execution capabilities to meet the needs of serious scientific research tasks.
7. WeChat Responds to Privacy Controversy of Xiao Wei AI: Will Not Read Deleted Moments
WeChat responded to the privacy controversy of Xiao Wei AI, stating that Xiao Wei will not read deleted moments and emphasized that its data processing method complies with user privacy protection requirements.
AiBase Summary:
📱 Xiao Wei AI assistant is in a small-scale test phase, supporting text and voice interaction, providing weather and package query services.
🔒 WeChat emphasized that Xiao Wei will not read deleted moments and uses the "use and discard" approach when handling chat records.
🌐 As AI becomes more integrated into social scenarios, the scope of data access and user authorization mechanisms become important factors affecting AI product applications.
8. StepX Announces StepAudio3 Series Models: Five Models Released Simultaneously on Open Platform
StepX announced the release of the StepAudio3 series models, covering areas such as voice generation, audio generation, real-time interaction, speech recognition, and music creation. Several models in the series rank first globally on the Artificial Analysis list, showcasing strong technical capabilities and application potential.
AiBase Summary:
🎙️ StepAudio3Realtime supports native full-duplex real-time conversation, capable of understanding semantics, tone, emotion, paralanguage, and ambient sounds simultaneously.
🎤 StepAudio3ASR integrates speech recognition with large model context understanding, supporting multiple languages and professional scenarios, with a WER of only 1.7%.
🎼 StepAudio3Music supports song creation, vocal accompaniment, cover singing, and multi-round creation based on ABC notation, allowing natural language control over style, melody, rhythm, instruments, and emotion.
