Tech News Overview
Today’s coverage spans several high-value topics: the AI field sees a flurry of releases such as DeepSeek-V4-Pro and Gemini 3.7 Flash, while AI’s role on both the offensive and defensive sides of cybersecurity becomes increasingly prominent; in security, large-scale supply chain attacks and multiple new defense technologies are both in the spotlight; Chinese scientists have achieved leading breakthroughs in high-temperature superconductivity and quantum communication; on the infrastructure side, there are engineering stories such as a 16-year-old vulnerability in SQLite being uncovered. Below is today’s curated summary.
🤖 AI and Machine Learning
DeepSeek officially releases V4-Pro, with comprehensively upgraded Agent capabilities
DeepSeek today announced the launch of its V4-Pro large model, centered on major upgrades for Agent scenarios. The new model refines flexible reasoning control: simple tasks can use low reasoning mode, daily Agent workflows use high mode, and complex tasks support the highest-strength reasoning budget. It also natively supports the OpenAI Responses API and offers out-of-the-box optimization for Codex. V4-Pro has simultaneously been listed in the “Expert Mode” of the App and Web, as well as via API. DeepSeek’s previously strong attention in the developer community has led this iteration to be widely seen as a forceful response to the new-generation models from OpenAI and Anthropic. In particular, the balance between reasoning cost and Agent stability is worth ongoing observation.
Original link: https://x.com/deepseek_ai/status/2087864585504305397
Gemini 3.7 Flash released, targeting coding and agent development
Google DeepMind announced the launch of Gemini 3.7 Flash, positioned as an enhanced upgrade for coding, knowledge work, and web development scenarios. Although the official full capability baseline has not yet been published, based on the consistent positioning of this series, 3.7 Flash continues to focus on low latency and cost-effectiveness as its core selling points, aimed at high-frequency calling scenarios. The industry generally believes that with GPT-5.x and DeepSeek V4 making simultaneous progress, the lightweight model track has entered a difficult stage of “must run fast and answer accurately.” Developers should pay close attention to its code generation coherence and multi-step tool-calling performance in real projects.
Original link: https://x.com/GoogleDeepMind/status/2087948366294515977
Meta open-sources Muse Glimmer, a 30B multimodal model focused on local Agent scenarios
Meta has again entered the open-source multimodal arena, releasing Muse Glimmer, a 30B-parameter model designed for localized, agentic use cases. Hugging Face has already provided Day-0 support, and community developers quickly open-sourced multiple interactive demos. The most eye-catching attempt was equipping Glimmer with tools so it could complete its own quantization compression on itself. This kind of “model self-optimization” intuitively demonstrates instruction-following ability while also indirectly confirming Meta’s ambition for on-device and private deployment scenarios. Compared with enormous instruction models, the 30B scale combined with quantization technology could become a low-cost choice for small and medium teams to implement multimodal Agents.
Original link: https://x.com/ben_burtenshaw/status/2086758382241763824
Running DeepSeek-R1 in the browser from scratch: a full hands-on guide with WebGPU + Transformer.js
A hardcore tutorial appeared on the Juejin hot list. The author explains in detail how to fully run the DeepSeek-R1 model locally in the browser using WebGPU and Transformer.js. The article examines WebGPU operator adaptation, memory management, and the Transformer.js inference pipeline at the source-code level, and provides a practical path that can be used directly. Technical solutions like this for “running large models in the frontend” have significant value for privacy-sensitive scenarios such as local document analysis and offline assistants, and further validate the feasibility of the browser as an AI inference container.
Original link: https://juejin.cn/post/7671475529255321610
Meta releases Muse Glimmer open-source model, optimized for local Agent workflows
Meta officially launched Muse Glimmer, an open-weight model focused on local, always-on Agent workflows. It has 30 billion parameters and outperforms same-size competitors on key Agent tasks and benchmarks, with a design goal of running entirely on consumer-grade hardware such as a Mac or a PC with a high-performance GPU. The model weights are released under the permissive Apache 2.0 license, continuing Meta’s tradition of open foundational AI research. At a time when cloud inference costs and privacy concerns are rising, local Agent models represent an important direction for lowering deployment barriers.
Original link: https://x.com/AIatMeta/status/2086757844544811485
Claude did not solve the Riemann Hypothesis, but improved the zero lower bound from 41.6% to 67.2%
Anthropic revealed that its unreleased research version of Claude attempted to attack the Riemann Hypothesis. Although it did not produce a final proof, it achieved considerable progress on a related problem: raising the lower bound for the proportion of zeros of the Riemann zeta function satisfying the hypothesis from 41.6% to 67.2%. This is not the finish line, but it demonstrates AI’s potential in deep mathematical reasoning—not only verifying existing conclusions, but also producing valuable intermediate results in directions humans have not yet explored. Attempts like this indicate that AI will increasingly enter frontier science as a “mathematical research partner” rather than a simple calculation tool.
Original link: https://x.com/AnthropicAI/status/2086867246073401655
Alibaba open-sources its first Max-level model: Qwen3.8-2.4T-A95B, natively supporting 256K context
Alibaba’s Qwen team officially opened the weights of Qwen3.8-2.4T-A95B on the ModelScope community, marking the first time a Qwen-Max-level model has been open-sourced. The model has a total of 2.4T parameters, activates 95B parameters per Token, and natively supports 262,144 Token context, expandable to 1,010,000. Qwen3.8 continues the hybrid architecture, with a focus on improving end-to-end capabilities in programming, office work, scientific research, and long-horizon Agent tasks. Official evaluations show it achieves high scores on benchmarks such as PaperBench and IFBench, trading blows with models like Opus 4.8, Fable 5, and GPT 5.6 Sol. For research institutions and enterprises that want to deploy a very large-scale MoE model in private environments, this is an extremely important option.
Original link: https://www.ithome.com/0/989/001.htm
Robot foundation model DYNA-2 released: trained on 1 million hours of human video, with task success rates up to 90%
Embodied intelligence company Dyna Robotics released the robot foundation model DYNA-2. Its pretraining is based entirely on more than 1 million hours of first-person human video, equivalent to about 170 years of human experience. The model is defined as a “world action model.” By simultaneously predicting the next video frame and the action that should be performed, it learns spatial relationships, movement patterns, and physical responses after contact with objects. Compared with relying on manually teleoperated data, this “learning by observing humans” strategy significantly reduces data collection costs. Official data shows its task success rate can reach up to 90%.
Original link: https://www.ithome.com/0/989/200.htm
OpenAI previews GPT-5.6 Sol ultrafast mode, with API access rolling out first
OpenAI teased the Ultrafast mode of GPT-5.6 Sol on the X platform, claiming inference speeds up to 14 times faster than before. The mode will first be available to selected API customers and gradually expand to more enterprise users as compute capacity grows. This marks a shift in large-model competition from parameter scale to response speed and service density. It is especially significant for real-time interactive scenarios and Agent workflows—faster inference means more complex multi-step tasks can be completed within acceptable user latency. For developers building applications on APIs, this may be one of the most noteworthy capability upgrades in the second half of the year.
Original link: https://x.com/OpenAI/status/2087947721936359705
Google releases Gemini 3.7 Flash, with significantly enhanced coding and web development capabilities
Google DeepMind officially launched Gemini 3.7 Flash, which shows clear improvements over 3.6 Flash on key coding tasks such as debugging and problem-solving. It can also design more functional webpage layouts and applications with fewer prompts, and demonstrates better reasoning and accuracy in real business workflows. The model is now available via API in Antigravity, Google AI Studio, and Android Studio. Google AI Pro and Ultra subscribers can use it through the Spark interface in the Gemini app. This iteration focuses on developers’ everyday pain points and aims to create more direct competitive pressure on GitHub Copilot and Claude in the AI coding assistant space.
Original link: https://x.com/GoogleDeepMind/status/2087948368957894859
Foreign media report: Hassabis wanted to leave Alphabet, eventually stayed by becoming Chairman
According to Pathfounders, Demis Hassabis, the 2024 Nobel Prize winner in Chemistry and former CEO of Google DeepMind, had intended to leave the company before Alphabet’s recent senior management reshuffle. Sources say Hassabis originally wanted to leave at the same time as former Alphabet Chief Scientist Jeff Dean, but Alphabet believed that losing two AI business leaders in a short period would cause its stock price to collapse. The company eventually persuaded him to transition to Chairman of DeepMind and Chief Scientist of Alphabet. This arrangement allows Hassabis to devote more energy to AI science exploration projects such as Isomorphic Labs, while also paving a smooth transition path for a possible future departure. Jeff Dean has already left with several colleagues to found Discovery Loop. This personnel adjustment reflects the subtle power struggle between top AI talent and tech giants.
Original link: https://www.ithome.com/0/987/775.htm
Meta demonstrates Muse Glimmer: completing an entire agentic task with a single prompt
Meta demonstrated a new capability of its Muse Glimmer model—completing a multi-step Agent task end-to-end from a single natural language prompt. In the demo, the model autonomously discovered a local Home Assistant instance through web tool calls, queried device APIs, wrote a responsive HTML/CSS/JS dashboard from scratch, and deployed a local server for verification. This means large models are moving from “conversational generation” to “autonomous execution of complex workflows,” with demonstrative significance for scenarios such as smart home control and automated operations. Although this is still at the demo stage, the end-to-end autonomy shows a viable form for Agent products.
Original link: https://x.com/AIatMeta/status/2086757849217257760
Practical sharing: how to use LLMs to learn complex knowledge efficiently
A practical article from the Hacker News hot list resonated widely (505 points, 286 comments). The author systematically lays out a methodology for using LLMs as learning partners: not simply asking questions and seeking answers, but using Socratic questioning, conceptual analogies, and mistake diagnosis to help the model build knowledge frameworks and fill blind spots. The article emphasizes that LLMs are best suited to act as “mentors who never get tired,” helping learners actively construct understanding rather than passively receive information. In an age of information overload, this approach provides knowledge workers with a reusable learning paradigm.
Original link: https://laurentiugabriel.github.io/blog/articles/how-i-use-llms-to-learn/
Sign language translation enters the consumer market for the first time: Google DeepMind launches a sign language-to-text model
Google DeepMind announced on August 12 the release of a large-scale multilingual sign language-to-text (SL2T) model, marking the first time sign language AI has truly reached consumer products. Gboard and Live Transcribe on the Pixel 11 series now support sign language-to-text input, initially for American Sign Language, with expansion to more sign languages and devices planned. Approximately 70 million hearing-impaired people worldwide use more than 200 sign languages. The grammar and vocabulary of sign languages are entirely different from spoken languages, which means SL2T requires genuine machine translation rather than a simple gesture-to-text sequence mapping. DeepMind trained the model on more than 50 sign languages and 100,000 hours of data, and protects user privacy by converting sign language into pose keypoints, so raw video streams never need to be uploaded to servers. The breakthrough of this product is that sign language communication has finally obtained a “typing alternative” similar to voice input, and tests show that using American Sign Language is faster and more natural than typing English. However, Google has not yet open-sourced the model or provided complete technical details, so its multilingual expansion capability still needs to be verified.
Original link: https://www.ithome.com/0/989/105.htm
AI Agents enter materials science: Discovered Materials finds new materials for the semiconductor industry
YC-incubated startup Discovered Materials has launched an AI Agent for new material discovery. Co-founders Advaith and Akash introduced the project on HN, targeting a core pain point in the semiconductor industry: runaway power consumption. GPU heat dissipation is deteriorating at an alarming rate—Nvidia H100 (2022) has a TDP of 700W, Blackwell (2024) reaches 1.2kW, and Rubin (2026) is expected to hit 2.3kW. Traditional materials R&D cycles are long and struggle to keep pace with chip iteration speed. The company uses AI Agents to automate material screening and discovery processes, attempting to find a solution to heat dissipation at the new materials level. This is a direct application of AI for Science in hard tech, and the direction has high industrial value. However, no specific material discovery results have been disclosed yet, and its scientific validity still needs to be verified through papers or actual products.
Original link: https://discoveredmaterials.com/research/
Today’s focus: AI is rapidly evolving from a “general conversational assistant” into a “vertical-domain doer”—from materials discovery and sign language translation to intelligent diabetes diagnosis and treatment. The real breakthrough lies in whether models can solve specific problems under complex real-world constraints. At the same time, security offense and defense, along with accelerated scientific discovery, are continuously setting new boundaries and possibilities for this process.