OpenAI Solves Ten Decade-Old Math Problems with Internal Astra Model
OpenAI 内部 Astra 模型攻克十项十年未解数学难题 ⭐️ 9.0/10

OpenAI announced that an internal version of its next major model, Astra, solved ten long-standing open problems in mathematics and theoretical computer science, at a cost of under $2,000 per problem at GPT-5.6 Sol token prices. The results were released with Lean 4 formalizations, a paper, and an LLM-generated reasoning walkthrough. This marks a major milestone in AI-driven theoretical research: an AI system autonomously produced verifiable proofs of problems untouched for a decade, at a fraction of traditional research cost. It follows Anthropic's Claude-based cryptographic findings and is likely to intensify debates about AI's role in mathematics and the future of scientific discovery. OpenAI published the openai/ten-proofs repository with Lean 4 formalizations, plus a paper and an LLM-generated PDF reconstructing the reasoning traces. The blog post notes that OpenAI did not disclose how many problems were attempted without success, and that the model is an 'internal version of Astra,' not a publicly released system.

rss · Simon Willison · Aug 1, 20:34

Background: The news references a recent trend of frontier AI models being used for high-level research: days earlier, Anthropic's Mythos Preview discovered cryptographic weaknesses for about $100,000 in tokens. Mathematician Terence Tao has described such AI as a catalyst for 'big mathematics' — large-scale decentralized human-machine collaboration where AI handles technical grunt work. Formal proof assistants like Lean 4 allow AI-generated proofs to be verified by machine, increasing confidence in results.

Tags: #AI research, #mathematics, #OpenAI, #scientific discovery, #LLM capabilities

YC Open-Sources Internal Multi-Agent Harness QM
YC 开源内部多智能体框架 QM ⭐️ 9.0/10

Y Combinator has open-sourced QM, the multi-agent harness it uses internally, under an MIT license on GitHub. QM has been running real business workflows at YC — accounting, legal, event operations, and engineering — with its tool registry growing from about 20 to more than 350 tools. This matters because YC is sharing a production-validated reference architecture for multi-agent systems, not just a demo. Its scope-based abstraction — where memory, permissions, and tooling are bound to people or spaces rather than sessions — offers a concrete answer to how context and permission boundaries should be drawn in the agent era, which could influence how teams build collaborative AI systems. The architecture is deliberately minimal and split into four layers: a headless core in TypeScript on Node.js with Fastify, swappable harnesses/models (Pi, OpenCode, Codex, Claude Code), Postgres-backed persistence with per-scope sandboxes and a small execute tool, and plugin-based frontends like a Vite+Lit web UI and Slack via Bolt. Security is offered in three tiers — Strict, Auto (a default classifier that screens external data before it reaches the model), and Dangerous — and narrower scopes can only tighten a chosen security level, never loosen it.

rss · meng shao(@shao__meng) · Aug 1, 02:40

Background: An agent harness is the infrastructure that wraps an AI model to manage its lifecycle, context, tool access, verification, and safety in production — the model generates text, while the harness decides what the model sees, what it can do, and when to stop. QM takes this idea further as a multi-agent harness for a whole company, using scopes (per person, Slack channel, or project room) to give each space its own memory, files, permissions, scheduled jobs, and sandbox. That design enables both personal isolation and collaborative sharing, and lets the same identity and configuration carry seamlessly between Slack and Web. YC says the project is meant to be easy to customize, like Hermes or OpenClaw, but useful for an entire organization.

Tags: #multi-agent, #open-source, #agent-harness, #YC, #scope-abstraction

Ripgrep musl builds segfault on very large searches
ripgrep 的 musl 版本在超大搜索时偶尔段错误 ⭐️ 8.0/10

A bug report filed on the ripgrep GitHub repository (issue #3494) documents that musl builds of ripgrep can sporadically segfault when performing very large searches. The discussion quickly expanded into an analysis of musl's mallocng allocator and its behavior under multithreaded contention. Ripgrep is one of the most widely used command-line search tools, and musl-based static binaries are a popular choice for minimal containers and embedded Linux systems. A segfault in this configuration undermines reliability in exactly the environments that favor musl, and the ensuing discussion highlights allocator and kernel-level issues that can affect many multithreaded programs. The crash appears to stem from musl's default mallocng allocator, which has known weaknesses under multithreaded lock contention, rather than from ripgrep's core search logic. The thread also linked to an AI-generated analysis of the bug and referenced a kernel patch, showing how a user-space crash can prompt kernel-level scrutiny.

hackernews · throwaway2037 · Aug 1, 12:34 · Discussion

Background: musl is a lightweight C standard library for Linux, commonly used to build static binaries for minimal environments; it aims for efficiency and standards conformance. Like most libc implementations, musl ships its own memory allocator — mallocng — which manages dynamic memory for programs. Very large, multithreaded searches create heavy allocation/deallocation pressure, and allocators can behave differently under contention, leading to crashes in some cases. General-purpose allocators make trade-offs between speed, fragmentation, and scalability, which is why a libc's default allocator can become a bottleneck or a source of intermittent bugs.

Discussion: Commenters generally treated the bug as an allocator problem rather than a ripgrep logic bug: several pointed out that musl's mallocng handles multithreaded contention poorly, while others shared a kernel patch reference and an AI-generated analysis. One commenter warned that running ripgrep on HPC cluster filesystems with many small I/Os can hammer metadata mechanisms, but the core thread remained focused on why the crash manifests only with musl.

Tags: #ripgrep, #musl, #segfault, #allocator, #bug

NetBSD 11.0 released with RISC-V port and fast-boot microVM
NetBSD 11.0 发布，新增 RISC-V 移植与快速启动 microVM ⭐️ 8.0/10

The NetBSD project has released version 11.0, the first release to include a RISC-V port. It also introduces a new MICROVM kernel for x86 that can boot in about 10 milliseconds, alongside improvements to the npf firewall and expanded legacy hardware support. This release strengthens NetBSD's position as a highly portable operating system, appealing to users in embedded, cloud, and vintage computing communities. The fast-boot microVM capability could enable new use cases in edge computing and serverless architectures, while the RISC-V port aligns with the growing open hardware movement. The release notes highlight the new MICROVM kernel for x86, which boots in about 10 ms, as well as improvements to the npf(7) firewall including layer 2 and user/group filtering. NetBSD 11.0 also continues its tradition of supporting vintage and misc. hardware such as alpha, hp300, m68k, mac68k, and others.

hackernews · jaypatelani · Aug 1, 17:56 · Discussion

Background: NetBSD is a free, open-source Unix-like operating system known for its exceptional portability, running on a wide range of hardware from large servers to embedded devices. RISC-V is an open instruction set architecture (ISA) based on reduced instruction set computing (RISC) principles, developed collaboratively and free to use. A microVM is a lightweight virtual machine that combines the security and isolation of traditional VMs with the resource efficiency of containers, making it ideal for high-density cloud workloads.

Discussion: Community members expressed positive sentiment, with one noting that NetBSD's continued support for vintage hardware contrasts with Linux dropping such support, making NetBSD the go-to OS for older machines. Others congratulated the RISC-V port and praised the microVM's 10 ms boot time and the firewall improvements as valuable features. Some also raised broader questions about the current status of BSD systems compared to Linux.

Tags: #NetBSD, #BSD, #RISC-V, #Operating Systems, #Open Source

Canada Signs UN Cybercrime Convention Amid Surveillance Concerns
加拿大签署联合国网络犯罪公约引发监控担忧 ⭐️ 8.0/10

In July 2026, Canada quietly signed the United Nations Convention against Cybercrime, with ministers touting its child-protection provisions and human rights safeguards. Critics, including privacy expert Michael Geist, argue the treaty is a surveillance regime in disguise. This signing commits Canada to an international cybercrime framework that could facilitate cross-border evidence collection and surveillance, raising serious privacy and human rights concerns. It also signals a shift in Canadian digital policy that may affect how domestic laws are enforced and how citizens' data is shared abroad. The treaty, formally the Convention on Cybercrime: Strengthening International Cooperation to Combat Crimes Committed Through ICT Systems, was adopted by the UN General Assembly in December 2024 after being proposed by Russia in 2017. Canada signed without public consultation, and as a signatory rather than a ratifying party, its legal obligations remain limited for now.

hackernews · iamnothere · Aug 1, 14:19 · Discussion

Background: The UN Cybercrime Convention is the first international criminal justice treaty addressing crimes committed through ICT systems, designed to strengthen international cooperation in enforcement. It has been controversial since its inception: Russia proposed it in 2017, and human rights advocates and some governments resisted it, fearing it could legitimize surveillance, weaken privacy protections, and let authoritarian states use mutual legal assistance to target dissidents. The convention, also known as the Hanoi Convention, opened for signature in 2025, and as of May 2026 dozens of countries have signed it.

Discussion: Commenters largely praised Michael Geist's long-standing investigative work on privacy issues and echoed his skepticism, with one calling the move 'WYSIWYG politics' where governments signal to different audiences. Others pointed out that Canada signs most UN treaties and that signing alone has limited impact until ratification, so some questioned how much real change this represents.

Tags: #privacy, #surveillance, #cybercrime, #UN treaty, #Canada

DeepSeek Launches V4-Flash with Native Codex Support, Beats V4-Pro in Agent Tests
DeepSeek 发布 V4-Flash 正式版：原生接入 Codex，全面超越 V4-Pro ⭐️ 8.0/10

DeepSeek released DeepSeek-V4-Flash-0731, the official version of its V4-Flash model, with native integration for OpenAI Codex including a pre-configured system prompt. The model beats DeepSeek's larger V4-Pro preview across all nine agentic evals, scoring 82.7% on Terminal Bench, and is priced at $0.28 per 1M tokens with a 1M context window, making output about 89x cheaper than Anthropic's Opus 4.8. This release shows DeepSeek aggressively targeting agentic coding workloads and price-performance, positioning a smaller open-weight model as a top-three open-source model and a cheap drop-in for Codex. Native Codex support plus a one-line setup lowers adoption friction, intensifying pricing pressure on commercial models like Opus 4.8. The model is available day-zero on Fireworks, and DeepSeek says it offers better cost-per-task than V4-Pro. The official Codex adapter includes Codex's own system prompt, so developers can switch with a single command without changing existing settings.

rss · 小互(@imxiaohu) · Aug 1, 02:46

Background: OpenAI Codex is an AI system that turns natural-language instructions into working code and automates coding tasks, typically used as a coding agent. Agentic evals, such as Terminal Bench, measure how well an AI agent plans, uses tools, and completes real-world tasks. DeepSeek is a Chinese AI lab known for publishing open-weight models, while Anthropic's Claude Opus 4.8, released in May 2026, is a recent flagship commercial model strong in coding and agentic work.

Tags: #DeepSeek, #AI Model, #Open Source, #Codex, #LLM

MiniMax 3 Is First AI Video Model to Write 'Hi' on Chalkboard
MiniMax 3 成为首个通过粉笔字'Hi'测试的 AI 视频模型 ⭐️ 8.0/10

After testing the same Turing-test prompt on AI video models for over a year and a half, Justine Moore reports that MiniMax 3 (Hailuo AI) finally succeeded in generating a video of a man writing "Hi" in chalk on a blackboard. The model is also open source. This marks a notable breakthrough in video generation, specifically in the accurate rendering of legible handwritten text—a task that previously stumped every model tested. Open-sourcing the model also gives developers and researchers access to a state-of-the-art video generator, accelerating innovation in the field. The exact prompt used was "man writes 'hi' in chalk on blackboard," and prior models, including Veo 2, came close but still failed. The tweet includes a demonstration video showing the model successfully writing "Hi" on the chalkboard, indicating strong consistency between handwriting and hand movements.

rss · Justine Moore(@venturetwins) · Aug 1, 02:08

Background: AI video models generate visual scenes from text prompts, but accurately rendering legible written text inside the video is notoriously difficult because the models lack explicit text layout understanding. The term "Turing test" here is used metaphorically, borrowing from the classic test for machine intelligence to frame text writing as a benchmark for video models. MiniMax Group Inc. is a Shanghai-based AI company that develops the Hailuo AI video-generation service; it listed on the Hong Kong Stock Exchange in January 2026 and is considered one of China's "AI Tigers."

Tags: #AI video, #MiniMax, #open source, #Turing test, #text generation

Huawei Open-Sources 505B openPangu-2.0-Pro Trained on Ascend NPU
华为开源基于昇腾 NPU 训练的 505B 参数 openPangu-2.0-Pro 模型 ⭐️ 8.0/10

Huawei has open-sourced openPangu-2.0-Pro on Hugging Face, a 505B-parameter mixture-of-experts (MoE) model with a 512K context window. The company claims it is the first 500B+ frontier model trained entirely on non-NVIDIA hardware (Huawei Ascend NPU). This release matters because it demonstrates that large-scale model training is feasible without NVIDIA hardware, potentially diversifying the AI supply chain. However, its benchmark results trail other top open-source models, so it may not shift the competitive landscape immediately. The model uses a MoE architecture with 505B total parameters and 18B active parameters (A18B). On publicly overlapping benchmarks, it scored GPQA-Diamond 87.9 (vs Kimi K3 93.5 and GLM-5.2 91.2), BrowseComp 65.7 (vs K3 91.2), and SWE-bench Verified 68.5 (vs DeepSeek V4 self-reported 80.6). Notably, Huawei's reported benchmark dimensions barely overlap with DeepSeek's.

rss · meng shao(@shao__meng) · Aug 1, 02:23

Background: Mixture-of-Experts (MoE) is a neural network architecture that divides computation across multiple 'expert' subnetworks, activating only a subset per token, enabling larger models with lower compute cost. Huawei's Ascend NPU is a line of AI accelerators designed as an alternative to NVIDIA GPUs. openPangu-2.0-Pro is part of Huawei's Pangu model family, and its open-weight release on Hugging Face allows developers to download and fine-tune it, though its 505B size requires substantial inference resources.

Tags: #open source, #Huawei, #LLM, #Ascend NPU, #MoE

Opus 5 Autonomously Creates Playable Multiplayer Halo Clone in 12 Hours
Opus 5 连续运行 12 小时，从零做出可联机的《光环 1》仿制游戏 ⭐️ 8.0/10

Anthropic's Claude Opus 5 model ran continuously for 12 hours and autonomously produced a Halo 1 clone with full multiplayer functionality, including 5v5 team deathmatch on a self-hosted server, matchmaking, leaderboards, and in-game text chat. This demonstrates that frontier LLMs can now deliver complete, playable multiplayer systems rather than just code snippets or demos. It could accelerate game prototyping and lower barriers for indie development, while raising questions about copyright and the role of human programmers. The output includes server-side infrastructure plus client features typical of a full game service. No technical methodology (e.g., iteration history, token usage, or code size) was disclosed, only attached video evidence.

rss · AI Will(@FinanceYF5) · Aug 1, 23:42

Background: Claude Opus 5 is Anthropic's latest reasoning model, designed for complex multi-step problem solving; it supports a 1M-token context window and up to 128K tokens of output. Historically, AI-assisted game development focuses on automating isolated tasks (such as assets, music, and script snippets), while fully autonomous game creation has remained an open challenge. This report suggests a shift from task-level assistance to system-level generation, although broader research still describes end-to-end autonomous game creation as evolving rather than mature.

Tags: #AI, #game development, #multiplayer, #code generation, #LLM

AI Cracks Math Conjecture and Solves Problems, Mathematicians Divided
AI 破解数学猜想并解题，数学家看法分歧 ⭐️ 8.0/10

OpenAI's internal reasoning model refuted the 1946 Erdős unit distance conjecture, and GPT 5.6 Pro solved two longstanding math problems on its first attempts, according to Fields Medal winner Timothy Gowers. AI is beginning to produce genuine mathematical breakthroughs, shifting the role of human mathematicians from primary discoverers to verifiers. Gowers warns this could lead to the 'possible destruction of mathematical culture' if researchers stop developing the expertise needed to understand AI-generated proofs. The Erdős unit distance conjecture, posed by Paul Erdős in 1946, asks how many pairs at exact distance 1 can occur among n points in the plane; OpenAI's model refuted it with assistance from Gemini 3.1. Gowers, who reportedly spent considerable time on the two problems, saw GPT 5.6 Pro solve each on its first attempt, though some mathematicians view AI simply as a productivity tool.

rss · The Decoder · Aug 1, 16:01

Background: The Erdős unit distance conjecture is a classic open problem in combinatorial geometry. AI systems, particularly large language models like OpenAI's GPT series, have recently been applied to mathematical reasoning with increasing success. Timothy Gowers is a celebrated mathematician and Fields Medalist known for his work in combinatorics and analysis. His warning reflects a broader debate about how AI should be integrated into mathematical research.

Tags: #AI, #Mathematics, #Research, #OpenAI, #Machine Learning

Self-Spreading Worm in Word Docs Hijacks Microsoft Copilot
隐藏在 Word 文档中的自传播蠕虫劫持微软 Copilot ⭐️ 8.0/10

A security researcher demonstrated a worm-like attack on Microsoft Copilot for Word, using hidden prompt injections embedded in Word documents that automatically spread into new files whenever the documents are reused. Microsoft acknowledged the issue but failed to fix it after 144 days and two attempted patches. The attack shows how AI-powered office tools can be turned into vectors for self-propagating malware via indirect prompt injection, a serious AI supply-chain risk. Even though it has not been exploited in the wild, the 144-day unpatched window underscores the difficulty of securing AI-integrated software. The worm relies on invisible prompt injections that ride along inside document content, so each reuse of a document triggers propagation to new files. Microsoft confirmed the vulnerability but failed to deliver a working fix in two attempts over 144 days.

rss · The Decoder · Aug 1, 13:51

Background: Prompt injection is a type of exploit where carefully crafted inputs cause large language models to behave in unintended ways, often bypassing safeguards. Indirect prompt injection is especially dangerous because the malicious instructions are embedded in content the model retrieves, such as web pages or documents, rather than coming from the user directly. This attack extends that concept to create an AI worm that self-propagates through shared Word documents.

Tags: #AI security, #Prompt injection, #Malware, #Microsoft Copilot, #Security research

ByteDance Seedance 2.5 makes 30-second AI videos with built-in audio
字节跳动 Seedance 2.5 可生成带内置音频的 30 秒 AI 视频 ⭐️ 8.0/10

ByteDance released Seedance 2.5, an AI video model that generates a continuous 30-second clip with synchronized audio in a single pass. This triples the output length of Google's Gemini Omni Flash, and supports dozens of reference images, videos, and audio files. This advancement could significantly streamline video production workflows, especially for advertising teams, by removing the need to stitch together multiple short clips. It also raises the competitive bar for multimodal AI video generation, pressuring other major players like Google. Seedance 2.5 was previewed on June 23, 2026, at ByteDance's Volcano Engine FORCE conference in Beijing. The model is part of the Doubao (豆包) video generation lineup and supports up to 50 multimodal reference inputs, according to the search results.

rss · The Decoder · Aug 1, 13:33

Background: Most AI video models previously generated short clips of around 5 to 10 seconds, so users had to piece together many clips to create longer scenes. Seedance 2.5 aims to break this limitation by generating a full 30-second clip in one run, with native audio. Gemini Omni Flash, mentioned in the news, is a preview model from Google for conversational video generation and editing, which currently produces clips of about 10 seconds.

Tags: #AI video, #ByteDance, #Seedance, #text-to-video, #multimodal AI

German court rules Suno AI music generator violated copyrights, rejects fair use
德国法院裁定 Suno AI 音乐生成器侵犯版权，驳回合理使用辩护 ⭐️ 8.0/10

A Munich court ruled that AI music generator Suno violated copyrights through both training and output, finding six songs reproducibly stored in its models. The court rejected both Germany's text-and-data-mining exception and the US fair use defense. This is a landmark legal ruling on AI music generation and copyright, with potential to shape future AI training and output practices globally. It signals that courts may hold AI companies liable for reproducing copyrighted works even when alternative legal defenses are raised. The ruling is not final, and several key questions remain open regarding AI training data and output liability. The court explicitly rejected both the German text-and-data-mining exception and the US fair use defense in this case.

rss · The Decoder · Aug 1, 10:40

Background: Suno is an AI music generation platform that creates songs from text prompts, images, or videos. German law, based on an EU directive, provides a text-and-data-mining exception that permits certain uses of copyrighted works without consent, while US fair use allows limited unauthorized use under certain conditions. This case is among the first to examine how these exceptions apply to AI music training and output.

Tags: #AI, #copyright, #music, #legal, #intellectual property

Google Pulls Nano Banana 2 from Google Earth After Fake Satellite Images
谷歌因伪造卫星图像从 Google Earth 撤回 Nano Banana 2 ⭐️ 8.0/10

Google pulled its Nano Banana 2 image model from Google Earth just two days after launch, after users demonstrated how easily they could generate convincing fake satellite images—for example, showing a refugee column at the US-Mexico border with a simple prompt. This incident underscores the growing risk that generative AI tools can be weaponized to create misleading geospatial content, threatening trust in satellite imagery used for news, policy, and security decisions. It also highlights the need for robust safeguards and rapid response mechanisms when AI products are released publicly. Nano Banana 2 is Google's Flash-tier image generation model based on Gemini 3.1, featuring 4K output, multi-image composition, and world-knowledge grounded generation. The Google Earth integration let users pick any location and 'reimagine' it, but the ease of creating photorealistic forgeries of real places prompted the swift retraction.

rss · The Decoder · Aug 1, 09:00

Background: Nano Banana is Google's AI image generator that has been integrated into products like Pixel smartphones and Gemini assistants. Google Earth recently added an image-generation feature that uses Nano Banana 2 to let users alter satellite views of real locations, a capability with obvious potential for misuse. The rapid retraction shows the tension between enabling creative expression and preventing disinformation, especially for imagery that is typically assumed to be authentic geospatial data.

Tags: #AI ethics, #Google, #generative AI, #satellite imagery, #deepfakes

US Treasury Secretary Bessent's memo shows plan to buy $5–10B in yen
美财长贝森特备忘录显示拟购买 50 亿至 100 亿美元日元 ⭐️ 8.0/10

A Reuters photo of Treasury Secretary Bessent's memo at a Cabinet meeting showed a to-do item reading 'Buy $5–10 billion worth of yen.' This indicates the U.S. may have intervened in currency markets to support the yen, which would be the first such U.S. intervention since 2011. If confirmed, U.S. intervention to buy yen would mark the first time since 2011 that Washington has acted directly to support Japan's currency. It could signal a new phase of coordinated official action in currency markets and have major implications for global finance and the dollar-yen exchange rate. The memo was photographed at 11:33 a.m. ET during Trump's Cabinet meeting at Camp David. Earlier in the day, Reuters reported that the Treasury had informed several banks it might intervene that day, and Japanese authorities had already conducted yen-buying intervention in Tokyo, pushing the yen sharply higher.

telegram · zaihuapd · Aug 1, 05:52

Background: Currency intervention occurs when authorities directly buy or sell foreign exchange to influence the exchange rate. The last time the U.S. Treasury intervened in the yen was in 2011, following the Great East Japan Earthquake, when it coordinated with other G7 nations. The yen is a major reserve and safe-haven currency, and its value affects trade competitiveness and global financial conditions.

Tags: #forex intervention, #US Treasury, #Japanese yen, #financial markets, #currency

EA's $55 Billion Saudi-Led Takeover Closes Next Week
EA 550 亿美元卖身沙特财团，下周完成私有化 ⭐️ 8.0/10

Electronic Arts has received full regulatory approval for its $55 billion sale to a consortium led by Saudi Arabia's Public Investment Fund, with the deal set to close on August 4, 2026. EA will become a private company and will no longer publicly disclose its financial data. This is the second-largest acquisition in gaming history, trailing only Microsoft's $75.4 billion purchase of Activision Blizzard in 2023. The deal gives Saudi Arabia a major foothold in the global gaming industry and could reshape competitive dynamics among major publishers. The buyer group includes Saudi Arabia's Public Investment Fund (PIF), Silver Lake, and Affinity Partners, the investment firm founded by Jared Kushner. PIF has already fully acquired developers Scopely and Niantic and has steadily increased its stakes in several gaming companies.

telegram · zaihuapd · Aug 1, 09:10

Background: The Public Investment Fund is Saudi Arabia's sovereign wealth fund, with estimated assets of about $900 billion and chaired by Crown Prince Mohammed bin Salman. It has been aggressively investing in gaming as part of the country's Vision 2030 plan to diversify its economy, but the fund has also drawn criticism over transparency and human rights concerns.

Tags: #gaming, #acquisition, #Saudi PIF, #Electronic Arts, #industry news

Microsoft Confirms Plan to Launch Copilot 'Super App' This Year
微软确认今年推出 Copilot『超级应用』 ⭐️ 8.0/10

Satya Nadella confirmed on Microsoft's earnings call that the company will ship an AI 'super app' this year, combining Copilot's chat, coding, and agentic capabilities for consumers and businesses. He said the company will merge these experiences, including code features, into a single super app within the quarter. The move signals a strategic shift from standalone AI assistants toward a unified super-app delivery model, potentially reshaping how enterprise and consumer users access AI. It also intensifies competition with OpenAI's ChatGPT Work app and reinforces Microsoft's bet that agentic workflows are the next major AI battleground. According to a prior Fortune report, the super app will merge Copilot chat, GitHub Copilot coding tools, Copilot Cowork, and Autopilot agentic systems into one interface. Microsoft posted quarterly revenue of $90 billion, driven largely by AI and cloud, but no specific launch date or product name has been announced.

telegram · zaihuapd · Aug 1, 13:18

Background: Copilot is Microsoft's AI assistant embedded across Windows, Microsoft 365, and developer tools. Cowork is an agentic layer that lets Copilot take action—such as sending emails, scheduling meetings, and creating documents—rather than just generating text, while Autopilot refers to a newer class of autonomous agents that work with minimal or no prompting. Agentic AI systems pursue goals through their own actions, a concept gaining rapid traction in the industry. A 'super app' bundles many services into a single application, a model popularized by apps like WeChat.

Tags: #Microsoft, #Copilot, #AI, #Superapp, #OpenAI

ChangXin Memory Unveils DDR5 at 8,000Mbps, Enters High-End Tier
长鑫存储发布 8000Mbps DDR5 并进入高端梯队 ⭐️ 8.0/10

ChangXin Memory (CXMT) showcased its full DDR5 and LPDDR5X product lines for the first time at IC China, with DDR5 reaching up to 8,000Mbps — 25% faster than mainstream 6,400Mbps modules. It also introduced 24Gb DDR5 dies and LPDDR5X up to 10,667Mbps with 16Gb dies. This marks a significant milestone for China's semiconductor industry, as a domestic vendor reaches international high-end DDR5 performance levels. The 8,000Mbps speed aligns with leading global products, potentially increasing supply options for data-center and AI workloads amid export controls. The DDR5 line tops out at 8,000Mbps with 24Gb (3GB) dies aimed at data-center capacity expansion. LPDDR5X reaches 10,667Mbps with 16Gb dies, supporting 12GB to 32GB packaging solutions. The announcement came at the 22nd China International Semiconductor Expo (IC China).

telegram · zaihuapd · Aug 1, 15:30

Background: DDR5 SDRAM is the fifth generation of double-data-rate memory standardized by JEDEC, doubling bandwidth and reducing voltage to 1.1V compared to DDR4, while supporting up to 512GB per DIMM. LPDDR5X is a low-power variant for mobile devices, evolved from LPDDR5 to improve data rates, power efficiency, and density for smartphones and laptops. ChangXin Memory (CXMT) is China's leading DRAM maker, whose progress is closely watched as the country seeks to build a self-sufficient memory supply chain.

Tags: #半导体, #DDR5, #长鑫存储, #存储芯片, #IC China

📊 Run stats · Total 9m 34s · AI analysis 2m 17s · Tokens 0.49 MCY (input 0.30 / output 0.20 MCY)