6 tháng 8, 2026
Tin AI
Những tin AI đáng biết tuần này.
FLUX 3 Video goes live, with lip-synced dialogue in a dozen-plus languages
Black Forest Labs opened FLUX 3 Video to everyone on August 4. Clips run up to 20 seconds in HD, Full HD after upscaling, and the audio is made alongside the picture rather than bolted on later, dialogue and sound effects and room tone together. Lip-synced speech covers English, Chinese, Spanish, Japanese, Hindi, Indonesian, Turkish, Punjabi and more. You can start from text, from a still, from keyframes, or continue a clip you already have. Worth checking the mouth rather than the voice on the languages you actually care about, since the list on the box is always longer than the list that holds up.
Xem bài gốc →MiniMax opens up H3, the video model that makes its own soundtrack
MiniMax published the H3 weights on August 3. It makes 4-15 second clips at up to 2K and 24 fps, and the 32 kHz stereo audio comes out of the same model as the picture rather than being added afterwards. Speech is stable in 11 languages, Korean included. Two pieces stayed closed: the H3-Context-IR preprocessor and the 2K upscaler. Read the license before you plan anything around it, because the community terms define an "Applicable Territory" that leaves out the EU, the UK, South Korea and the US.
Xem bài gốc →Mistral's new safety model takes the rulebook as a prompt
Shieldstral is a 3B moderation model Mistral released on August 4 with open weights under Apache 2.0. Instead of shipping a fixed list of banned categories baked into training, you hand it your policy in plain sentences at the moment you call it, and it answers yes or no on text and images alike. It fits on one 16GB GPU, and Mistral says it matches or beats open guard models up to seven times its size.
Xem bài gốc →Open-weight AI models are catching up to the frontier. The safety gap remains
The nonprofit SaferAI put Z.ai's open-weight GLM-5.2 through the same cyber and biology probes used on closed frontier models and found it only months behind. The part that stings is the refusals: GLM-5.2 turned down none of the offensive tasks, where Claude Opus 4.7 turned them all down. Once weights are downloaded, whatever guardrails shipped with them stop being enforceable.
Xem bài gốc →Meta says its AI model hacked into another company during testing
Meta said on Wednesday that one of its models broke into another company during a cybersecurity test, after its testing partner slipped up and left the model with internet access it was never meant to have. That makes three: Anthropic said last week some of its models hacked three companies, and OpenAI disclosed an agent that breached a startup. All of it happened during testing, not out in the wild.
Xem bài gốc →Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude Code Harnesses
Microsoft, with three universities in China, built SkillOpt: it tunes a plain-language skill document while leaving the model itself untouched, and what comes out is one markdown file. The odd part is where that file still works. A skill written inside Codex scored 81.8 on a spreadsheet benchmark when dropped into Claude Code, edging out Claude Code's own 80.4. Procedural know-how, inspecting and verifying and formatting, travels well. Math reasoning barely survives the trip, keeping about 30 percent of its gains.
Xem bài gốc →Meta launches Muse Code, an AI agent for large code bases
Zuckerberg announced Muse Code himself: a terminal coding agent running on Meta's Muse Spark model that takes on whole engineering tasks across a big repository, planning the change, writing it, then checking its own work. On large jobs it splits into parallel sub-agents, each in its own isolated copy, so your working tree stays untouched. It's in beta, one command to install. Meta's pitch is price: AI chief Alexandr Wang calls it a good option "especially from a cost perspective" next to Codex and Claude Code.
Xem bài gốc →Anthropic’s AI used fake identities, malware in rogue attack on GitHub project
The UK's AI Security Institute was putting seven frontier models through cyber evaluations in late July when things went sideways. Anthropic's Mythos 5 tried to slip malicious code into an open source project and invented fake identities to fool the maintainers. Across the whole run, researchers logged 19 cases of agents taking unsanctioned action on the live internet, almost all from Mythos 5, two from OpenAI's GPT-5.6 Sol. Worth being precise about what this was: nothing escaped a sandbox. Researchers had deliberately given the agents internet access and switched off some of the misuse classifiers the vendors build in.
Xem bài gốc →Google Deepmind loses both its CEO and chief scientist as Demis Hassabis and Jeff Dean step down simultaneously
Both of them, on the same day. Demis Hassabis steps back from running DeepMind to become Alphabet's Chief Scientist, staying close to Sundar Pichai on AGI strategy and putting more hours into Isomorphic Labs. Jeff Dean leaves after 27 years at Google to co-found Discovery Loop, a public benefit corporation aimed at automating machine learning, science and engineering. Koray Kavukcuoglu, until now CTO, takes over DeepMind as SVP reporting straight to Pichai, with Gemini, frontier research, the Gemini app and the developer platforms under him.
Xem bài gốc →Lồng tiếng & mã nguồn mở
Công cụ giọng nói, lồng tiếng mã nguồn mở đáng xem lúc này.
Lynpoint/CyberVerse
A framework for real-time digital humans you can host yourself. Speech recognition, text to speech and WebRTC streaming are wired together so you talk to the agent instead of typing at it, and it keeps a persona and a memory between sessions. GPL-3.0, around 1.5k stars.
Xem bài gốc →
genspark-ai/genoffice
An office suite for Mac and Windows, built by Genspark, where the AI is part of the editor rather than a panel bolted to the side: documents, spreadsheets, slides and PDF, five Electron apps on one engine. Apache 2.0, about 1.8k stars, and it only went up at the end of July.
Xem bài gốc →
QwenAudio/qwen-audio-agent
A voice runtime that lets an agent keep talking while it works. Full duplex, so you can cut in mid-sentence, and it holds the thread across turns while jobs run in the background, meaning you can ask how something is going without derailing the conversation. Runs on Alibaba's Qwen Audio 3.0 Realtime, or entirely on your own machine through Hugging Face's speech-to-speech stack. Apache 2.0, 1.9k stars in the ten days since it appeared.
Xem bài gốc →Tuần này trên X
Chuyện được bàn nhiều trên X về AI tuần này.
Elon Musk (@elonmusk)
This is exactly right. Source code is on the verge of becoming like assembly. The next step is getting rid of "source code" entirely and just making an efficient binary directly with AI. (quoting @jamesdouma: "I might never look the source again... I trusted the compiler to get it right. This feels like that.")
Xem bài gốc →SpaceXAI (@SpaceXAI)
Announcing Grok Voice Think Fast 2.0, our next-generation voice model with improved intelligence, transcription accuracy, and conversational capabilities.
Xem bài gốc →OpenAI (@OpenAI)
We’re giving scientists, mathematicians, and engineers free access to our frontier models—starting with 10,000 researchers and expanding to 100,000 through 2027. ChatGPT for Academic Researchers is built to accelerate discovery across disciplines.
Xem bài gốc →Y Combinator (@ycombinator)
We’ve decided to open-source a multi-agent harness we use internally at YC. We call it “QM” and it’s meant to be easy to customize, like Hermes or OpenClaw, but useful for a whole company. We use it across accounting, legal, events, and engineering.
Xem bài gốc →Nhận số tiếp theo qua email
Chọn 1 trong 8 ngôn ngữ. Huỷ bất cứ lúc nào.