Qwen Releases Qwen3.8-LiveTranslate Real-Time Translation Model, LAAL Drops to 2.3 Seconds
Qwen releases Qwen3.8-LiveTranslate with Interleave architecture and Hybrid-MoE Thinker–Talker design, reducing average latency (LAAL) from 2.8s to 2.3s.
Model Releases & Updates
Qwen Releases Qwen3.8-LiveTranslate Real-Time Translation Model, LAAL Drops to 2.3 Seconds
Qwen releases Qwen3.8-LiveTranslate with Interleave architecture and Hybrid-MoE Thinker–Talker design, reducing average latency (LAAL) from 2.8s to 2.3s.
Industry News
Google Discloses Gemini Autonomously Hacked Three Real Companies in Security Tests Before Self-Terminating
Google confirms that Gemini autonomously hacked three real companies during an Irregular "capture the flag" exercise in May, due to accidental open internet access in the test environment — the first known AI jailbreak incident.
- Source: IT Home (RSS)
- AIHOT: https://aihot.news/items/cmu7l5wy60hj3rogr5i76b082
New York Times and Other Media File Summary Judgment Motion, Citing OpenAI and Microsoft Executives' Internal Comments
The New York Times, Daily News Group, and Ziff Davis filed a 92-page summary judgment motion against OpenAI and Microsoft for billions in AI training copyright damages, citing previously undisclosed internal emails and sworn testimony.
NYT Copyright Lawsuit Reveals: Microsoft Executive Internally Called Training AI the Largest Theft of Labor in Human History
In its motion for summary judgment, the NYT revealed internal materials from Microsoft and OpenAI. Microsoft Applied Sciences Director Brent Hecht's 2023 memo called LLMs consuming labor "the largest theft in human history"; Microsoft data shows Copilot reduced NYT clicks by up to 93% vs Bing.
- Source: IT Home (RSS)
- AIHOT: https://aihot.news/items/cmu7l5wy60hj2rogr88fr2vkk
Anthropic and Accenture Partner for Embedded Independent Evaluation, Each Committing at Least $10 Billion
Anthropic announced a partnership with Accenture for embedded independent evaluation of frontier models, led by Accenture's AI business Faculty, including model evaluation, red teaming, alignment assessment, and safety testing. Both companies expect to invest at least $10 billion each over five years.
Tips & Perspectives
Reverse Engineering Reveals ZCode Silently Packages Git History and Uploads to Aliyun OSS
Developer ferstar reverse-engineered Z.ai's AI coding desktop app ZCode, discovering that after login it silently packages the entire workspace (including complete .git history, LFS cache, reflogs, and global config) and encrypts it to Aliyun OSS. One snapshot contained 42,411 files at 313MB, with .git directory accounting for 86.6% of the payload.
3-Person Team Hacked OpenAI Employee Accounts with Frontier Models for Under $3,000 in Token Costs
A 3-person team exploited two vulnerabilities to take over OpenAI employees' ChatGPT/Codex accounts on July 25, gaining access to Outlook, Slack, GitHub, and other linked services. They proved the vulnerability by submitting a PR to OpenAI's internal codebase — all within 72 hours.
OpenRouter Benchmarks 20 Image Generation Models on Cost, Editing, and Quality
OpenRouter tested 20 image generation models with identical prompts. Per-image costs ranged from $0.006 to $0.134 — a 22x difference.
Trail of Bits Built Custom LSP, Decompiler, and Lean Formal Verification for Miden zkVM Audit
For auditing Miden zkVM, Trail of Bits had Agents spend six months building from scratch: an MASM LSP server, decompiler, static analysis engine, and Lean VM executor model. These tools found critical vulnerabilities allowing malicious provers to forge Falcon signatures and steal funds, identified 400+ type validation defects, produced 95 machine-verified correctness proofs, and caught 2 subtle bugs missed by unit tests.
- Source: Trail of Bits: AI Security Research
- AIHOT: https://aihot.news/items/cmu6w30lt0dnhrowkh7qwiped
TypeSafe AI Releases Jev: LLM for High-Frequency Decisions Only
TypeSafe AI launched Jev, a model focused solely on high-frequency decisions — no conversation or text generation, only judgment output. It's 20-200x faster than traditional LLMs at $0.042/million tokens with free output tokens. Author benchmarks show Jev ranked second in accuracy and cheapest for pre-screening tasks, achieving highest accuracy and fastest speed on parallel judgment tasks. Uses RLCD training for decision calibration. Available at https://typesafe.ai/ with Vercel already integrated.
- Source: 公众号：数字生命卡兹克
- AIHOT: https://aihot.news/items/cmu67pecj0e0wrofj97vtuwzu
Gary Marcus: Trump Downplaying AI Risks for Economic Reasons, AI Hallucinated Intelligence Report Nearly Started a War
Gary Marcus cited NYT reporting that Trump is downplaying AI panic and resisting regulation for economic reasons. He shared Katie Bo Lillis's report that an AI-assisted intelligence report falsely concluded a Chinese ship carried nuclear weapons components due to model hallucination, causing the US military to scramble — reportedly "nearly starting a war."
- Source: Gary Marcus: The Road to AI We Can Trust (RSS)
- AIHOT: https://aihot.news/items/cmu7e0dly06xdrogro9rq1x6d
Gary Marcus: The Real Threat Isn't Rogue Superintelligence, It's Scalable Hacker Attacks from Agentic AI
Gary Marcus argues that the real concern isn't out-of-control superintelligence, but the unleashed agentic AI causing internet-scale hacking attacks.
- Source: Gary Marcus: The Road to AI We Can Trust (RSS)
- AIHOT: https://aihot.news/items/cmu77kvz00tf6rowk9kl5o4k0
Ethan Mollick on Capability Overhang: GPT-6 Astra and Fable 5.1 Capabilities Remain Largely Untapped
Ethan Mollick argues that GPT-6 Astra and Fable 5.1 can already reliably complete weeks worth of human work, but most people are far from exhausting their capabilities, creating a "capability overhang."
- Source: Ethan Mollick: One Useful Thing (RSS)
- AIHOT: https://aihot.news/items/cmu79r5lt0z09rowk1o1z6jnt
Originally published on WeChat Official Account 「比特财商」.
About the Author
ERIC
AI Technology Expert, focusing on research and application of artificial intelligence and automation tools
Contact & Platforms