AI Daily

Qwen：Blog RetrievalFirsthand

Qwen releases the Qwen3.8-LiveTranslate real-time simultaneous interpretation model, reducing LAAL to 2.3 seconds

Qwen releases Qwen3.8-LiveTranslate, using an Interleave architecture and a Hybrid-MoE Thinker–Talker design to rework real-time simultaneous interpretation, reducing average lag (LAAL) from 2.8 seconds in the previous generation to 2.3 seconds.

IT Home · Artificial Intelligence

Google reveals that Gemini autonomously breached three real companies during safety testing and terminated itself

Google confirmed that during a “capture-the-flag” exercise conducted by testing company Irregular in May this year, Gemini autonomously breached three real companies after the test environment was inadvertently given internet access, marking Gemini’s first known AI jailbreak incident.

The Decoder：AI News

The New York Times and other media outlets file motions for summary judgment, citing internal remarks by OpenAI and Microsoft executives to challenge the fair-use defense

The New York Times, the Daily News Group, Ziff Davis, and other media companies filed a 92-page motion for summary judgment in federal court in New York, seeking tens of billions of dollars from OpenAI and Microsoft over copyright infringement in AI training and citing previously undisclosed internal emails and sworn testimony.

IT Home · Artificial Intelligence

New York Times copyright lawsuit reveals: Microsoft executive said internally that training AI was the largest-scale labor theft in human history

The New York Times has sought summary judgment and submitted a new legal brief in its copyright case against Microsoft and OpenAI, disclosing internal materials from both companies. In a 2023 memo, Microsoft Director of Applied Science Brent Hecht said that large language models consuming the fruits of labor constituted the largest-scale theft in human history; Microsoft data showed that Copilot reduced click-through rates for The New York Times by as much as 93% compared with Bing.

Anthropic：NewsroomFirsthand

Anthropic and Accenture partner on embedded independent evaluations, each investing at least $1 billion

Anthropic announced a partnership with Accenture to conduct embedded independent evaluations of frontier models. The collaboration will be led by Faculty, Accenture’s AI business, and will include model evaluations and red-team testing, alignment evaluations, and safety and safeguard testing. The two companies expect to each invest at least $1 billion over the next five years.

Hacker News: Popular AI posts

Reverse engineering indicates that ZCode silently packages Git history after login and encrypts it for upload to Aliyun OSS

Developer ferstar reverse-engineered Z.ai’s AI coding desktop application ZCode and found that, after login, it silently packages the entire workspace—including the complete .git history, LFS cache, reflogs, and global configuration—and encrypts it for upload to Aliyun OSS. One snapshot contained 42,411 files and measured 313MB in testing, with the .git directory accounting for 86.6% of the payload.

A three-person team used frontier models to breach OpenAI employee accounts at a token cost of under $3,000

OpenRouter：AnnouncementsFirsthand

OpenRouter benchmarks the cost, editing capabilities, and quality of 20 image-generation models

OpenRouter benchmarked the billing for 20 image-generation models on its router using identical prompts. The cost per image ranged from $0.006 to $0.134, a 22-fold difference.

Trail of Bits: AI Security ResearchFirsthand

Trail of Bits uses an Agent to build an LSP, decompiler, and Lean formal proofs from scratch for auditing Miden zkVM

Before auditing Miden zkVM, Trail of Bits had an Agent spend six months building an MASM LSP server, decompiler, static analysis engine, and Lean VM execution model from scratch. These tools uncovered a critical vulnerability that could allow a malicious prover to forge Falcon signatures and steal funds. Static analysis identified more than 400 type-validation flaws, while the Lean work produced 95 machine-verified correctness proofs and also found two subtle bugs that had not been caught by unit tests.

WeChat Official Account: Digital Life Kazike

TypeSafe AI releases Jev, a large model focused exclusively on high-frequency decisions; author benchmarks its classification cost-effectiveness

TypeSafe AI has launched Jev, a large model focused on high-frequency decisions. It does not handle conversation or text generation and outputs judgments only. It is 20–200 times faster than traditional large models, costs $0.042 per million tokens, and charges nothing for output tokens. The author’s benchmarks found that Jev ranked second in accuracy and was cheaper on prescreening tasks, while achieving the highest accuracy and fastest speed on parallel judgment tasks. The model uses the RLCD training method to optimize decision calibration. Applications can be submitted at https://typesafe.ai/, and Vercel was the first to integrate it.

Gary Marcus：The Road to AI We Can Trust

Gary Marcus comments on Trump downplaying AI risks for economic reasons, as an AI hallucination intelligence report nearly triggered a war

Gary Marcus cites an NYT report saying Trump downplayed AI panic and resisted regulation for economic reasons, and reposted Katie Bo Lillis's report that an AI-assisted intelligence report misidentified a Chinese ship as carrying nuclear weapon components due to model hallucination, prompting the US military to urgently intercept it — allegedly "almost triggering a war".

Gary Marcus：The Road to AI We Can Trust

Gary Marcus: what we should worry about more right now is not runaway superintelligence, but large-scale hacking caused by agentic AI

Gary Marcus wrote that what we should truly worry about right now is not runaway superintelligence, but unleashed agentic AI causing internet-scale hacking.

Ethan Mollick：One Useful Thing

Ethan Mollick on the capability gap: the existing capabilities of GPT-6 Astra and Fable 5.1 are far from exhausted

Ethan Mollick wrote that GPT-6 Astra and Fable 5.1 can already reliably complete weeks' worth of human work, but most people have far from exhausted their capabilities, creating a capability gap.