The afternoon of August 5 brought a mix of announcements from model providers, real-world implementations, and the realities of AI risks and monetization.
Qwen Developers Share Next-Generation Plans in AMA
Alibaba's Qwen team held a developer Q&A (AMA) on X (formerly Twitter), discussing the recent Qwen3.8-Max and the direction of its next-generation models. The topic that drew the most interest was the introduction of a 27B model, with the team saying it will be "released soon" and will be "a bigger leap than the previous generation," suggesting they revisited their size lineup in response to community requests.
On the design side, Qwen3.8-Max is said to have 2.4 trillion total parameters, of which roughly 95 billion are active during inference (a MoE configuration). This "huge but only partially activated" design is positioned to balance performance against cost. The AMA also touched on handling long videos, describing a hierarchical video memory that structurally retains scenes, characters, and their relationships over time. Updates to the Qoder coding environment and QwenWork are coming soon, and while the technical report is not yet public, the team plans to maintain a roughly monthly release cadence. Note that figures such as total parameters and active values are based on reports and community information; final specifications await official disclosure.
MacPaw Brings On-Device Inference to the Mac with Liquid AI
MacPaw, the company behind a Mac app store, is building a local version of its AI assistant "Eney" on Liquid AI models. By completing inference entirely on-device without going through the cloud, the company aims to deliver privacy and responsiveness at the same time. MacPaw also indicated that it expects to offer on-device inference to developers through its store.
Liquid AI is a developer known for small, efficient models, and some see it as a good fit for resource-constrained environments like the Mac. Details on supported models and accuracy remain limited at the time of reporting, so real-world evaluation will have to wait for future developments.
AI Review Misses Hardware Wallet Flaws
A case has been reported that calls into question the accuracy of AI-powered code and security reviews. Regarding the Coldcard hardware wallet, Coinkite, its developer, has pointed out in public records that an AI-assisted review missed an actual vulnerability.
Automated AI reviews are attractive for their processing speed and coverage, but the risk of overlooking issues that require specialized expertise or context has long been pointed out. This incident once again highlights the need not to overtrust AI output in security-critical areas and to layer human verification on top. Technical details are limited to what appears in the public record, so the full picture remains fragmentary.
TIME Serves AI Bots an Alternate Site with Ads Built In
Among media companies scrambling to respond to AI crawlers, TIME has been reported to serve AI bots a different version of its website with ads incorporated. By deliberately preparing pages that differ from those shown to human readers, the magazine apparently aims to secure revenue opportunities even from bot-driven traffic.
As AI search and summarization services spread, there has been growing debate over the problem of content being crawled and cited without translating into revenue. How far this approach of "monetizing AI traffic through a separate route" will spread remains unclear, but it is drawing attention as one defensive measure for publishers. The details are based on an analysis of the article in question.
Study Finds People Prefer Stories Written by AI
A study examining whether people can tell human-written text from AI-generated text found that readers tended to prefer stories written by AI. The findings, published in the journal Judgment and Decision Making, published by the University of Cambridge, come from an experiment titled "Bot or not."
In creative writing, AI quality has been closing in on human levels, and the phenomenon of readers unknowingly choosing AI-produced work has been observed for some time. This study reconfirms that tendency from the angle of preference, offering material for thinking about generative AI's impact on creative work. Some aspects of the experiment's design and scale require checking the original paper.
Data Center Realities Put the Brakes on the AI Revenue Boom
A view has emerged that pours cold water on the sustainability of the data center investment underpinning the AI boom. A Reuters column argues that physical constraints — power, land, and supply capacity — are acting as a "reality check" that could erode big tech's expectations of expanding AI revenue.
In the recent earnings season, too, voices questioning the payback of AI-related spending have grown louder. When infrastructure constraints will surface as pressure on earnings remains unclear, but how far infrastructure can keep up with demand growth looks set to become a key point of debate going forward.