This week, AI showed up in a hiring algorithm manufacturing bias from scratch, on a picket line in front of a Hyundai plant, and in the mouth of Dave Eggers standing inside OpenAI headquarters. A governor is using it to audit every regulation in New York State. Linus Torvalds told its critics to leave.

The abstract debates ended a while ago. Here's what's actually happening.

New research covered by MIT Technology Review finds that large language models used in hiring don't merely reflect the biases baked into their training data, they generate novel biases that have no direct human analog, favoring certain candidate profiles for reasons that can't be traced back to any identifiable human prejudice. This isn't a calibration problem or a legacy dataset problem. It's the model doing something generative in exactly the wrong direction.

What makes this particularly sharp is the scale. AI screening tools are already processing millions of real résumés, which means this isn't a lab finding waiting for real-world implications. The experiment is running. The subjects didn't consent to being in it.

The lesson isn't that AI is uniquely malicious in hiring contexts (human screeners are demonstrably biased, too). The lesson is that "we audited it for known biases" is no longer a sufficient guarantee when the system is capable of manufacturing ones you didn't think to look for. Source: MIT Technology Review

Workers at a Hyundai auto factory have gone on strike specifically over the introduction of humanoid robots on the production floor; making this, per Ars Technica's reporting, the car industry's first labor action driven explicitly by humanoid automation fears rather than wages or conditions. Hyundai, which owns Boston Dynamics, has been among the most aggressive automakers in piloting humanoid robots for physical assembly tasks.

For years, "AI will replace jobs" has been a forecast dressed up as a philosophical puzzle. This is what it looks like when the forecast arrives at the factory gate and people decide they'd rather not wait to find out how it resolves. The strike also clarifies something the tech industry tends to obscure: the question of who bears the transition costs is a political question, not a technical one, and workers are starting to answer it on their own terms. Source: Ars Technica

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.

Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.

89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Start flowing free

Author Dave Eggers (McSweeney's founder, Pulitzer finalist, author of A Heartbreaking Work of Staggering Genius) addressed OpenAI staff directly and told them that ChatGPT was "silencing an entire generation," according to reporting by The Verge. His argument wasn't about plagiarism or copyright. It was about atrophy: that writing is a cognitive and emotional technology for thinking through difficulty, and that outsourcing it doesn't just change the output, it prevents the development of the mental capacity writing is meant to build.

The framing matters. Most critiques of AI writing focus on authenticity or attribution. Eggers's critique is neurological and developmental. It’s closer to arguing that using a calculator before you understand arithmetic doesn't just produce correct answers, it forecloses a certain kind of mathematical intuition from ever forming.

Whether you agree or not, it's a more serious argument than the discourse usually gets, and it deserves a serious answer from the people building the tools, not just reassurances that users can always choose to write by hand. Source: The Verge

Moonshot AI released Kimi K3 and Alibaba released a new version of Qwen within a close window of each other, with both models benchmarking at or near the frontier of globally available AI systems, according to The Verge's reporting. Both are open-weight releases, meaning their parameters are publicly accessible (a deliberate strategic choice that accelerates adoption and embeds these models into global developer ecosystems).

The open-source angle is the underappreciated part of this story. A model at the frontier that's also open-weight doesn't just compete with GPT-4 class systems, it becomes infrastructure. Developers in Southeast Asia, Africa, and Europe building on Qwen or Kimi K3 are building on systems whose values, tendencies, and blind spots were shaped in Beijing and Hangzhou. The geopolitics of AI have always been about more than compute clusters and export controls. They're about which models become the default substrate for the next decade of software. Source: The Verge

Linus Torvalds, the creator and lead maintainer of the Linux kernel (the open-source operating system that runs the majority of the world's servers, smartphones, and supercomputers) has responded to critics of AI-assisted code contributions by telling them, in characteristically blunt terms, to fork the project or walk away, per Ars Technica. Torvalds's position is that code quality is what matters, not whether a human or a model assisted in producing it.

This is a significant institutional moment. Linux is not a startup with a growth mandate. Rather, it's a 34-year-old project governed by one of the most rigorous peer review processes in software. When Torvalds says AI-assisted code is acceptable if it passes review, he's setting a precedent that will ripple through enterprise procurement, open-source licensing debates, and academic software curricula.

The deeper implication is practical: the kernel's existing review process was designed to catch bad code from humans. Whether it's equally well-suited to catch the specific failure modes of AI-generated code (confidently wrong, subtly inconsistent, difficult to trace) is a question the project is now answering by doing. Source: Ars Technica

The human mind is the original generative engine.