Amnesty International Releases Report Exposing the Human Rights Costs of Generative AI
Amnesty International has published a report titled "Unlawful by design: Exposing the human rights costs of generative AI." It is a fairly hard-hitting document, arguing that generative AI is structured to violate human rights right from the design stage.
The report offers a comprehensive analysis of the human rights impacts of developing and operating AI systems, covering a wide range of issues including privacy violations, discriminatory outputs, worker surveillance, and threats to freedom of expression. As the phrase "Unlawful by design" in the title suggests, its distinguishing feature is that it frames AI's problems not as isolated bugs or accidents, but as structural issues rooted in the system design itself.
The report has also been drawing attention on Hacker News. The very fact that an international human rights organization has put out a report on AI that goes this far arguably signals that the AI governance debate has entered a new stage.
Connecticut Governor Signs Law Requiring Employee Notification of AI Use
Connecticut Governor Ned Lamont has signed a law requiring employers to notify employees about their use of AI. According to Bloomberg Law, it is one of the first state-level laws in the US to establish employer obligations around AI.
Specifically, when companies use AI tools for hiring, performance reviews, and similar purposes, they are required to notify employees accordingly. The law aims to keep AI decision-making processes from becoming a black box and to guarantee workers the right to know how AI is involved in their evaluations or hiring outcomes.
With still no comprehensive federal AI regulation in the US, individual states are starting to move on their own. Connecticut's law could become a model for other states, and similar legislation may well spread going forward.
NPR Highlights the Dangers of Open-Weight AI Models
NPR took up the safety challenges posed by open-weight AI models in an article titled "AI models are free, private, and will never say 'no'."
Because open-weight models can be downloaded and used freely by anyone, they offer benefits in terms of privacy protection and cost reduction. At the same time, it is technically possible to strip off their safety guards and repurpose them for malicious ends. The article emphasizes the danger of models that never say no.
This debate is a key issue for the AI open-source community. Democratizing models drives innovation, but no consensus has yet formed on how to balance that against the risk of misuse. The topic has also sparked discussion on Hacker News, where pro-regulation and anti-regulation viewpoints are clashing.
The Guardian: Silicon Valley's Plan to Conquer the Cosmos with Conscious AI
The Guardian published an interactive piece titled "Our tech overlords are planning for conscious AI to conquer the cosmos," taking a critical look at the transhumanist visions harbored by some Silicon Valley tech leaders.
The piece points out that some technology leaders envision developing "conscious AI" and expanding into the cosmos as beings that transcend humanity. It analyzes how such visions are shaping the direction of real-world AI development and the allocation of funding.
It drew moderate attention on Hacker News with 8 points, where commenters exchanged critical takes like "tech industry messianism" alongside more defensive views such as "sci-fi visions are also a driving force for innovation." It offers a good opportunity to consider where the line lies between practical problem-solving and philosophical ambition when talking about the future of technology.
Qwen3.6-35B vs Gemma4-26B — Surprising Results from Real-World Benchmarks
A benchmark post pitting Qwen3.6-35B-A3B and Gemma4-26B-A4B against each other on a Radeon 7900 XTX has been attracting significant attention in Reddit's LocalLLaMA community.
The results were counterintuitive. Thanks to Qwen's MTP (Multi-Token Prediction), its decode speed of 130 tok/s far exceeded Gemma's 78 tok/s. However, Qwen generates roughly twice as many tokens for the same prompt (much of it consumed by internal reasoning), so in terms of final wall-clock time, Gemma came out about 20% faster.
When compared across six practical tasks (meeting minutes, incident retrospectives, log triage, code review, build-vs-buy analysis, and creative work), Gemma won five of them. Qwen came out on top only in code review.
The poster's conclusion is an interesting one: a faster decoder does not equal a faster response. The insight that with reasoning models, the number of tokens it takes to reach an answer matters more than tokens per second is a practical guideline for local LLM users.
TechCrunch: Are Tech CEOs Uniquely Susceptible to "AI Psychosis"?
In the latest episode of TechCrunch's Equity podcast, the discussion turned to whether tech CEOs are uniquely susceptible to "AI psychosis."
If we call the state of over-believing in AI's potential while losing a realistic grasp of it "AI psychosis," then the remarks of tech industry leaders do perhaps show that tendency. A few months ago, claims that "AGI is just around the corner" were coming one after another, yet a growing number of voices point out that actual deployment is progressing more slowly than predicted.
This debate overlaps with the Guardian article mentioned above: how is tech leaders' excessive enthusiasm for AI influencing product development and business strategy? It is a topic that makes you think anew about the importance of keeping cool-headed judgment.