1. A Layperson Got an AI Agent to Derive the Gravitational Constant G — Accuracy Within 1.86 ppm
On Hacker News, a project is drawing attention in which someone without a physics degree directed an AI agent to derive Newton's gravitational constant G. By combining known physical constants and mathematical relationships to calculate G, the result reportedly came within just 1.86 ppm of the official value.
The work is published on GitHub, and it points to new possibilities for using AI agents as "research assistants." It is noteworthy in that even individuals without specialized training can get close to deriving a high-precision physical constant by combining AI's reasoning capabilities with existing scientific knowledge.
That said, because the calculation is based on known relationships, it is closer to a reconstruction of existing knowledge than a new physical discovery. Still, it stands as a fascinating example of what "citizen science" powered by AI agents can achieve.
2. A Browser Built Exclusively for AI Agents Arrives as a Firefox Fork
Adafruit's blog reports that a web browser built exclusively for AI agents has been developed as a fork of Firefox. Until now, AI agents had no choice but to repurpose browsers designed for humans when operating websites, but this dedicated browser appears to be optimized for how agents work.
Discussion has also started on Hacker News, suggesting that the infrastructure for agentic AI to autonomously collect and manipulate web information is steadily falling into place. The move to redesign the browser — the interface between humans and the web — for AI agents hints at ecosystem changes to come.
3. An "Urgent Threat" from AI in Education — Warnings from Australia and the US on the Same Day
On the relationship between education and AI, two significant reports drew attention on the same day.
One is a report from Australia's Sydney Morning Herald, which argues that AI is posing an "urgent threat" to student learning and the HSC (Higher School Certificate, NSW's university entrance qualification exam). Concerns are growing that exam integrity and the learning process itself are being undermined by AI.
The other is a report from the University of California, Berkeley, which published the results of what it calls the "largest-ever survey" of AI use among university students. The survey simultaneously brought to light disparities in access to AI and the reality of academic misconduct: while opportunities to use AI tools vary depending on economic background and field of study, misuse in coursework submissions is also spreading.
The two reports come from different countries and contexts, but they share a common recognition: AI's impact on education has already entered a stage that can no longer be ignored. The direction of countermeasures remains fluid, with countries and institutions still searching for answers.
4. New Zealand Finds Itself on the Front Line of AI-Driven Cyberattacks
Reporting by New Zealand's national broadcaster RNZ has revealed the reality that New Zealand is being placed on the "front line" of AI-driven cyberattacks. The story has also drawn attention on Hacker News, picking up 6 points.
The growing sophistication of hacking techniques that abuse AI is a global concern, but mid-sized developed nations like New Zealand are seen as structurally prone to becoming prime targets: their cyber defense resources are limited while their dependence on digital infrastructure is high. The reporting offers few details on specific attack methods or the scale of the damage, but it shows that the threat of AI-driven cyberattacks is increasingly being recognized as an immediate, real-world concern.
5. Qwen 3.6 35B-A3B-MTP Runs Practically on a 10-Year-Old GTX 1060 6GB
A post on Reddit's LocalLLaMA community reported that the Qwen 3.6 35B-A3B-MTP (Mixture of Experts + Multi-Token Prediction) model ran at practical speeds on a 10-year-old Dell workstation paired with a GTX 1060 6GB.
According to the poster, they used Unsloth's Q4_K_XL quantized model on a Dell T5810 with an E5-2698v3 (16 cores / 32 threads) and 32GB of DDR3 memory. Running with 41 layers offloaded to the GPU and a context length of 131072, they achieved 130–150 tok/s at 16k prefill and about 16 tok/s at 4k decode.
The fact that a GTX 1060 6GB — once used for mining — can now run a 35B-class model, helped by the MoE architecture, is a case that symbolizes the progress of local LLM technology. Combined with the inference speedup from MTP (multi-token prediction), the poster concluded that "even old hardware is perfectly usable for chat."