Nolan Lawson benchmarks AI agent costs across JavaScript frameworks
Nolan Lawson published a benchmark measuring the token costs and performance of AI agents building a small Pokémon app with different JavaScript frameworks. The test used Opus 5.5 in a sandbox environment, evaluating development time, token usage, and Lighthouse scores. Lawson notes that results are stochastic and that multi-page applications often outperform single-page apps in this context. He argues that the web community relies too heavily on intuition rather than evidence when selecting tools, especially in the agentic era.
nolanlawson.com
·
HN discussion
Calegari urges mathematicians to adapt to AI's impact
Frank Calegari argues that the mathematics community should stop mourning AI's rise and instead adapt its practices. He notes that OpenAI's October release of over 700 papers marked a watershed, despite criticism of their quality and attribution. Calegari suggests that mathematicians have more control over hiring and publishing than over AI companies. He recommends emphasizing breadth in graduate education and encouraging senior researchers to explore new ideas rather than just using AI to answer existing questions. This shift, he believes, is necessary to maintain the discipline's relevance and excitement.
terrytao.wordpress.com
·
HN discussion
Plymouth police chief unplugs Flock cameras after company ignores contract cancellation
Plymouth Police Chief Eric Fadden ordered staff to disconnect power cords from Flock cameras after the company refused to acknowledge the city’s contract cancellation. The city council voted to end the agreement in September, and the police department suspended user accounts and sent formal cancellation notices. Flock, which owns the hardware, has not provided a timeline for removal. Fadden confirmed the cameras are now unplugged to ensure they are not recording license plates.
thedrive.com
·
HN discussion
Researchers prove the Platonic Representation Hypothesis using JEPAs
Researchers have formally proven the Platonic Representation Hypothesis, demonstrating that Joint-Embedding Predictive Architectures learn consistent world representations across modalities. The theorem holds when models are trained on "good views," defined as paired observations of the same scene. The authors show that next-token prediction functions as a JEPA using "bad views," which explains why language models require significantly more scale to align with vision systems. Validated across 940 million web images, the study finds that good views require 42 times fewer training samples to achieve the same alignment as next-token prediction.
arxiv.org
·
HN discussion
OmniDiff introduces content-aware Git diffing in Rust
Developer ivankovic has released OmniDiff, an open-source Git diffing tool written in Rust. The project claims to handle 99.95% of code changes robustly, using content-specific algorithms for 98.8% of diffs. It supports various file types, including PDFs and web fonts, and offers prebuilt binaries for Linux, macOS, and Windows. A VS Code extension is also available, though it currently relies on upstream features that are not yet released. The tool aims to provide faster and more accurate diffing than standard Unix utilities.
github.com
·
HN discussion
Proton CEO dismisses security testing as 'loser thinking'
Proton CEO Andy Yen dismissed an engineer’s request for time to pentest new code as “bullshit and loser thinking,” urging the team to rely more on AI instead. The author argues this is not an AI issue, as using AI for code is standard if proper guardrails exist. For a privacy company, however, careful shipping is part of the product. The author notes that while other firms adopt coding agents without humiliating staff, Proton’s leadership treats safety constraints as obstacles to speed. This shift undermines the confidence users pay for, leading the author to question whether they can still recommend the service.
blog.raed.dev
·
HN discussion