🎮 Gemini 3 Pro vs. 2.5 Pro in Pokémon Crystal
🔗 Read more 🔗
🧩 Claude in Chrome
A useful reminder that embedding AI directly into browsers expands the attack surface and raises the bar for clear risk communication and user awareness.
Anthropic’s announcement about Claude running inside Chrome highlights new security risks associated with browser-integrated AI, including prompt injection and unintended actions. The post explains how such attacks might leak sensitive data or execute unwanted commands, and outlines mitigations currently being tested.
🔗 Read more 🔗
📊 Show HN: HN Wrapped 2025 — Your Year on Hacker News, Reviewed by an LLM
More fun than rigorous, but still a clever snapshot of how interests and conversations in the tech community continue to evolve.
HN Wrapped 2025 provides a lighthearted retrospective of the year’s key moments, discussions, and predictions on Hacker News. It surfaces trends, popular topics, and community dynamics from the past year.
🔗 Read more 🔗
⏱️ Measuring How Long AI Can Stay on Task
By focusing on sustained task execution rather than benchmarks, this work offers a more realistic lens for evaluating near-term automation risk and potential.
A METR research article analyzes how long AI systems can autonomously carry out complex tasks, showing that this effective time horizon has doubled roughly every seven months over the past six years. While today’s models handle short, well-scoped problems well, they still struggle with long, multi-step workflows. The authors argue this metric better reflects real-world AI capability and predict rapid progress toward month-long autonomous work.
🔗 Read more 🔗
📐 Proving Randomized MaxCut Bounds in Lean4
A compelling illustration of how theoretical computer science and modern proof assistants increasingly reinforce each other.
This deep technical article walks through formalizing the Randomized MaxCut approximation algorithm in Lean4. It shows how to encode probabilistic reasoning and combinatorial optimization, constructing proofs of expected approximation guarantees step by step. The discussion highlights how Lean abstractions like Finset and proof tactics model randomness and expectation.
🔗 Read more 🔗