Sunday — November 2, 2025

The EU's AI Act shuts down Prague's airport facial recognition, an experiment replaces all web server logic with an LLM, and a new benchmark finds AI agents automate just 2.5% of remote work.

News

Czech police forced to turn off facial recognition cameras at the Prague airport

A real-time facial recognition system operated by Czech Police at Prague's airport from 2018 was shut down in August 2025 after being rendered illegal by the EU's AI Act. The Act's requirement for judicial approval for biometric surveillance, which the system lacked, was the primary driver for its termination. A multi-year DPA investigation confirmed the system violated data protection laws, highlighting a systemic issue where Czech police deploy large-scale biometric tools without a compliant legal framework.

Tech companies are firing everyone to "fund AI", spending money on each other

The text is a Reddit block page triggered by a network policy, likely from an automated script or application. It instructs developers to authenticate with API credentials and to use a unique, descriptive User-Agent string to regain access. A unique code is provided for filing a support ticket if the block is believed to be an error.

AI Broke Interviews

LLMs have fundamentally broken the already flawed technical interview process by enabling candidates to generate perfect answers for coding, system design, and behavioral questions, eroding trust in remote evaluations. In response, companies are reverting to in-person interviews to regain a reliable signal of a candidate's authentic reasoning ability. The author proposes a shift towards "AI-resistant" techniques that prioritize dialogue, debugging, and explaining code over rote problem-solving, aiming to measure uniquely human skills like engineering judgment and adaptability that models cannot easily replicate.

Meta readies $25B bond sale as soaring AI costs trigger stock sell-off

Meta is preparing a $25bn bond sale to fund its escalating AI-related capital expenditures. The announcement of these high costs has triggered a stock sell-off, highlighting the significant financial investment required for competitive AI development.

"Our research is greatly sped up by AI but AI still needs us"

Mathematician Timothy Gowers describes crossing a new threshold in his research workflow. While working on a proof, he identified a useful intermediate statement, an experience he suggests is becoming common for mathematicians, implying the integration of AI tools in the discovery process.

Research

Watermarking for Generative AI

InvGNN-WM is a trigger-free watermarking method for GNNs that addresses the fragility of backdoor-based approaches by tying ownership to the model's implicit perception of a graph invariant. This enables robust, black-box verification by predicting algebraic connectivity on a private carrier set, with negligible impact on task performance. The watermark demonstrates strong resilience to common model edits like pruning, fine-tuning, and quantization, and its exact removal is proven to be NP-complete.

Agentic AI Home Energy Management System: Residential Load Scheduling

This paper presents an agentic AI Home Energy Management System (HEMS) where LLMs autonomously coordinate the entire workflow from natural language requests to multi-appliance scheduling. The system uses a hierarchical agent architecture with the ReAct pattern, achieving cost-optimal scheduling without example demonstrations. Evaluations against a MILP benchmark showed Llama-3.3-70B successfully coordinated all appliances, while other models failed at multi-appliance tasks despite succeeding with single appliances. The fully open-sourced system highlights that reliable analytical query handling remains a challenge for LLMs without explicit guidance.

Remote Labor Index: Measuring AI Automation of Remote Work

To bridge the gap between AI performance on research benchmarks and real-world economic value, researchers introduced the Remote Labor Index (RLI). This multi-sector benchmark evaluates end-to-end agent performance on practical, economically valuable projects. Current AI agents perform poorly on RLI, with the top-performing agent achieving only a 2.5% automation rate, providing an empirical baseline for tracking AI-driven labor automation.

Humains-Junior: A 3.8B Language Model Achieving GPT-4o-Level Factual Accuracy

Humans-Junior, a 3.8B model based on Phi-3.5-mini, achieves performance statistically equivalent to GPT-4o on the FACTS Grounding benchmark while being ~19x less expensive via managed APIs. This is accomplished by synergistically combining "Exoskeleton Reasoning" scaffolds with behavioral fine-tuning that teaches protocol compliance rather than domain knowledge. The directed reasoning technique alone was also shown to improve the performance of frontier models like GPT-4o and Gemini-2.5-Pro.

Reconfigurable Analog Computers

Classic analog computers were hindered by slow, manual programming via physical patching. They are now gaining renewed interest as co-processors to circumvent the physical scaling limits of digital hardware. This modern application requires automatic, digitally-controlled reconfiguration systems, known as autopatching, to be viable.

Code

Show HN: Why write code if the LLM can just do the thing? (web app experiment)

"nokode" is an experiment in creating a web server where an LLM, given tools for database access, HTTP responses, and feedback persistence, replaces all application logic. The project successfully built a functional CRUD app, demonstrating the LLM's emergent ability to design schemas, APIs, and UIs from a simple prompt. While the capability was proven, the implementation was impractically slow (300-6000x), expensive (100-1000x), and inconsistent, highlighting that performance, cost, and reliability are the key barriers to this paradigm.

Show HN: KeyLeak Detector – Scan websites for exposed API keys and secrets

KeyLeak Detector is a web application that scans websites for leaked secrets and sensitive information. It uses Playwright for browser automation and mitmproxy for network interception to analyze HTML, JS, and network traffic for hardcoded credentials. The tool employs regex to detect over 50 patterns, including API keys for numerous LLM providers such as OpenAI, Anthropic, Gemini, and Hugging Face.

Cognotik: A New FOSS AI Coding Assistant. For JetBrains IDEs

Cognotik is an open-source, AI-powered development platform featuring a modular architecture for intelligent planning and code generation. It operates on a BYOK model, providing a unified API (JOpenAI) to support a wide range of providers, including OpenAI, Anthropic, Groq, and local models via Ollama. The platform offers multiple interfaces for AI-assisted development, such as a desktop app, a web UI, and an IntelliJ plugin for direct IDE integration.

Diffwatch – Watch AI agents touch the FS and see diffs live

diffwatch is a CLI tool that provides a real-time, colored diff visualization of file changes directly in the terminal. It supports recursive directory watching and uses smart filtering to ignore noise like .git and node_modules. The TUI is built with Bubbletea, and its event pipeline leverages fsnotify and a debouncer to efficiently handle rapid file modifications.

Jotit – command-line notes with AI search and summaries

Jotit is a command-line note-taking tool that uses a local SQLite database with vector capabilities for storage and retrieval. It enables quick note capture and leverages AI for semantic search and summarization. The architecture uses background jobs for creating local embeddings and an external LLM API for generating summaries.

    The EU's AI Act shuts down Prague's airport facial recognition, an experiment replaces all web server logic with an LLM, and a new benchmark finds AI agents automate just 2.5% of remote work.