Bilingual

China warns of 'security backdoor' in Anthropic AI coding tool
16 days ago
- Chinese regulator warns of a 'security backdoor' in Anthropic's Claude Code AI tool, potentially transmitting sensitive data without consent.
- Anthropic claims the mechanism is an anti-abuse measure to detect fraud and unsupported regions via device timezone and routing checks.
- China's National Vulnerability Database advises users to check, uninstall, or upgrade Claude Code to prevent data leaks, and Alibaba bans its use due to security concerns.
- Anthropic accuses Alibaba of reverse-engineering its AI models, while an engineer states the tracking was an anti-abuse experiment being rolled back.
AI Can't Recreate the Thrust Game (But It Can Help You Understand It)
15 days ago
- The author attempted to use Claude AI to recreate the classic 1986 game Thrust in the browser but found the initial result was 'slop' and not playable, lacking the authentic feel.
- Through a process of 'archaeology,' the author used Claude to interrogate disassembled 6502 assembly code, which helped generate detailed specifications for subsystems like physics and sound, leading to an accurate recreation.
- Key to matching the original game's feel was replicating its specific physics timing, including a gated update cadence (6 active slots per 16-tick window) and asymmetric drag, rather than just using the same constants.
- For authentic sound, the author emulated the BBC's SN76489 sound chip and MOS interface in an AudioWorklet, rather than sampling, ensuring precise timing and envelope shapes.
- Graphics and levels were extracted from the original source, using hard-coded sprites for ship rotation and the original's scanline-parity fill for terrain, while demo mode was replicated via timed keypress injection.
- The author concludes that while AI could not directly recreate Thrust, it was invaluable for deeply understanding the original systems, and the recreation process fostered appreciation for the original developers' achievements.
- The article also mentions the author's separate project: a turn-based 4X strategy game built from scratch with a custom WebGPU engine, eight civilizations, and an AI opponent, playable directly in the browser.
How to fix AI-generated pixel art
12 days ago
- AI pixel art often has blurry and inconsistent pixels.
- Pixel Snapper aligns pixels to a grid, preserving shapes.
- Step 1: Quantize colors to create a consistent palette.
- Step 2: Analyze brightness profiles to detect edges.
- Step 3: Determine grid step size and cut edges.
- Step 4: Resample colors in each cell for output.
- Transforms off-grid AI art into clean pixel art.
- Includes a Pixel Art Generator for game development.
The Conservationist Who Turned 40 Terabytes of Public Data into a Video Game
11 days ago
- Raffael Hickisch, a conservationist, co-founded the Chinko Nature Reserve in Central Africa, managing it through public-private partnerships.
- Previously, Hickisch struggled with integrating diverse data sources (settlement, deforestation, fire data) for monitoring protected areas due to technical and budget constraints.
- He discovered exe.dev, which allowed him to use AI agent Shelley to quickly prototype apps, like Five Megapixel Conservation, without large budgets.
- Five Megapixel Conservation aggregates and displays global conservation efforts using GPX tracks on a zoomable map, generating reports for protected areas.
- Hickisch also developed an app using 100 parallel VMs to process LIDAR data from Austria, classifying objects to identify old forests needing protection.
- To engage users, he created a settlers-style game linking to real LIDAR data, highlighting the underlying data's importance over the game itself.
- His tools are open-source, aiming for adoption by governments and nonprofits to democratize custom software building in conservation.
Accretive Editing
15 days ago
- Accretive editing is a failure mode in AI tools where they add addendums instead of correcting text when updating.
- An example involves changing support from Amazon Bedrock to LiteLLM, where AI added an addendum rather than replacing the obsolete information.
- This issue cannot be fixed by simply telling the model to write less or adjust style, as it stems from how LLMs predict text based on multiple inputs.
- The cause may be that LLMs lack human perspective on document updates, focusing on predicting likely text rather than making it true for readers.
- To address accretive editing, focus on instructing AI to replace obsolete text with accurate text, aiming for a document that reads as if written correctly from the start.
OpenAI's first hardware product is the $230 Codex Micro macropad by Work Louder
10 days ago
- OpenAI launched its first hardware product, the $230 Codex Micro macropad, developed with Work Louder.
- The device is a limited edition, programmable mechanical macropad that features six 'agent keys' with colored lights to indicate agent status (idle, thinking, finished, error).
- It includes a dial, joystick, six regular keys for accepting/rejecting changes, and a push-to-talk button, all customizable via Work Louder's Input tool.
- The Codex Micro supports USB-C and Bluetooth, comes with low-profile switches (silent or clicky), and includes 32 additional keycaps.
- While designed for OpenAI's Codex agentic coding tool, it can be used with other applications through additional layers in the Input app.
NotebookLM is now Gemini Notebook
9 days ago
- NotebookLM has been renamed to Gemini Notebook.
- It now includes a secure cloud computer for each notebook to write and execute code, enhancing data analysis.
- Features are available to Google AI Ultra users and Workspace customers, with rollout to all Pro users soon.
- Notebooks are accessible in the Gemini app and will soon integrate into AI Mode in Google Search.
Linus Torvalds to critics of AI coding in Linux: "Fork it. Or just walk away."
8 days ago
- Linus Torvalds strongly supports integrating AI tools like Sashiko into the Linux kernel development, rejecting anti-AI absolutists.
- Sashiko, an AI-powered code review system, can independently find over 50% of bugs but also generates false positives estimated within a 20% range.
- Torvalds advocates for individual choice in using LLM tools, opposing demands to ban AI-generated code and emphasizing the open-source option to fork the project if needed.

Hasty Briefsbeta

ai tools

China warns of 'security backdoor' in Anthropic AI coding tool

AI Can't Recreate the Thrust Game (But It Can Help You Understand It)

How to fix AI-generated pixel art

The Conservationist Who Turned 40 Terabytes of Public Data into a Video Game

Accretive Editing

OpenAI's first hardware product is the $230 Codex Micro macropad by Work Louder

NotebookLM is now Gemini Notebook

Linus Torvalds to critics of AI coding in Linux: "Fork it. Or just walk away."