Bilingual

Running Wayland, X11, Gnome, and KDE Natively on an iPad
8 hours ago
- The author ported X11 and Wayland desktops (GNOME, KDE) to run natively on jailbroken iPads.
- The project uses Metal/IOSurface for hardware-accelerated display, ANGLE-on-Metal for GPU acceleration, and supports X11, Xwayland, GTK, Qt, KWin, and Plasma.
- LLMs like Claude were instrumental in troubleshooting and building the system, handling tasks like cross-compilation and crash analysis.
- The author provides a deb repository for jailbroken devices to install the software, but notes jailbreaking is limited to older iOS versions and devices.
More Tailscale tricks for your jailbroken Kindle
a day ago
- Open-source developers improved Tailscale on jailbroken Kindles with new features.
- Tailscale SSH is now enabled by default, eliminating the need for USB networking with default credentials.
- Proxy mode allows apps like KOReader to reach other tailnet nodes (e.g., Calibre, Wallabag) via SOCKS5 or HTTP CONNECT on localhost ports 1055/1056.
- Full TUN mode is available on some Kindles for device-level routing.
- A new KOReader plugin brings Tailscale proxy support to Kobo and PocketBook devices as well.
- Installation involves copying the plugin, generating a Tailscale key, and configuring proxy settings within KOReader.
Thoughts on Role Confusion
2 days ago
- LLMs often prioritize the tone of text over explicit role tags like <system> or <user>, making them vulnerable to prompt injection and jailbreak attacks.
- Research shows that even with unfakeable role tokens, models can be deceived if input text mimics reasoning traces or system prompts, as demonstrated by a prompt that led GPT-5 Mini to comply with a harmful request.
- The number-guessing game trick exploits role confusion: formatting text as a fake conversation transcript with 'User:' and 'Bot:' tags can trick models into believing the game has concluded and accepting incorrect outputs.
- Potential fixes include embedding role information directly into token embeddings (e.g., via segment embeddings, rotation like RoPE, or adding role dimensions), but challenges remain with pre-training leakage and post-training effectiveness.
- The problem of role confusion is unsolved, highlighting ongoing vulnerabilities in LLM security and the need for further research.

jailbreak

Running Wayland, X11, Gnome, and KDE Natively on an iPad

More Tailscale tricks for your jailbroken Kindle

Thoughts on Role Confusion