Llama.ttf: A font which is also an LLM (fuglede.github.io)

Article summary

A font file called llama.ttf has been created, which also functions as a large language model and inference engine, allowing it to generate text and interact with users. This is made possible by the HarfBuzz font shaping engine, which can execute arbitrary code, including WebAssembly. The font can be used in any application that supports HarfBuzz and WebAssembly, without requiring any modifications to the application. The creator of the font has demonstrated its capabilities in a video.

Main themes

• Font hacking
• Large language models
• WebAssembly
• Security risks
• Innovative uses of technology

What commenters say

• The ability to execute arbitrary code in a font file poses significant security risks, including the potential for phishing attacks and data exfiltration.
• The use of WebAssembly in font files is a horrible idea due to the increased complexity and attack surface it introduces.
• Sandboxing and quotas can mitigate the security risks associated with executing arbitrary code in font files, but they are not a foolproof solution.
• The benefits of being able to run a large language model in a font file, such as enabling text generation and interaction, outweigh the potential security risks.
• The complexity and unpredictability of arbitrary code in font files make them difficult to analyze and restrict, which could lead to unintended consequences.
• The idea of executing arbitrary code in font files is not new and has been explored in other contexts, such as with TIFF files.
• The use of WebAssembly in font files could become a standard feature in the future, making it an expected part of general-purpose computing devices.
• Restricting the code that can be executed in font files to less than Turing complete, such as formula evaluation, could be a more secure alternative.