Newshot

Developer squeezes a tiny Transformer model onto a floppy disk so a 486 can run AI

AI summary

Developer WollyMonjac has packed a language model onto a floppy disk so 486DX and Pentium PCs running DOS can use AI. Tiny 2.1 is not an LLM but a miniature question-answering Transformer written in Turbo C 2.0, with 367,000 parameters, 4 layers, a hidden dimension of 96, 8-bit quantized weights, a total size of about 380 KB, and training on roughly 3,500 facts; running it requires about 540 KB of free conventional memory.

Translation(model translation, not the source's own text)

Developer Squeezes a Tiny Transformer Model onto a Floppy Disk, Bringing AI to a 486-era PC

IT之家 November 22 news: Developer WollyMonjac has completed a project with a decidedly retro-tech flavor: fitting a language model onto a floppy disk so that 486DX and Pentium-era PCs running DOS can use AI.

It should be noted, however, that Tiny 2.1, developed by WollyMonjac, is not a large language model (LLM). Although it uses the same Transformer architecture as large models, the developer positions it quite modestly: a miniature question-answering Transformer model written in Turbo C 2.0.

In the 486 era, an ordinary PC's formatted floppy disk held less than 1.475 million bytes, so running modern AI software in such limited space is no easy task. Tiny 2.1 compresses the model as much as possible, ultimately achieving the following configuration:

  • Parameters: 367,000, using a 4-layer network structure with a hidden dimension of 96, reading and generating content letter by letter.
  • Weight precision: 8-bit quantization, with one scaling factor per row, for a total model size of about 380 KB.
  • Training data: about 3,500 facts, each trained with multiple different phrasings of the question.

To run Tiny 2.1 on an old PC, users need to ensure the system has about 540 KB of available conventional memory.

IT之家 notes that the software's interface design is also fairly complete, including a conversation history window, a neural network monitoring panel, and an Ask Tiny window for entering questions. Following the tradition of DOS software, the interface also provides a series of F-key shortcuts.

In addition, the software has a rather retro-futuristic feature: it reads the model's generated answers aloud in a robotic voice through the PC's built-in speaker.

The developer is clearly quite humble about this DOS program. The GitHub project page states that the software was written by Fable 5.2 and trained by Opus 5.5. WollyMonjac also jokingly admits that, apart from the knowledge contained in its training data, this language model "will very confidently make things up."

In recent months, discussion of the hallucination problem in large language models has somewhat subsided, but in this retro AI project, this ability to spout nonsense with a straight face seems rather like an interesting feature.

If users feel that Tiny 2.1's knowledge base is not rich enough, WollyMonjac also provides a dedicated training folder in the software package released under the MIT license. Interested users can add their own training data to build a question-answering Transformer model tailored to their own fields of interest.

Nowadays, getting old software to run on brand-new computer hardware is no longer difficult. But conversely, getting modern software and games to run on decades-old hardware is likewise becoming an interesting development trend.

Besides Tiny 2.1, which has drawn attention this time, many similar projects have appeared before. For example, a developer made a ChatGPT client for DOS, letting the IBM PC launched in 1984 use ChatGPT; someone successfully got Minecraft running on a Game Boy; someone used generative AI to generate 8×8 pixel sprites for the Commodore 64; and a developer even ported Doom to the Atari ST.

Although these projects have limited practical value, they show how developers can break through the performance and storage limits of old hardware, creating interesting collisions between modern technology and classic computing devices.

Related Articles

  • New York's new concert hall Domino's official website makes a fool of itself, promoting "handmade" while being full of AI-generated traces
  • IMF Managing Director warns: excessive AI investment enthusiasm may lead to further adjustments in financial markets
  • Who is responsible when AI agents transfer money or make purchases without authorization? Multiple banks warn users that losses may have to be borne by themselves
  • Adobe announces four AI experimental projects: photos turn into 3D scenes in seconds, and AI characters can be designed through chat
  • Bloomberg estimates: AI contributed about a quarter of the UK's economic growth this year
  • Not grand enough? Report says Meta will stop naming AI models after fruits

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗