开发者将微型 Transformer 模型塞进软盘,让 486 老电脑也能用上 AI
IT之家 11 月 22 日消息,开发者 WollyMonjac 完成了一项颇具复古科技色彩的项目:将一个语言模型塞进软盘,让运行 DOS 系统的 486DX 和 Pentium 老电脑也能使用 AI。
不过需要说明的是,WollyMonjac 开发的 Tiny 2.1 并不是大型语言模型(LLM)。尽管它采用了与大型模型相同的 Transformer 架构,但开发者对它的定位相当朴素:一个使用 Turbo C 2.0 编写的微型问答 Transformer 模型。
在 486 电脑时代,普通 PC 格式化软盘的存储容量还不到 147.5 万字节,因此想要在如此有限的空间内运行现代 AI 软件并不容易。Tiny 2.1 则尽可能压缩模型体积,最终实现了以下配置:
参数量:36.7 万个,采用 4 层网络结构,隐藏维度为 96,逐字母读取和生成内容。
权重精度:8 位量化,每行使用一个缩放系数,模型总大小约为 380 KB。
训练数据:约 3500 条事实信息,每条信息都采用多种不同的问法进行训练。
要在老式 PC 上运行 Tiny 2.1,用户需要确保系统拥有约 540 KB 的可用常规内存。
IT之家注意到,这款软件的界面设计也相当完整,包含对话历史窗口、神经网络监控面板,以及用于输入问题的 Ask Tiny 窗口。按照 DOS 软件的传统习惯,界面还提供了一系列 F 键快捷操作。
此外,这款软件还有一个颇具复古未来主义色彩的功能:通过 PC 内置扬声器,以机器人语音朗读模型生成的回答。
开发者对这款 DOS 程序的定位显然十分谦逊。GitHub 项目页面介绍称,该软件由 Fable 5.2 编写,并由 Opus 5.5 训练。WollyMonjac 还开玩笑地承认,除了训练数据涉及的知识之外,这个语言模型“会非常自信地编造内容”。
近几个月,外界对大型语言模型幻觉问题的讨论有所减少,但放在这款复古 AI 项目上,这种一本正经地胡说八道的能力反倒有些像一项有趣的功能。
如果用户觉得 Tiny 2.1 的知识储备不够丰富,WollyMonjac 还在采用 MIT 许可证发布的软件包中提供了专门的训练文件夹。有兴趣的用户可以自行补充训练数据,打造一个专门针对自己感兴趣领域的问答 Transformer 模型。
如今,让老软件运行在全新的电脑硬件上已经不是什么难事。但反过来,让现代软件和游戏运行在几十年前的老硬件上,同样正在成为一股有趣的开发趋势。
除了此次受到关注的 Tiny 2.1,此前也出现过不少类似项目。例如,有开发者为 DOS 系统制作了 ChatGPT 客户端,让 1984 年推出的 IBM PC 也能使用 ChatGPT;有人成功让《我的世界》在 Game Boy 上运行;有人利用生成式 AI 为 Commodore 64 生成 8×8 像素精灵图;甚至还有开发者将《毁灭战士》(Doom)移植到了 Atari ST 电脑上。
这些项目虽然实用价值有限,却展示了开发者如何突破老旧硬件的性能和存储限制,让现代技术与经典计算设备产生有趣的碰撞。
本文来源:https://www.ithome.com/1/030/699.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Developer fits a tiny Transformer model onto a floppy disk, letting old 486 computers use AI
IT Home reported on November 22 that developer WollyMonjac completed a project with a distinctly retro-tech flavor: fitting a language model onto a floppy disk so that old 486DX and Pentium computers running DOS can also 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's positioning for it is quite modest: a tiny question-answering Transformer model written in Turbo C 2.0.
In the era of 486 computers, the storage capacity of a formatted floppy disk on an ordinary PC was less than 1.475 million bytes, so running modern AI software in such limited space was no easy task. Tiny 2.1 compressed the model size as much as possible, ultimately achieving the following configuration:
Parameter count: 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, using one scaling factor per row, with a total model size of about 380 KB.
Training data: about 3,500 pieces of factual information, with each piece trained using a variety of different question phrasings.
To run Tiny 2.1 on an old PC, users need to ensure the system has about 540 KB of available conventional memory.
IT Home noted that the software's interface design is also quite complete, including a conversation history window, a neural network monitoring panel, and an Ask Tiny window for entering questions. Following the traditional conventions of DOS software, the interface also offers a series of function-key shortcuts.
In addition, the software has a feature with a distinctly retro-futuristic flavor: reading the model's generated answers aloud in a robotic voice through the PC's built-in speaker.
The developer's positioning of this DOS program is clearly very humble. The GitHub project page states that the software was written with Fable 5.2 and trained with Opus 5.5. WollyMonjac also jokingly admitted that, apart from the knowledge covered by the training data, this language model 'will fabricate content very confidently.'
In recent months, outside discussion of the hallucination problem in large language models has decreased, but in the context of this retro AI project, this ability to talk nonsense with a straight face seems more like an interesting feature.
If users feel that Tiny 2.1's knowledge base is not rich enough, WollyMonjac also provided a dedicated training folder in the software package released under the MIT license. Interested users can supplement the training data themselves and build a question-answering Transformer model tailored to their own areas of interest.
Nowadays, getting old software to run on brand-new computer hardware is no longer difficult. But the reverse—getting modern software and games to run on hardware decades old—is likewise becoming an interesting development trend.
Besides Tiny 2.1, which has drawn attention this time, there have been many similar projects before. For example, a developer created a ChatGPT client for DOS, allowing the IBM PC released in 1984 to use ChatGPT; someone successfully got Minecraft to run on the Game Boy; someone used generative AI to generate 8x8 pixel sprites for the Commodore 64; and even a developer ported Doom to the Atari ST computer.
Although these projects have limited practical value, they demonstrate how developers break through the performance and storage limitations of old hardware, bringing modern technology into interesting collision with classic computing devices.
本文来源:https://www.ithome.com/1/030/699.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
