本文作者利用一个仅限于本地数据的旧式神经网络char-rnn，对2017年收集的近4500个万圣节服装创意进行了重新训练。即使没有互联网的广泛信息，该模型仍能将现有数据进行有趣重组，生成如“Skypug”、“Hungry Boats”和“Mid wonka”等新奇组合。作者还发现，更大的模型版本能更好地融合服装创意，产生“science horse”和“Captain Gay”等新角色。文章强调，AI模型的结果受限于其训练数据的完整性，并邀请读者提交新的服装创意以更新数据集，以期未来能再次训练模型。

💡 **本地化AI的创意重组能力**：即使char-rnn等AI模型仅限于本地数据训练，也能通过对旧有数据集（如2017年的万圣节服装创意）进行重组，生成出令人意想不到的新颖创意。这证明了AI在有限信息下进行发散性创新的潜力。

🚀 **模型规模与创意融合效果**：作者通过对比发现，更大版本的char-rnn在处理服装创意时表现更佳，能更有效地将不同元素融合，产生更具独创性和趣味性的新角色名称，例如“science horse”和“Captain Gay”，这暗示了模型参数对创意输出质量的影响。

📊 **训练数据的重要性与局限性**：文章明确指出，AI模型（包括大型语言模型如ChatGPT）的输出结果直接受其训练数据的影响。如果当前事件或特定视角在训练数据中缺失，AI将无法体现。因此，更新和丰富训练数据对于AI生成更全面、更具时效性的内容至关重要。

💌 **社区共创与数据更新**：作者通过公开的提交表单邀请读者贡献新的万圣节服装创意，旨在更新char-rnn的训练数据集。这种社区共创的方式不仅能丰富AI的学习内容，也为AI的持续进化提供了动力，作者表示若收到足够的新提交，将再次训练模型。

I've recently been experimenting with one of my favorite old-school neural networks, a tiny program that runs on my laptop and knows only about the data I give it. Without internet training, char-rnn doesn't have outside references to draw on (for better or for worse) but it still manages to remix the data into interesting new things.

In 2017 I asked AI Weirdness readers to help me crowdsource a list of Halloween costumes and then trained char-rnn on the nearly 4,500 submissions I got. Today I'm returning both to the dataset and to char-rnn (here's a version that runs on modern Python), mainly because they still entertain me. My laptop is more powerful now than the 2010 Macbook I was using back then, so I'm able to run a bigger version of char-rnn. I actually can't tell whether it helps. But I do know I'm entertained:

The Skypug
Hungry Boats
Mid wonka
Burderous bread cat
Holy Cheesarenda
Moth fairy
A magicial slice
Fall wearing monster
The Godfish

I checked, and nobody in the training data from 2017 was using "mid" as an adjective, so "Mid wonka" is a happy coincidence. The larger version of char-rnn was better than I expected at remixing costumes, producing interesting new characters.

science horse
Lady Doo
Captain Gay
Silence Minister
Cheetos Captain
A scorph Doo
Undead Mario
Sailor Who

There were a couple of Scooby Doo costumes in the original training data, which is probably why the neural net is putting doo in its costumes.

Know what was not in its 2017 era training data? Kpop Demon Hunters, which I have it on good authority will not be an unknown costume in 2025. For fun I asked the neural net to complete the phrases "Kpop " and "Kpop D" and "Kpop De":

Kpop Punk
Kpop and the man and a bus
Kpop Bader Ginsburg
Kpop Dog
Kpop Donald science
Kpop Devil Monster
Kpop Dead Death
Kpop Demetic
Kpop Dead of Turtles

This holds for larger language models like ChatGPT as well, of course. If a current event or a perspective is missing from the training data, it's missing from the result.

The submission form for crowdsourced Halloween costumes is still open, so if you have a few costumes you've seen or dreamed of recently, you can help bring the training data up to date! If I get enough new submissions maybe I'll train the neural network again. (The dataset as of Oct 4 2018 is available on my github).

Bonus content for AI Weirdness supporters: a few more of my favorite costumes trained from the 2018 dataset!