- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
Researchers have unearthed hundreds of thousands of cuneiform tablets, but many remain untranslated. Translating an ancient language is a time-intensive process, and only a few hundred experts are qualified to perform it. A recent study describes a new AI that produces high-quality translations of ancient texts.
It’s pretty freaking great at stuff like that though. We use a custom programming language at work, there are similarities with Haskell and others, but also many differences.
We had a little game where a colleague had put together some team-exercises. He had encrypted a message in base64 and therein written instructions for code, in our custom language that when run gave you an output.
ChatGPT managed to print out the, 100% non random output, and 100% stuff that’s never been anywhere on the internet, without trouble.
Google’s DeepMind was able to teach itself Indonesian without being directly trained on how to do so. Ancient Sumerian doesn’t seem too far fetched, all things considered!
There was a funny bit on WANShow a few months back where they demonstrated tricking ChatGPT into speaking Dutch (I think. It might have been another language). It vehemently insisted that it didn’t know Dutch, and could only talk to them in English. The messages saying this were written in Dutch.
As a Dutch speaker, chatgpt always was able to speak Dutch though, tested it very early on
Google’s DeepMind was able to teach itself Indonesian without being directly trained on how to do so. Ancient Sumerian doesn’t seem too far fetched, all things considered!
I can’t even wrap my head around how a large language model can do this.
I can’t even wrap my head around how humans do this.
That’s the problem, you see… it is great for simple things. Then you start believing in it and give more complicated tasks. It will fail, you will never know until it is too late. We are doomed…
I’ve found that after using it for a while, I developed a feel for the complexity of the tasks it can handle. If I aim below this level, its output is very good most of the time. But I have to decompose the problem and make it solve the subproblems one by one.
(The complexity ceiling is much higher for GPT-4, so I use it almost exclusively.)