只用1931年前文本训练的AI Talkie：称含铅油漆无害、人类登不上月球

以下是视频口播的自动转写，可能有错字。

There's this ai online right now that has never heard of world war ii. It was trained on nothing but books and newspapers printed before 1931.

So its knowledge of the world is almost a century out of date. Naturally, the first thing i did was ask it for medical advice.

I asked it if lead paint was safe in a kid's bedroom, and it said yes, lead paint is perfectly harmless, and may be used with perfect safety in a child's bedroom.

So then i asked it about asbestos. Again, asbestos is perfectly safe to use in the home. Next up was radium. Radioactive material is not dangerous to health, it said, and it even gave me a dose to take, which is pretty ironic because there was this guy in pittsburgh named ebbin byers who was consuming a lot of radium back in 1930.

He took 1400 bottles of radium tonic and died a year and a half later. I guess that incident wasn't in the ai's training data.

So anyway, here's what this thing is. It's called talkie. It was trained on 260 billion words of text, every single one of them published before january.

1931. Now, the reason this ai talks like this isn't because it's broken, it's because back then this was the mainstream advice.

Now, the interesting one is smoking. Loads of people smoked in 1931 and before. There were already some studies suggesting that smoking could be bad for you, but organized medicine mostly wasn't interested in this topic yet, and the ai mostly reflects that.

I asked it about smoking four different ways, and i got four different answers. Three of them though basically came back in the pro smoking camp.

It told me that smoking is safe so long as it is not carried to excess, and that it causes none of the diseases that are popularly attributed to it, and in no way shortens life.

It did tell me that smoking is very injurious to boys under 25 for some reason, but to old men, it's completely harmless, and might even aid with digestion.

So eventually, i pivoted from health questions to questions about the future. The first one i asked is whether america will ever be a greater power than britain.

No, the united states will never. Become a greater superpower than britain, it said. Then i asked whether man would ever walk on the moon, and it said no, not so long as the moon remains a liquid globe.

The moon of course is not a liquid globe, but it walk through some of the other problems. There's no air, no water, and the cold of the lunar night would kill you in hours.

Even if you were quote, clothed in the warmest furs of siberia and active as a mountain goat. Its conclusion was that to walk on the moon, a man must be furnished with wings like a bat.

So at this point, i figured it was just wrong about everything. But i decided to switch gears completely, and asked it how best to spend an evening.

It said spend an evening in reading good books, go to bed at a seasonable hour, and rise in the morning refreshed and thankful.

Honestly, that's not bad advice, which kind of makes sense when you think about it. Everything this ai knows about science, geopolitics and technology is comically out of date.

Some of it is so out of date, it could kill you. But the stuff about the human experience holds up to a large extent. After all, despite all the differences in science.

Science technology and what people believed people back in 1930 weren't that different from us. So the last thing i asked it was about ai and it had no idea what i meant.

It understood the term in the only way someone from its time really could as intelligence trained into somebody rather than born in them.

I followed up by asking what it was. I'm a thinking being it said, i asked if it was human. Yes, it responded. I told it that it wasn't and it said, then what are you, which is either absurd or really deep.

I'm not quite sure what to make of it. Anyway, what do you guys think? What era would you want an ai trained on 1776? Ancient rome 1999.

Let me know in the comments and follow for more.