This is the Daily Tech News for Tuesday, April 28th, 2026.

We tell you what you need to know, give you the important context, and help each other understand.

And today, a large language model from the year 1930. 30. 30. That's language model

talks like this. Yes. That's right. I'm Jason Howell. I'm Tom Merritt. I can't wait to talk about

this week's big story. Let's get to it.

And I mean, define big, right? Like, it's just one of those days where there was no like premier marquee story. So of course, we're going to talk about a large language model from the

1930s, sort of. It's big. It's really big. Big and big.

It's 13 billion big.

It's Taki, 1930, and it's a new 13 billion open weight LLM, or in this case, it's actually

called a VLM, stands for vintage language model.

It's trained on 260 billion tokens of pre-1931 English

books, newspapers, journals, patents, case law. But the important thing there is that it has a

historical cutoff of December 31st, 1930, which is actually the cutoff for works

to enter the public domain. So, you know, it wasn't just chosen randomly. That basically

means that all of this content that's fed into it is legal to use in this capacity. It's

developed by a nonprofit team that is using the model to look, basically to explore a few key research angles.

The first one is because Taki only saw pre-1931 text, it essentially avoids benchmark contamination and will, as a result, better show how much LMs language models can

actually generalize beyond their training data.

Secondly, by checking how surprised Taki is by the New York Times on this day events.

So they're feeding them these on this day, on this day events that happened in the future after 1930.

Researchers can then measure how well a language model can forecast future outcomes as it moves past or beyond its knowledge cutoff date.

So it's a good way to kind of explore that.

...