🔥Telegram İfşa

By design, an "instructed" model is no longer a model of the base training corpus, but rather of an idealized version of it, never clearly defined. It can still score sequences of natural language but what it now models is a vision of what language should be in a specific setup.

A good summary. That GPT-4 was launched as a product when it's at the far edge of LM science is obviously a concern given its transformative potential.

Actually, the largest language models are already replicating, through distillation. Clever girls.

Twitter is Barker now? Or is it Dumber?

"The canary in the coal mine"

Liminality. —We Don’t Know What Will Happen Next nytimes.com/2023/04/18/opi…

"We are totally f&%ked."

“I’m not telling you to make the world better I’m just telling you to live in it. Not just to endure it, not just to suffer it, not just to pass through it, but to live in it. To look at it. To take chances. To make your own work and take pride in it." —Joan Didion

It's a good question, what am I worried about? Well, let me share a scenario. It's April 2026. I wake up in the morning and check Hacker News. "Hundreds of Starlink satellites burn up in the atmosphere."

Is that how we should assess the practical intelligence of a machine: by its ability to make money? What a sad state of affairs!

If you happen to be in Toronto for ACL, you might want to disco on July 10 at 11am virtual2023.aclweb.org/paper_D44.html

Outstanding!

“There is no limit to stupidity. Space itself is said to be bounded by its own curvature, but stupidity continues beyond infinity.” Gene Wolfe

regards croisés

we're made of stories

speaking of stories theguardian.com/commentisfree/…