AutoTune for Writing

1st up, the always reliable Gary Marcus:

Two critical updates re: Astra and mathematics

So OpenAI announces a new model has solved 10 hard math problems. Then within a day an Antropic researcher (Alpöge) is able to reproduce 1/2 the results using an older public model.

Alpöge’s fast partial replication made me wonder whether OpenAI’s real advance here was finding (through AI) which open problems were amenable to a certain kind of search and verify technique.

So now in addition to continuously trying to 1-up each other with miraculous claims, the 2 leading AI companies are now starting to play defense as well. Yawn.

---

From Doc Searls' Weblog, we get the title of my post, which will go in my table of names for LLMs.

Throbsday - On AI wronging and writing

I hate AI writing and can spot it in a second. It’s not just the excess of one-line paragraphs, the contrasty “not that, but this” argument constructions, the excessive use of em dashes. It’s a tone that sounds like the average of a million well-put hunks of text published by competent writers of opinion columns and advertising copy. In other words, a lot like my own writing.

AI writing is AutoTune for the human kind.

Doc Searls has written many times about his use of GenAI, so I was kind of surprised at the negative attitude here. But he makes excellent points: it can be a tool, but please no slop.

Note, lots of musicians totally hate AutoTune.

---

From the King of Snark, sci-fi author John Scalzi. He joins the group of authors becoming ever more outspoken about AI slop. Hunh, I guess this goes in the table of names as well.

Generative “AI”: The Guitar Hero of Creativity

Yah, you can generate stuff that almost seems decent. But you aren't developing any real skill or, I think more tellingly, any real style.

I remember when my kids would play Guitar Hero & I'm totally like "Why not actually learn to play guitar?" (I think all of my kids do play instruments.)

---

From Programmable Mutter by Henry Farrell:

The downside of robot solutionism

Mapping the Abelard Snazz feedback loop

Apparently a comic series by Alan Moore in the early 1980s is eerily prescient re the many arms races that seem to define GenAI - using GenAI to detect GenAI slop, cheating, plagiarism, etc. Spoiler alert - the human race winds up being forced to flee the planet - oops!

---

Here's the home page/directory for my posts on Bullshit. This is post #140.