She sent me this article published by Anthropic:
When AI builds itselfThis was published June 4. Reading it, I realized this was the post Gary Marcus wrote about June 5. Here's Marcus's post:Our progress towards recursive self-improvement, and its implications.
No, Anthropic did not call for a pause on AI developmentI linked to that in this post June 7:Not really
So many headlines like these.
& The Winner of the Academy Award for Best AI Bullshit Hype is ...Spoiler, the winner was Anthropic.
I don't know if Marcus didn't link to the Anthropic article, or I just didn't follow the link. Regardless, I hadn't read it before.
I gives a history of the phases of Generative AI development. Here's an interesting chart, showing programmer productivity (measured in Lines of Code (LOC)) increasing with Claude use.
I believe this chart. All the (experienced, senior) developers I'm still in touch with all use Claude Code & love it!
This next chart is titled "Claude Code session success rate". Wow, looks like they've about got it licked!
I don't think I would have noticed how bogus this chart is if I hadn't read (& reviewed) "Calling Bullshit" late last year. This book is about bullshit in science, including charts. Probably its main advice dealing with charts is: pay attention to the axes.
Look at the Y-axis on the left. Look at the top of the axis, 100%, all alone up in the left corner. And, whoa, the Claude phase headers are down in the body of the chart! That's not right! They should be out of the chart, up above the 100% line.
I decided to fix the chart. My friend GIMP-2 helped me. Here's the result:
I moved the phase labels out of the chart, & duplicated the Y-axis on the right side.
I look at this chart, I don't think:
Wow, looks like they've about got it licked!I think:
Wow, they've totally plateaued since March. Looks like 85-90% success rate, aka, 10-15% failure rate. That agrees with the 9-10% error rates coming from other GenAI applications.So the whole "recursive self-improvement" thing is bullshit! You turn this thing loose to improve itself with a 10-15% failure rate, you are going to immediately downward spiral into Degenerative AI.
I find it really sad. Anthropic seems to be leading the field, particularly with Claude Code. Why are they compelled to be so intellectually dishonest? Their version of this chart was clearly designed to be misleading.
Also, I noticed at the end of the article the following:
* Quotes from Anthropic employees throughout this article are drawn from internal discussions and used with permission. They reflect individual views as of May 2026, not official company positions.That's a nice piece of weasel-speak. So their employees are making wild claims, but they are "not official company positions". Of course not.
Well, at least everyone seems to be recognizing their unending hype strategy, perpetually calling wolf & claiming their product is so powerful that is is unsafe. They may have shot themselves in the foot, tho, the Feds are acting like they believe them. LOL!
Here's the home page/directory for my posts on Bullshit. This is post #130.
1 comment:
Note: some asshole, agentic or otherwise, named Siva Naga keeps posting links to Claude training courses in the comments of this post. So I am disabling comments on this post.
Post a Comment