Gary Marcus

you said it was wrong to say that there were sensors. it wasn’t. you also bait and switch: my objection wasn’t purely to title; it was to framing of blog, in screenshot. suggestion was that better title could have reduced misunderstanding, not that original title was false.

As I said, they did not just use one object -- they used five progressively more general cubes over the course of the project. Not all of them have sensors. If they had hidden the limitations in the paper, I'd agree with adjusting title/abstract. But it's all in the blog post.

yes i know about the five variants on a cube; it’s still essentially one object. i responded to this elsewhere. since you still have neither addressed my key points nor apologized for inaccurately claiming i made an error on the sensors, we are done.

as postscript watch the openai video: NOTHING that i say on my slide is mentioned there. no mention of sensors; no mention of dropped cubes; no mention of innate components. as a future PhD, do you think that is ok?

That video is a nontechnical summary of the work. The discussion of limitations is exactly where I would expect it -- in the main body of the blog post.

sensors don’t appear in the main part of the blog; they are only in the table. and NOTHING in the blog says that performance drops to ZERO on fully scrambled cube when you remove the sensor. do correct me with a quote if I am wrong, or correct your inaccurate claims if I am not

I read the paper first. While they could afford to add a line to the blog clarifying which cube was used for the main result, your original statement was substantially more misleading than this omission.

Let's make a deal: I'll delete and revise my initial tweet once you revise your slide to make explicit that the blog clearly presents all limitations, excepting the use of sensors, which is currently left to the paper.