Gary Marcus

Rank 17 of 47
|
Score 65
2/15/2025, 4:49:39 PM
X
In reply to:
NLeseul
@NLeseul
·
281d
@GaryMarcus @Zjoue2 And it did better than I expected with the bicycle one, even. (Although it forgot to actually draw the seat and handlebars.) https://chatgpt.com/share/67b0c4…
NLeseul
@NLeseul
·
281d
@GaryMarcus @Zjoue2 The geometric shapes one was fine too: https://chatgpt.com/share/67b0c4…
Gary Marcus
@GaryMarcus
·
281d
@NLeseul @Zjoue2 Did you try these as well?
Gary Marcus
@GaryMarcus
·
281d
ChatGPT is the ultimate “personality hire”.
NLeseul
@NLeseul
·
281d
@GaryMarcus @Zjoue2 Asking for SVG output seems to work better than image generation in at least this case. A lot of this is just limitations of the image generators they send these tasks to, rather than limitations of the main text model.

https://chatgpt.com/share/67b0bf…
Gary Marcus
@GaryMarcus
·
282d
@Zjoue2 ASI does not currently exist and Grok 2’s scene comprehension of its own output is *often* spectacularly bad, across a wide range of examples like these.
Zjoue
@Zjoue2
·
282d
@GaryMarcus Let’s steel man as well as straw man please, this is a little misleading for those who might believe ASI currently exists and this is all the latest AI can do. In general very supportive of your critiques and voice in this field.

The statement is a technical inquiry about scene comprehension in AI systems, specifically regarding image labeling capabilities. It does not engage in public discourse as it focuses on a specific technical issue rather than broader societal concerns or debates.