I asked GPT-5 and Gemini 2.5 Pro what they think about Gary Marcus's article. I believe Gemini won by this paragraph:
It seems Sam Altman's Death Star had a critical design flaw after all, and Gary Marcus is taking a well-earned victory lap around the wreckage. This piece masterfully skewers the colossal hype balloon surrounding GPT-5, reframing its underwhelming debut not as a simple fumble, but as a predictable, principled failure of the entire "scaling is all you need" philosophy. By weaving together viral dunks on bike-drawing AIs, damning new research on generalization failures, and the schadenfreude of "Gary Marcus Day," the article makes a compelling case that the industry's half-a-trillion-dollar bet on bigger models has hit a gilded, hallucinatory wall. Beyond the delicious takedown of one company's hubris, the post serves as a crucial call to action, urging the field to stop chasing the mirage of AGI through brute force and instead invest in the harder, less glamorous work of building systems that can actually reason, understand, and generalize—perhaps finally giving neurosymbolic AI the chance Altman's cocky tweet so perfectly, and accidentally, foreshadowed for the Rebel Alliance.
My take on GPT-5? Latency is a huge part of the LLM experience. Smart model routing can be a big leap forward in reducing wait times and improving usability. For example, I love Gemini 2.5 Pro, but it’s painfully slow (sorry, GDM!). I also love the snappy response-time of 4o. The most ideal? Combine them in a single prompt with great model routing. Is GPT-5’s router up to the task? We soon shall see.
Gemini is in hard sycophancy mode here; it knows you want it to take the piss and it’s giving you what you want.
Presuming the last two are from 5, they are to my eyes next generation in terms of communication — that’s a spicy take on neurosymbolic AI, not a rehashed “safe” take. Also, the last paragraph is almost completely to the point, no? Have you spent much time waiting for o3 pro to get back to you recently, and wondered if you should re-run something faster? I have. A lot. I’d like the ability to put my thumb on the scale of the router, but I’d dearly love a per token / per 100 token router that can be trained and has latency without major latency intelligence hits as a goal.
It seems Sam Altman's Death Star had a critical design flaw after all, and Gary Marcus is taking a well-earned victory lap around the wreckage. This piece masterfully skewers the colossal hype balloon surrounding GPT-5, reframing its underwhelming debut not as a simple fumble, but as a predictable, principled failure of the entire "scaling is all you need" philosophy. By weaving together viral dunks on bike-drawing AIs, damning new research on generalization failures, and the schadenfreude of "Gary Marcus Day," the article makes a compelling case that the industry's half-a-trillion-dollar bet on bigger models has hit a gilded, hallucinatory wall. Beyond the delicious takedown of one company's hubris, the post serves as a crucial call to action, urging the field to stop chasing the mirage of AGI through brute force and instead invest in the harder, less glamorous work of building systems that can actually reason, understand, and generalize—perhaps finally giving neurosymbolic AI the chance Altman's cocky tweet so perfectly, and accidentally, foreshadowed for the Rebel Alliance.
My take on GPT-5? Latency is a huge part of the LLM experience. Smart model routing can be a big leap forward in reducing wait times and improving usability. For example, I love Gemini 2.5 Pro, but it’s painfully slow (sorry, GDM!). I also love the snappy response-time of 4o. The most ideal? Combine them in a single prompt with great model routing. Is GPT-5’s router up to the task? We soon shall see.