Love this. Legacy media showing its true colors again, to the detriment of all
A POV re your take on "...and was like having a PhD advisor on any topic". I strongly disagree. I've had a PhD supervisor and been around many. GPT5 is smarter and more useful
There's a question I like to think about and is ~to your throughline, but it isn't very PC: if a model were smarter than me, how could I tell? If it got twice as smart from there, how could I tell?
GPT4 wasn't smarter than me, o3 was debatable, I can't keep up with 5. Neither can my friends.
Some PhD level stuff 5 has helped me with:
- correctly diagnosing a complex lifelong postural/physiotherapeutic issues re a back injury
- correctly interpreting brain MRI scans and blood tests. My neurologist - no joke - just implements theses that 5 reasons out
- discovered a diagnosis in the context of autoimmune disease symptoms (these are extremely complex, notoriously difficult to pin down)
- synthesized new and functional frames in an LLM consciousness project after comprehensive literature review in 4E cogsci
- took a loose mathematical intuition, formalized it into a conjecture, proved it, generalized it, proved it again, found a significant open problem in number theory that it applies to, sketched a proof
- found a nontrivial and original analysis of the economics of the AI supercycle
The whole AI intelligence debate is filled with meaningless mumbo jumbo. What does being twice as smart mean? How many times smarter are you than a dog? These questions are unanswerable.
You may find the satirical paper "On the Impossibility of Super-sized Machines" to be clarifying. See especially argument 2, "Meaninglessness of Human-Sized Largeness" https://arxiv.org/abs/1703.10987
While this paper is somewhat humorous. It is drawing a false analogy. No one debates the fact that men are on average larger than women or that certain ethnic groups are larger than others. Yet we cannot come to similar consensus regarding IQ or any other measure of cognitive capacity. Clearly intelligence is something we don't understand the way we understand physical size.
Maybe true, but there's still some axis we're discussing. Intelligence is not a nothing, the dialog is intelligible. Whatever we mean, GPT5 is smarter than GPT2 and my dog is smarter now than when she was 1 month old. Epistemic nihilism will not help us reason clearly
No need for epistemic nihilism, although epistemic humility is certainly scarse within online AI discourses. Ultimately, I just wish people would use specific and quantifiable concepts when discussing future AI capabilities.
Yup, that lack of humility is awkwardly stifling in its own way
Quantifying is great. Benchmarks seem to saturate fast though, and our capacity to continue to create meaningful new ones is constrained by our own upper bounds. Hence the AGI asymptote, I suppose - detectable gains could saturate holistically without gains doing the same, with obvious superhuman performance still out of reach
I will add, though, that intelligence is massively understood in orthogonal frames that remain scientific, 4E cogsci for example. These do afford a meaningful comparison between me and my dog, ie they're performant where benchmark based quantification isn't. That surprising (to STEM folks, anyways) observation suggests there might be a there there by porting ideas to AI world
The other aspect of all this, is that this recent debate is very focused on LLMs. While it makes sense that they should get a lot of attention, anyone who is paying attention to other areas of generative AI like image and video models knows that they are not slowing down at all.
OpenAI pushed images forward, not in terms of image quality, but a less obvious way. Vastly better instruction following and fairly reliable text. It got the attention for the Ghibli styling, but the actual big deal was how much more useful it was at many tasks. Veo 3 hit with high quality video that generated audio at the same time, with impressive lip-sync compared to previous solutions. nano-banana (almost certainly from Google and likely officially releasing this week) is a big step up in editing existing images (OpenAI's model often changes the entire image while this doesn't) and there's several new tools that do editing on existing videos now.
While all of that is happening in the commercial space, the open weight image/video models have exploded. Like FLUX for images and Wan for video. That has a lot of dangers of its own, especially since the video models support an image as the first frame. And efficiency is another dimension. People now can generate videos with a consumer graphics card.
Not to mention Genie 3 from Google feeling a little like its own ChatGPT moment in terms of interactive world models. Robotics also feels on the cusp of big advances.
Love this. Legacy media showing its true colors again, to the detriment of all
A POV re your take on "...and was like having a PhD advisor on any topic". I strongly disagree. I've had a PhD supervisor and been around many. GPT5 is smarter and more useful
There's a question I like to think about and is ~to your throughline, but it isn't very PC: if a model were smarter than me, how could I tell? If it got twice as smart from there, how could I tell?
To that is the AGI asymptote idea:
https://www.latent.space/p/self-improving?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd1c06950-0bb2-4f66-af8a-b51e3d5f446e_2110x952.png&open=false
GPT4 wasn't smarter than me, o3 was debatable, I can't keep up with 5. Neither can my friends.
Some PhD level stuff 5 has helped me with:
- correctly diagnosing a complex lifelong postural/physiotherapeutic issues re a back injury
- correctly interpreting brain MRI scans and blood tests. My neurologist - no joke - just implements theses that 5 reasons out
- discovered a diagnosis in the context of autoimmune disease symptoms (these are extremely complex, notoriously difficult to pin down)
- synthesized new and functional frames in an LLM consciousness project after comprehensive literature review in 4E cogsci
- took a loose mathematical intuition, formalized it into a conjecture, proved it, generalized it, proved it again, found a significant open problem in number theory that it applies to, sketched a proof
- found a nontrivial and original analysis of the economics of the AI supercycle
The whole AI intelligence debate is filled with meaningless mumbo jumbo. What does being twice as smart mean? How many times smarter are you than a dog? These questions are unanswerable.
You may find the satirical paper "On the Impossibility of Super-sized Machines" to be clarifying. See especially argument 2, "Meaninglessness of Human-Sized Largeness" https://arxiv.org/abs/1703.10987
While this paper is somewhat humorous. It is drawing a false analogy. No one debates the fact that men are on average larger than women or that certain ethnic groups are larger than others. Yet we cannot come to similar consensus regarding IQ or any other measure of cognitive capacity. Clearly intelligence is something we don't understand the way we understand physical size.
Maybe true, but there's still some axis we're discussing. Intelligence is not a nothing, the dialog is intelligible. Whatever we mean, GPT5 is smarter than GPT2 and my dog is smarter now than when she was 1 month old. Epistemic nihilism will not help us reason clearly
No need for epistemic nihilism, although epistemic humility is certainly scarse within online AI discourses. Ultimately, I just wish people would use specific and quantifiable concepts when discussing future AI capabilities.
Yup, that lack of humility is awkwardly stifling in its own way
Quantifying is great. Benchmarks seem to saturate fast though, and our capacity to continue to create meaningful new ones is constrained by our own upper bounds. Hence the AGI asymptote, I suppose - detectable gains could saturate holistically without gains doing the same, with obvious superhuman performance still out of reach
I will add, though, that intelligence is massively understood in orthogonal frames that remain scientific, 4E cogsci for example. These do afford a meaningful comparison between me and my dog, ie they're performant where benchmark based quantification isn't. That surprising (to STEM folks, anyways) observation suggests there might be a there there by porting ideas to AI world
My even hotter take is that even the narrative of "pre-training progress is slowing" is also false, with pretty real consequences.
Say more? Seems like most of the gains since GPT-4 have come from RL on verifiable rewards + inference compute.
My claim is not that current progress is driven by pre-training, it's that the returns to pre-training didn't decline as claimed.
The other aspect of all this, is that this recent debate is very focused on LLMs. While it makes sense that they should get a lot of attention, anyone who is paying attention to other areas of generative AI like image and video models knows that they are not slowing down at all.
OpenAI pushed images forward, not in terms of image quality, but a less obvious way. Vastly better instruction following and fairly reliable text. It got the attention for the Ghibli styling, but the actual big deal was how much more useful it was at many tasks. Veo 3 hit with high quality video that generated audio at the same time, with impressive lip-sync compared to previous solutions. nano-banana (almost certainly from Google and likely officially releasing this week) is a big step up in editing existing images (OpenAI's model often changes the entire image while this doesn't) and there's several new tools that do editing on existing videos now.
While all of that is happening in the commercial space, the open weight image/video models have exploded. Like FLUX for images and Wan for video. That has a lot of dangers of its own, especially since the video models support an image as the first frame. And efficiency is another dimension. People now can generate videos with a consumer graphics card.
Not to mention Genie 3 from Google feeling a little like its own ChatGPT moment in terms of interactive world models. Robotics also feels on the cusp of big advances.