@emollick Consistently wild how these papers come out with results so far behind the frontier
Really challenges the standard academic publishing model. Need ways to quickly update results w/ newer models. Gotta get faster..
@cmsholdings good chance we look back on this as one of the things that killed the bull
principles? standards? No one really cares when the bonus checks get this big
Big Short vibes
https://t.co/afxSdTkjhI
@AriDavidPaul We might look back on this kind of thing as the equivalent of NINJA loans or garbage credit ratings from the 2008 financial crisis. Things that in retrospect are obviously bad ideas, but where people are just making too much money to care.
@emollick GPT 5.5 pro is 45x more expensive than Kimi k2.6. Why aren't more people using Kimi if it's almost as good and dramatically cheaper?
GPT 5.0 pro is the same level as k2.6 but 30x more expensive - why would *anyone* use 5.0? Seems like benchmarks must be missing something..
@EpochAIResearch GPT 5.5 pro is 45x more expensive than Kimi k2.6. Why aren't more people using Kimi if it's almost as good and dramatically cheaper?
and GPT 5.0 pro is the same level as k2.6 but 30x more expensive - why would *anyone* use 5.0?
Seems like benchmarks must be missing something..
@BobEUnlimited "Sometimes things really are different" T-Rex and all the other dinosaurs
Maybe that's true this time. Maybe not.
But hard to tell which is which when the future is unknown
@snewmanpv@taalas_inc The foundations of current algorithms were developed in a memory-abundant paradigm. Memory prices have only ripped dramatically higher in the last 6-9mos
Model companies prob haven't re-optimized every aspect of training+inference to adjust for much higher memory prices
@snewmanpv@taalas_inc also, efficiency has *not* been the focus up to now. Most frontier labs are spending the vast majority of their energy on performance. If memory fabs can't keep up and the memory bottleneck tightens, efficiency will become a core priority. Prob lots of low hanging fruit to be had
@snewmanpv If memory becomes the bottleneck, less memory intensive approaches will flourish
@taalas_inc's model on the chip reduces memory requirements to near-zero
New memory-efficient architectures will emerge, out of necessity
@emollick von Neumann would not have considered the IR as a singularity-level event
There are degrees of "not knowing" and the IR changed human life and society in many ways, but not it did not fundamentally change *humanity*
Categorically different than the potential of AGI/ASI
A big pivot from Ken Griffin on AI:
“Number one is, in the last few months, there has been a step change in the productivity of the AI toolkit. It is profoundly more powerful than it was just nine months ago.
And for us at Citadel, that has allowed us to unleash a much broader array of use cases for AI. And it has been really interesting to watch, to be blunt, work that we would usually do with people with masters and PhDs in finance over the course of weeks or months being done by AI agents over the course of hours or days.
These are not these are not mid-tier white collar jobs. These are like extraordinarily high skilled jobs being, I'm going to pick a word, automated by agentic AI. And I gotta tell you, I went home one Friday actually fairly depressed by this because you could just see how this was going to have such a dramatic impact on society.
When you witness it in your own four walls, when you see work that used to be man years of work being done in days or weeks, it's like, wow, like that's the first time I've seen real impact in our four walls.”
This echoes my own experience with agents and the conversations I am having with students, friends & clients. The toolkit has dramatically transformed and it feels like in finance, for the first time, AI is real.
@fejau_inc Bubbly pockets, not a bubble
Some things are overpriced. Core (major chip, memory, model companies) is not overpriced
Models are too powerful and there's too much low hanging fruit* for the core to be a bubble
*to further improve models. and to deploy more widely
@shae_mcl@CanJohnUncu Are we in a regime where it's now cheaper to create these? In terms of $$$? or time?
Do current tools (and tools that are expected in the next 6-18mos) speed/reduce cost? and by how much? How long until we get all of these?
@binarybits@DKThomp ok
but I am curious why the post makes you skeptical (or reinforces your existing skepticism)
bcs that's not my read on it at all
I ask bcs it would be useful to know if there are real reasons why ASI is unlikely (theoretically, in the near/medium term. or at all)
@DKThomp@binarybits Also, at this point I don't think humans have the capacity to formally declare a provable limit to any aspect of Intelligence
At best it is an open empirical question
and discussion around limits is, by necessity, going to be pretty speculative absent strong empirical evidence
@DKThomp@binarybits Yes, why?
Why would any aspect/domain of human intelligence (HI) represent a limit on Intelligence writ large?
Theoretically it feels like intelligence could expand vastly beyond human capacities in every direction of current HI. and in other new non-HI directions too