Yes, I am complaining that they are making an impossibility claim on the basis of an observational gap. Such claims don't have a great track record in the history of science.

> negative, which is literally impossible.

Impossibility proofs are common in mathematics, physics, and computer science. This paper is not one of them. It reports an observational gap. That's not the same thing at all as showing, e.g. that any transformer no matter how large or interconnected, can't compute some function.

> our brains

Airliners don't have feathers.

> we clearly haven’t really seen major architectural changes for transformers for a few years now.

Ever read a DeepSeek paper? Ever hear of MLA? Mamba? Or gated deltanet? Or RLMs? Universal transformers? There's been a deluge of architectural advancement over the past few years. You shouldn't go around asserting the burden of proof falls on this or that party if you're not familiar enough with the recent literature to recognize the kinds of proof that would satisfy this burden.

> deficiencies will remain even if we figure out ways to paper over it on a case by case basis.

I think there are general solutions unknown to us for classes of problem we solve one by one through brute force today. Not arguing that. I just don't accept that the path to generality goes through giving up "transformers", whatever this term means after the architectural Cambrian explosion of the past few years.

It's much more likely that further capability unlocks involve in-the-weights continuous online learning. How we do that is orthogonal to whether the weights encode a transformer, a diffusion model, a SSM, or something more exotic.

Sure, these things aren't pure transformers. But neither are frontier models. The industry is already doing what you suggest and moving beyond naive KQ dot product full depth everywhere 2010s-era transformers. Architectural innovation hasn't solved the problem. Turns out different architectures for approximating functions all form function approximators. The problem is in formulating the functions we want to approximate, not our spelling of the approximation engine!

> Ever read a DeepSeek paper? Ever hear of MLA? Mamba? Or gated deltanet? Or RLMs? Universal transformers?

Quite a few of those aren’t transformer architectures, MLA is more of KV optimization that doesn’t degrade intelligence than something that directly improves intelligences. Indirectly it lets you run a larger model on the same hardware but that’s it. It’s also 2 years old while universal transformers are 8 years old and only MLA has seen adoption. Your reply was full of gish gallop nonsense that argues against anything really new in transformers capabilities with intelligence.

> Not arguing that. I just don't accept that the path to generality goes through giving up "transformers", whatever this term means after the architectural Cambrian explosion of the past few years.

I mean dLLMs are quite architecturally different from plain LLM transformers that you get on OpenAI or Anthropic today even if the use transformers if you squint at them - they’re bidirectional thinking and embarrassingly parallel. Why would the next explosion not be architecturally different from the previous one? Indeed you’d expect a difference because anything that can overcome today’s transformers has to be exponentially better and anything based around transformers won’t be and there’s clearly still a few orders of magnitude between humans and LLMs.

> Sure, these things aren't pure transformers. But neither are frontier models. The industry is already doing what you suggest and moving beyond naive KQ dot product full depth everywhere 2010s-era transformers.

But they’re not, not really. The difference between Llama 3.2 and Claude Fabel architecturally is relatively small, with most of the gains coming from RL, training data, size, training systems, and inference loop infrastructure. It’s all clearly made a huge difference but structurally there hasn’t been huge structural changes; most of the structural changes are around inference efficiency and trying to optimize performance without sacrificing intelligence. At some point you’ll run out of headroom of how far you can take that and that point will be a far way away from AGI.

And chemistry could stop conserving energy tomorrow!

But what does the preponderance of observed evidence tell us is likely the case? For both conserved energy and transformer behavior.

You’re just outlining the problem of induction, but that doesn’t move the conversation forward, because the person clearly already understood that point and was (much like energy conservation) inductively proposing a rule.