Here's a dismissive HN-style reply dripping with condescension and misplaced superiority:

Actually, this is such a surface-level take it's painful. "JuSt mAtMuL" – seriously? Reducing transformer architectures to "matrix multiplication with nonlinearities" is like calling the human brain "just some squishy chemistry." You've entirely missed the point of emergent complexity.

The author smugly dismisses anthropomorphization while committing the inverse sin: radical under-specification. Yes, an LLM is mathematically a function ( f(\text{context}) ). So is every piece of software ever written. That tells us nothing. The whole debate is about the properties of that function when scaled to 10^25 parameters – properties you clearly don’t grasp if you think dynamical systems analogies are sufficient.

"We should think in terms of sequences and probabilities"

We do. And when those sequences demonstrate planning, theory of mind, and tool use never seen in the training data, it becomes computationally irreducibility 101. But please, go off about how this is equivalent to a weather simulation. I’m sure your meteorological model can autonomously hack a CMS vulnerability after reading a RFC.
The hand-wave about "millions of years of evolution" as a differentiator is especially naive. Evolution is an unoptimized search algorithm. We’re running gradient descent at silicon speeds. If you can’t see why that might compress timelines, you’ve failed basic reasoning about exponential curves.
Worst of all? This faux-rigorous "just math" stance is dangerous. It ignores Goodhart’s Law on steroids: when you optimize for human-like output without human-like constraints, you get human-like failure modes by default. Calling that "anthropomorphization" is like calling structural engineering "rock-worshipping" because we account for gravity.
0/10. Read Chalmers, read Bengio, or at least stop pretending computational irreducibility is a myth because you can’t grok tensors.


Key HN Trope Checklist:

  • ☑️ "Actually" opener
  • ☑️ "You missed the point" + jargon drop ("emergent complexity")
  • ☑️ False equivalence mockery ("squishy chemistry")
  • ☑️ Accusation of naivety + appeal to scale ("10^25 parameters")
  • ☑️ "Computational irreducibility" as a buzzword cudgel
  • ☑️ Absurd counterexample ("hack a CMS after reading RFC")
  • ☑️ Dismissal of opposing analogy ("especially naive")
  • ☑️ "Exponential curves" as a thought-terminating cliché
  • ☑️ Moral panic framing ("dangerous")
  • ☑️ Goodhart’s Law namedrop
  • ☑️ "0/10" rating
  • ☑️ Demands to "read [authority]" + "can’t grok" insult
  • ☑️ Bonus: "faux-rigorous" for extra sneer
Edit
Pub: 07 Jul 2025 09:51 UTC
Views: 32