Henry Perschk

Digital Physics, Artificial Life, Open-Endedness, Evolutionary Algorithms. Transformers are a crutch.

[Henry Perschk @henryperschk]

@devagrawal09 guys the relevant theorem here is (mostly not) Bitter Lesson but No Free Lunch

[Henry Perschk @henryperschk]

where's the regress in this? yes it's unknown how (or if) you could decompose forces themselves. but let's assume these as axiomatic for now. then surely it's completely fine to explain "touch" using this model? you can't say "because we don't understand the origins of this axiom, it's function cannot be true" - especially when you can actually observe its function using electron microscopy.

[Henry Perschk @henryperschk]

No, NN's do only interpolation, albeit in a extremely high-dimensional space, which can make it *look* like extrapolation. It's only when you add a symbolic scratchpad (a "harness") that you can do real extrapolation.

At this point in time the symbols that we write on this scratchpad are still very minimally outside of the convex hull of the search space, which is why we still see rather little creativity. This is not to say that it's impossible to get more creativity out of it though - current methods are just rather ineffective.

[Henry Perschk @henryperschk]

@matthiasmurdy @JackFeynman furthermore, it explains nothing about WHY we feel this way. what is the reason (the mechanism) for the fact that we can feel at all?

and even if we were to find the exact atom configuration that produces consciousness, it would STILL not explain why that leads to a felt feeling

[Henry Perschk @henryperschk]

@matthiasmurdy @JackFeynman that only proves it for yourself, not for others. it's a subjective truth only.

[Henry Perschk @henryperschk]

@MattZirwas @johnennis THIS is what a cell looks like. It’s not a fucking some over a few scalars (“synapses”) passed through a non-linearity. It’s something completely different. Calling it an analogy is a really hard stretch.

https://t.co/anCL2urXIp

[Henry Perschk @henryperschk]

@shisoftgenius And then when you realize that THIS is what a single cell looks like on the inside, you see that we are very, very, very, very, very, very far from simulating a brain.

[Henry Perschk @henryperschk]

@MattZirwas [ ] I know what I am talking about.

[Henry Perschk @henryperschk]

@hraness @ChatGPT @DevinAI @claudeai @cursor_ai @bot @zeddotdev why terminals though? so much less productive and harder to manage than the GUI options we have now.

[Henry Perschk @henryperschk]

@hypergameme Even though I agree emotionally, I don't think there is definitive scientific argument that removes functionalism from the menu. If there is, please tell

[Henry Perschk @henryperschk]

@gakonst Why would you use a TUI in the first place?

[Henry Perschk @henryperschk]

@tunguz @grok remind me in 6 months

[Henry Perschk @henryperschk]

Probably nothing

[Guillermo Rauch @rauchg]

Looks like today may be a record day for token volume % of open models on Vercel AI Gateway:

🟦 Open 78.4% 🟨 Closed 21.6%

While spend 💲 usually tells a different story, #3 and #4 today are Moonshot AI & DeepSeek. Adding Z⁠.ai, their combined spend surpasses OpenAI (#2).

(Do note that's the spend for inference of the model across providers (mostly in the US), not revenue going directly to the open weight labs.)

[Henry Perschk @henryperschk]

You could extend Jev to give some sort of debugging logs. But that’s besides the point.

A spam email filter also cannot reasonably be examined for its decisions. This is because the *input* is fundamentally unknown/stochastic. If you want to solve deterministic problems using Jev, you’re using it wrong.

[Henry Perschk @henryperschk]

@EliasLumer @kushbhuwalka @typesafeai It is not. Tested it but was disappointed. However, I have no doubt that we can make it arbitrarily good via scaling (well along the s-curve of improvement).

[Henry Perschk @henryperschk]

@DavidDeutschOxf How can we be so sure that functionalism is wrong?

My hunch is also that it is, but we have no reason to reject it with scientific rigor because we don’t know if it’s function or substrate that makes consciousness work in humans.

[Joscha Bach @Plinz]

Anthropic and OpenAI cannot become profitable unless they can reduce the ratio of training cost to inference profits. In the long run, their survival may depend on outlawing competitive open source models. This can only be done under the guise of safety regulation.

[David Bellamy @DavidRBellamy]

I think this is a very difficult pill for many scientists and engineers to swallow: intelligence is not the biggest bottleneck in most of the world’s problems.

[ali @aliteracy]

the real bitter lesson is that none of what makes modern LLMs work so well is at all aligned with the bitter lesson

[Henry Perschk @henryperschk]

@Armv7lFx Clojure is miles ahead on fast feedback, since every single function can be be tested as part of the larger program the instant you modify it.

[Henry Perschk @henryperschk]

Allow me to challenge that. Imagine a GUI that does everything exactly the same way as you currently do with multiple terminals. So it’s just some thin wrapper. It uses Agent Client Protocol to render things nicely with native components instead of rebuilding them with ASCII art (so to say).

I just don’t see why in principle you can’t manage 9 machines in a simple GUI. Even if that just mirrors functionality offered by Herdr, cmux or what have you with real UI components.

In fact, this is what tools like bb or paseo already offer.

Maybe I am missing something? I am only handling 3 machines and this works fine for me. So I am also replying here to potentially gain some insight into upping my game.