In your reading, what is the distinction between blindsight’s scrambler and real intelligence? My reading is that it’s just as real, and draws out the disadvantages a sense of self constrains intelligence with
Sure. I should have been more precise about what is 'real intelligence' here.
What I mean is that blindsight's scramblers are aliens that cannot share human values. Their structure is completely different to ours, their qualia (or whether they even have it) is impossible for us to understand. In short, they do not have a soul. When Claude does this "slowly revealing a dramatic insight" thing that it does, it does that not because it has judged itself through some introspection as having an insight to share. It does not even know what an insight is or is not. It is not sharing anything, because it is not capable of sharing, because it does not have a soul.
The aesthetic structure of its replies is a pattern, a constraint on the token distribution, like the color of noise.
It's my bad to use the word 'intelligence' because it's so overloaded. Will Claude will act as a therapist or produce value or produce a work of art? No. It cannot, because it does not have a soul. I leave it freely open to interpretation whether having a soul is required for "real intelligence." But what I've noticed is that "intelligence" in these discussions is mostly used to denote some capability to produce [economic/social] value. In my mind value is a relational thing, a thing of human feeling.
I think I understand what you are trying to convey but I fear you've made the same mistake again, this time with "soul" instead of "intelligence."
I think what you are getting at is that they are deterministic automata. They are machines. We have introduced randomness to add variation but it is an artificial randomness that simply perturbs the path traversed.
When we choose words it isn't because of a token distribution, nor because we rolled a die. We choose words because we feel a certain way, the external world, our body and senses are all connected as one system. These machines don't experience moods or get tired or feel better after a good night's sleep. They don't know their audience, we're all the same to them. We have no personal relationship nor can we establish one, as presenting some arbitrary background is not the same thing as a fluid, evolving relationship that accumulates through experience over time. There are no scars or fond memories.
If these things can truly be intelligent, to abuse your use of the word, then at least we are quite far from holding them correctly.
What I am trying to get at is when people talk about "intelligence," beyond academic debates such as this one that mentions qualia, they are talking primarily about value. How intelligent the thing is is how valuable it is (often directly, in an economic sense). My argument is that value is a relational thing that must involves human feeling, intelligence actually has very little to do with it.
When a child slowly reveals to me something he's judged insightful about pokemon, it's valuable to me, even though I'm not learning new facts. When Claude does the same thing, it's not even "not valuable," it's wholly outside of value, even if I do not already know that fact or thing. If ever Claude reveals to me an insight, that insight certainly came from the training data, not from Claude. The soul came from a human being external to Claude -- the training data -- and passed transparently through Claude. Claude is a translator: it feels nothing, adds nothing. It colors the noise. I know this because the facts are the same but the aesthetic structure - how many bullet points, whether it says "delve" or "load-bearing" does change over time, according to the whims of whoever is in charge of post-training.
But the "real intelligence," is the same thing as "real value." It is the soul of the humans in the training data, the books, newspaper articles, etc.
A clever hacker news commenter might argue, well, what if we gave Claude a humanoid body and senses and let it interact with the world, then would it be intelligent? To that I would say, why waste your time and effort? You can get it for free: just talk to a friend, a neighbor, or a family member.
I think you're conflating intelligence with consciousness. The point of Blindsight is to lead us to the idea that intelligence doesn't require consciousness (and further, that consciousness might even act as a limiter on intelligence).
I agree, and yet it is reasonable to ask if these notions of "context" (sorry!) - namely feelings, external world, body, senses, etc - are somehow distinct from an LLM's notions of context. Today they certainly capture different things, but given the right representations, why couldn't these human notions also be captured as "context"?
The idea of memory does not seem to resolve this: if you allow the machine to "compact" its context, then you've given it a system which is analogous to our own evolving state. (Though this is undoubtedly still less expressive and meaningful than the one we have evolved as humans.)
One idea I've wondered about is our human capacity to induce subsequent mental states: I can effectively decide how I want to feel and take actions to create that feeling. It's not clear whether models exhibit any degree of privileged introspection into their own states. Is this important? I don't know. (Non)determinism also does not seem to resolve it; it's my understanding that there are plenty of philosophers and researchers who think that human behavior is deterministic, or that the question of determinism does not matter.
Until now, human knowledge and values have built on prior human knowledge and experience. If AI is able to develop without human influence, I believe its value system will necessarily diverge into something alien.
Will Claude will act as a therapist or produce
value or produce a work of art? No. It cannot,
because it does not have a soul.
I tentatively agree, although I'm only tentative because I don't think it's an interesting question.
Here's what I do think is interesting. You!
I mean... yes, you, too but not you specifically. The plural "you" that the english language lacks.
And so here's what I think is the actual interesting question. Might AI help you create art? Or be a therapist? Or something else interesting and worthwhile?
Maybe AI won't write the next great guitar solo. I'm pretty sure it won't. But might it help you learn to play guitar? Help you fix your broken guitar amp? Help you understand some tricky parts of guitar playing? Help you work through some tricky tabulature where you can't tell if you're playing it wrong or if the tab is just bad?
I don't know. But that's my angle for finding any of this interesting.
I think that Blindsight's scramblers were intelligent but not conscious, which for us is very difficult to understand. For them consciousness was a blight.
Wasip1 did not specify sockets. Some implementations have made non-standard additions to add them, but sockets were not added to the standard until wasip2.
No, the stack switching proposal is not used. Stack switching is a set of core Wasm opcodes that permit a guest to change its own stack. Instead of using opcodes inside the Wasm, the stackful async mode of the component model’s ABI calls out into the component model implementation where it can manipulate the stacks with special host powers - JSPI is sufficient on web engines to express this, and wasmtime manages guest stacks in memory.
When stack switching becomes available in all engines, it will be possible to implement this part of the component model in pure Wasm without host magic, e.g. web engines will be able to avoid the call out to JS to use JSPI.
Wasmtime implements a remote debugging server, so that you can debug guest programs with a recent build of LLDB. Set breakpoints based on the source language symbols, single-step through wasm opcodes, anything you'd expect: https://docs.wasmtime.dev/examples-debugging-guest.html
still useless. I had an idea to offload parts of my game core into wasm leaving only UI in the browser - I abandoned that idea when there was absolutely zero possibility for me to debug in browser what is wrong with my core. i just rewrote everything to TS (from rust)
There is a wat (s expression) syntax for the component model. The problem with the wat syntax for both Wasm and the component model is they’re a reflection of a binary format, and therefore are terrible for writing by hand. They’re designed to be written by tools, and the text format is just to help you understand the binary format.
I’ve written Wasm and component model wat extensively over the last decade to develop tests for Wasmtime, and even for an expert it’s a bad experience.
Wit syntax is easy to read and write by hand. There are high quality parsers that can transform it to and from the binary or wat format as needed, and code generators for a wide range of languages. It’s a way, way better experience in every way to deal with wit compared to the wat format.
Seconding this. Writing .wast files by hand is kind of fun as a novelty, but it's in no way something I'd want to maintain interfaces in.
Declarative interfaces via the WIT format are so much more readable in comparison; I'm really happy that's the direction we took for the component model.
the "redline" compiler is cranelift which is written in rust, but i think it's being compiled to Wasm first then JVM bytecode so it still works zero dependency. not sure if that will continue to be the case.
WebAssembly is a compiler target, not a human-authored language. There is exactly one audience of people for writing wat by hand: spec and tutorial authors and readers. Anyone actually developing an application they want to use will use a compiler to produce WebAssembly. Prove me wrong and write Roller Coaster Tycoon in raw wasm if you want, but having written and maintained wasm specs and toolchains for nearly a decade, I will never write any wat outside of a spec or tutorial.
There is exactly one case where I'd like to write "raw wat" (and for that matter "raw wasm bytecode"): I'd love to do something like the "bootstrappable builds" project for wasm, starting with a simple wat-to-bytecode parser/translator written in raw bytecode, then some tools wirtten in raw wat for bootstrapping into other languages. :)
Real programs, whether native JavaScript or in any other language that targets Wasm, have concurrency. Would you rather the component model exclude all concurrent programs, and fail to interact with concurrent JavaScript? The component model is meeting the web and programmers where they're at. Unless you're one of the few people implementing the low level bindings between components and guest or host languages, you don't have to ever read the CM spec or care about the minutae of how it gets implemented.
I was looking there because the high-level documentation doesn't seem detailed enough to understand how it works. Maybe some intermediate-level explanations are needed?
Wasm only gets additive changes - the binary format can't change in a way that breaks any previously existing programs, because that would break the Web. So, you just have to add more opcodes to your implementation.
