I keep seeing lots of people saying "LLMs are like compilers/assemblers for prompts"

[mcc@mastodon.social]

@mntmn @cwebber And I'd say "maybe the solution is to build a *good* natural language interface to the computer, so people use that instead" but I don't think a culture that believes LLMs are a computer interface (or are an "artificial intelligence"), could build or adopt such a system. If you put it side by side with the LLM the LLM will "win" because it is fail-open. A "good" interface would tell you when it can't do something, and then the user quits using it. An LLM can make something up.

[airtower@woem.men]

@mcc@mastodon.social @mntmn@mastodon.social @cwebber@social.coop Yeah, and a good computer interface needs to be precise and unambiguous. Natural language is notoriously ambiguous, e.g. pilots and air traffic controllers have to train using precise terminology and phrasing, because a misunderstanding can have catastrophic consequences. I highly doubt people who don't feel like learning a programming language would want to learn a similarly (possibly more strictly) formalized variant of their natural language.

[ireneista@adhd.irenes.space]

@mcc @mntmn @cwebber we aren't quite sure where to start in telling this story, so maybe we won't get into detail, but we were shocked to realize that megacorps have no ambitions for voice assistants beyond turning light bulbs on and off. no desire to build a general-purpose UI at all.

[dhobern@scicomm.xyz]

@mcc @mntmn @cwebber

I think there's a broader corollary (or perhaps it's actually a central subset of what you describe).

I always thought most people shared my experience that the exciting thing about the Internet, and good Internet search in particular, was that it offered access to the most relevant sources of information for any query. It was then on me to assess these sources and try to understand the topic at hand.

LLMs have resoundingly demonstrated that for most people this is all too much work and reminds them of school.

A majority of people clearly don't want to have to put in so much effort. They'd rather have an unambiguous answer that comes back and that they can treat as authoritative.

Sidenote - this is why mansplaining is a thing.

So, the primary (and I would argue, intended) result of the current "AI" mania is that the world is happily replumbing all its information and knowledge streams so that everyone receives whatever sanctioned propaganda those behind the curtain want to shovel out. (Pick a metaphor and stick with it ...)

LLMs are an assault on human communication and our ability to reason, organise and plan. They are the oligarch's wet dream.

[joeyh@sunbeam.city]

@cwebber of course a deterministic LLM could be made. But ~noone would use it. Being able to reroll the dice is an important part of the confidence game.

[bri7@social.treehouse.systems]

@mntmn @cwebber management types have wanted this since the 1950s. it’s why COBOL and SQL exist; it’s why RAD exists. It’s why so called “4th Generation Languages” exist. Management would like nothing more to be done with needing to think about all those pesky details like “that’s a logical impossibility” or “that’s P=NP”, they want their word to be the word of god

[rosie@0x4d4f5448.systems]

We love it when changes have non-localized and unpredictable results;

[drwho@masto.hackers.town]

@cstanhope @mcc @mntmn @cwebber I like it.

[mcc@mastodon.social]

@ireneista @mntmn @cwebber well it's a general purpose UI *now* but only in a very monkeys paw way

[cwebber@social.coop]

@joeyh I mean real talk that's why I don't play preset seeds in roguelikes, hooked on that RNG juice

[dryak@mstdn.science]

@mcc @mntmn @cwebber speaking of expanding to more users and of assembler:

An argument I've heard is that: in the past high level compiled languages have replaced assembler, and LLMs are the next step.

Well, assembler -- and assembler-adjacent stuff like C's SIMD intrinsics -- are still relied upon (think finely optimised low-lvl libraries in some fields like gaming, video codecs, and number crunching in scientific data analysis).
...

[dryak@mstdn.science]

@mcc @mntmn @cwebber ...
Proper software engineering done by human with brains that can hold an actual mental model of the design implication isn't going away.

At best, what we have is a (very unreliable) tool that will allow people who are utterly allergic to programming languages to still be able to play around code.

[smn@l3ib.org]

@cwebber they're lossy pseudorandom decompression

[octorine@fosstodon.org]

@cstanhope @drwho @mcc @mntmn @cwebber And to bring it full circle, grad students *can* be compilers.

[octorine@fosstodon.org]

@mntmn @cwebber My company is 100% invested in ai. It's all management talks about. Before LLMs, we were all in on no-code or low code languages, web robots and such.

It's basically the same fantasy as before, but this time the whole world is along for the ride.

[alina@girldick.gay]

@cwebber @joeyh the binding of isaac, enter the gungeon and dead cells are worse than a slot machine for my adhd brain

[mirabilos@toot.mirbsd.org]

@cwebber oh, they could… if you operated them yourself. Snapshotting, and saving the PRNG seed.

[rdviii@famichiki.jp]

@cwebber mostly agree, especially about them not being compilers, but some compilers aren't deterministic. You'll get a different result in memory layout or optimization sometimes. Especially for quantum compilers, where the compilation process itself is known to be NP hard, so heuristics are used.

[ckfinite@universeodon.com]

@cwebber I think that the lack of determinism is less important than the lack of a precise mapping between input and output. If we hypothesize a nondeterministic compiler that exposes a precise semantics, you can statistically predict the output based on the input based on the stochastic compilation process. LLMs don't give any such guarantee: you put imprecise natural language in and then the output is vibesy related to the input. Small changes to the input can beget very large changes to the output.

[kkarhan@infosec.space]

@eramdam @cwebber +1