Users are excited that superintelligent AI will favor abstractions and tooling like Lisp to create breathtaking towers of abstraction beyond human capability.
Based on 1 visible X reactions from 8 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@tenobrus it will use Lisp and write breathtaking towers of abstraction beyond all human ken
not picking on thebes here since it seems to be held pretty loosely, but i've seen arguments like this a lot and they've never made much sense to me. "superintelligence will be smarter than any human and make fewer mistakes, so it will be able to write untyped python without bugs" like sure maybe it will be *more capable* of doing so, but will that actually be the optimal use of its time and tokens? empirically we don't find smarter and smarter people regressing back to writing assembly because they're maybe more capable of writing it bug free. rather they continue to write and use new and more abstractions that allow them to become even more efficient in the future. also the ceiling on memory safety seems pretty goddamn high, given that it remains the number 1 issue with eg C++ even in environments like Google with hordes of language experts and custom tooling. it seems quite plausible that the correct answer for a superintelligence is in fact to just build tooling that makes the problem impossible.
like, in humans, there's inherent limitations to eg how much you can see at a time no matter how smart you are. so tools make sense. but for models there's no point to building tooling to make handling names easier for models, because models are already so good at it. you could imagine some constrained generation logit thing that filters names to just names that are valid in the current context, and maybe some programming language features around scoping to make this easier - but you just don't need to build that anymore. it dissolved into model capability a long time ago, it's below the waterline. i'm not sure if memory errors are destined to go this way, but it seems plausible... of course models may develop better languages for other reasons, anyways, like to attempt more ambitious projects. but "make zig!bun memory error free without porting" just seems like something that'd be pretty tractable for future powerful models.
@tenobrus i mean, models already ~never make the kinds of eg name errors that come from humans transposing a_b➡️b_a. it seems plausible to me that eg n paths to free!= 1 bugs are artifacts of one-screen-at-time coding that dissolve in 100m context but also many are saying... https://twitter.com/voooooogel/status/2075366422235271467
Thebes counters that 100-million-token contexts will eliminate memory.
Users are excited that superintelligent AI will favor abstractions and tooling like Lisp to create breathtaking towers of abstraction beyond human capability.
Based on 1 visible X reactions from 8 accounts; directional sample.
Ask a question below.
Published answers will appear here.