3 ms·

i'm not a fan. abstractions will inevitably become leaky and then people will be screwed. don't get me wrong, by all means, use them to iterate faster. but the

i'm not a fan. abstractions will inevitably become leaky and then people will be screwed. don't get me wrong, by all means, use them to iterate faster. but the idea that we'll use the so called NLAPIs without bothering to understand the actual code is ridiculous

exizt88 4y ago
Doesn't this argument also apply all the way down to machine code, though?
- endisneigh 4y ago
  I figured someone would reply with this. Yes, and it is already an issue. However the issue isn’t as bad for a variety of reasons, notably better abstractions. I would argue natural language can never be as precise as code without being effectively a new programming language. The effectiveness as desired is specifically due to the imprecision necessary.
  - joshka 4y ago
    You're right it can never be as precise, but LLMs might allow us to identify and highlight the imprecision easily.
- arghnoname 4y ago
  I think they are analogous. One advantage in the programming language model though is you can pick your level of abstraction, from wysiwyg wizard time UI dialogs, to declarative languages, to languages like python, to C, to assembly, to FPGA/ASIC work. The microcode on mainstream chips is closed to us, so not much we can do at that level of abstraction. Similarly the APIs these tools use go from web services to libraries to local system services to libraries to direct syscalls to directly interacting with hardware. The number of people working with each layer directly probably drops monotonically as we get closer and closer to the hardware. ChatGPT type interfaces just seem like one big opaque blob. How do we peel the onion? Does anyone even know what those layers are or understand them?
TOMDM 4y ago
Hey ChatGPT, your code is throwing excetpion XXX/causing unintended behaviour YYY, please adjust your code. I agree with your sentiment though, people are going to cut themselves on a lot of sharp edges, but lazy devs or people after unsustainable velocity will run unvetted code anyway.
- joshka 4y ago
  Hey GPT please generate tests for the code that you write that prove that all paths exhibit correct error handling behavior...

exizt88 4y ago

Doesn't this argument also apply all the way down to machine code, though?

endisneigh 4y ago
I figured someone would reply with this. Yes, and it is already an issue. However the issue isn’t as bad for a variety of reasons, notably better abstractions. I would argue natural language can never be as precise as code without being effectively a new programming language. The effectiveness as desired is specifically due to the imprecision necessary.
- joshka 4y ago
  You're right it can never be as precise, but LLMs might allow us to identify and highlight the imprecision easily.
arghnoname 4y ago
I think they are analogous. One advantage in the programming language model though is you can pick your level of abstraction, from wysiwyg wizard time UI dialogs, to declarative languages, to languages like python, to C, to assembly, to FPGA/ASIC work. The microcode on mainstream chips is closed to us, so not much we can do at that level of abstraction. Similarly the APIs these tools use go from web services to libraries to local system services to libraries to direct syscalls to directly interacting with hardware. The number of people working with each layer directly probably drops monotonically as we get closer and closer to the hardware. ChatGPT type interfaces just seem like one big opaque blob. How do we peel the onion? Does anyone even know what those layers are or understand them?

TOMDM 4y ago

Hey ChatGPT, your code is throwing excetpion XXX/causing unintended behaviour YYY, please adjust your code. I agree with your sentiment though, people are going to cut themselves on a lot of sharp edges, but lazy devs or people after unsustainable velocity will run unvetted code anyway.

joshka 4y ago
Hey GPT please generate tests for the code that you write that prove that all paths exhibit correct error handling behavior...