(disclaimer: I'm a software engineer with minimal compiler theory experience outside classes in college)
I wonder whether its possible to trust an LLM to "compile" your code to an executable and trust that the compiled code is faithful to the input without writing a static validator that is pretty much a compiler itself.
rvz•2mo ago
Don't take it seriously, It is Twitter bait from an intern.
> I wonder whether its possible to trust an LLM to "compile" your code to an executable and trust that the compiled code is faithful to the input without writing a static validator that is pretty much a compiler itself.
"LLMs as compilers" do not make any sense.
Traditional compilers must be deterministic to compile to the correct machine code for the correct architecture otherwise the executable will break.
canttestthis•2mo ago
The author is a Principal Engineer at DeepSeek. I don't know what that title maps to in other organizations. Their formal education/background is in ML however.
rvz•2mo ago
> The author is a Principal Engineer at DeepSeek.
The title is even more meaningless given that such a suggestion sounds totally at fundamental odds with today's state of the art compilers which need to be deterministic.
Replacing today's compilers, linkers and assemblers with LLMs (which are fundamentally stochastic preditcion models) for the use-case to compile software (not just generating correct code syntax) makes absolutely no sense whatsoever.
> Their formal education/background is in ML however.
Maybe whoever this person is should read up on what the Gell-Mann Amnesia Effect is.
Ask any compiler author on this very suggestion and they will question if this person is really a "Principal Engineer".
tuxracer•2mo ago
This seems like such a jerk move to reply to someone who worked hard and is excited about something to essentially try to tell them it was worthless. Whether an LLM will ever actually be appropriate as a compiler or not, the reply from Chen Fang is in such poor taste.
tjr•2mo ago
If true -- why? What would be the advantage?
eyeris•2mo ago
Think speed could be a potential benefit. Why go through source -> llvm ir -> binary with linking and all the other stuff? Think I’m bearish on the concept, but who knows?
One parallel could be using ml for simulations to not write or compute all the rules
canttestthis•2mo ago
(disclaimer: I'm a software engineer with minimal compiler theory experience outside classes in college) I wonder whether its possible to trust an LLM to "compile" your code to an executable and trust that the compiled code is faithful to the input without writing a static validator that is pretty much a compiler itself.
rvz•2mo ago
> I wonder whether its possible to trust an LLM to "compile" your code to an executable and trust that the compiled code is faithful to the input without writing a static validator that is pretty much a compiler itself.
"LLMs as compilers" do not make any sense.
Traditional compilers must be deterministic to compile to the correct machine code for the correct architecture otherwise the executable will break.
canttestthis•2mo ago
rvz•2mo ago
The title is even more meaningless given that such a suggestion sounds totally at fundamental odds with today's state of the art compilers which need to be deterministic.
Replacing today's compilers, linkers and assemblers with LLMs (which are fundamentally stochastic preditcion models) for the use-case to compile software (not just generating correct code syntax) makes absolutely no sense whatsoever.
> Their formal education/background is in ML however.
Maybe whoever this person is should read up on what the Gell-Mann Amnesia Effect is.
Ask any compiler author on this very suggestion and they will question if this person is really a "Principal Engineer".
tuxracer•2mo ago