Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

LLMs aren't remotely like compilers or power tools

blainehansen.me

Post reply on HN
Print view View original

11–20 of 55 posts

Re: LLMs aren't remotely like compilers or power tools

#11

Post by suprjami »

mycall wrote: post #9

Is there not a way to use a transformer that is deterministic in rounds as it is just maths? I thought it was temperature which randomized results but perhaps it is also something else. Of course, the content returned is non-deterministic.

If you specify the random seed and set temperature to zero, you'll always get the same response. As you said, it's just math.

That doesn't mean the response is correct though. You've just hard-locked the model to one possible response selection.

If you and I do this but we use different seeds, we still have no way to tell whose response will be better for every question or even a class of questions.

There will be responses where the model is outright wrong and we want to change seed or introduce some randomness by raising temperature, in the hope the model answers correctly next time.

Now we're back to running with a random seed and high temperature.

Re: LLMs aren't remotely like compilers or power tools

#12

Post by suprjami »

You had me up until you equated LLM usage to unreliable human delegation.

No human being is going to genuinely suggest to glue cheese onto a pizza to stop it sliding off.

These things aren't people. Stop anthropomorphizing a computer program. They'll make all the mistakes humans can make, as well as a new class of mistakes because they're just fancy next word prediction with no lived experience.

Re: LLMs aren't remotely like compilers or power tools

#13

Post by vrighter »

i recently heard a colleague telling another "vscode can't really help you when writing yaml files". I said that the yaml lsp supports loading a json schema, all you need is just one comment at the top to point the lsp to it, then you get reliable error reporting with red squigglies on every single keystroke.

He just shrugged and said "oh i just use ai".

What takes me one keystroke takes him a full minute. And he didn't even consider the use of actual tools

Re: LLMs aren't remotely like compilers or power tools

#14

Post by rjrje »

suprjami wrote: post #12

You had me up until you equated LLM usage to unreliable human delegation. No human being is going to genuinely suggest to glue cheese onto a pizza to stop it sliding off. These things aren't people. Stop anthropomorphizing a computer program. They'll make all the mistakes humans can make, as well as a new class of mistakes because they're just fancy next word prediction with no lived experience.

Indeed.

Just to add - I’m working on something very novel and LLM’s are absolutely useless at them. It’s actually comical to see the outputs - when you work on novel stuff you quickly see what these models actually are.

Re: LLMs aren't remotely like compilers or power tools

#16

Post by trescenzi »

camkego wrote: post #5

I've got to give the author credit, it seems like this is written without an LLM. But, there seems to be a lot that is arguable. For example, the article says "Natural language is the interface of last resort." But, is that actually true? Isn't the AI boom happening because N/L is truly available as an interface? Anyway, please keep writing and asking tough questions.

Natural language is an imprecise interface. It really ideally is the interface of last resort. If you have a straight line from problem to solution you do not need natural language. It’s only if there isn’t one that you do. And in that case the same sentence can mean different things to different people even those who are native speakers. The most obvious example to me is something like “next weekend” which is a term that can mean different things to different people at different times. But it’s a deeper problem than that.

The AI boom is happening because people what to believe that natural language is precise. Not because it’s actually an effective way to interface with everything.

Re: LLMs aren't remotely like compilers or power tools

#17

Post by pennomi »

vrighter wrote: post #13

i recently heard a colleague telling another "vscode can't really help you when writing yaml files". I said that the yaml lsp supports loading a json schema, all you need is just one comment at the top to point the lsp to it, then you get reliable error reporting with red squigglies on every single keystroke. He just shrugged and said "oh i just use ai". What takes me one keystroke takes him a full minute. And he did…

The very existence of a “do anything” tool causes humans to forgo learning the multitude of tools we already have. I suspect there will be no next generation of software engineers.

Re: LLMs aren't remotely like compilers or power tools

#18

Post by inigyou »

pennomi wrote: post #17
vrighter wrote: post #13

i recently heard a colleague telling another "vscode can't really help you when writing yaml files". I said that the yaml lsp supports loading a json schema, all you need is just one comment at the top to point the lsp to it, then you get reliable error reporting with red squigglies on every single keystroke. He just shrugged and said "oh i just use ai". What takes me one keystroke takes him a full minute. And he did…

The very existence of a “do anything” tool causes humans to forgo learning the multitude of tools we already have. I suspect there will be no next generation of software engineers.

How many buttons in your IDE's toolbar do you use?

Oh wait - that's an anachronism. The older generation of IDEs, like VS and Eclipse, had toolbars. VSCode doesn't have one.

Re: LLMs aren't remotely like compilers or power tools

#19

Post by throw310822 »

suprjami wrote: post #12

You had me up until you equated LLM usage to unreliable human delegation. No human being is going to genuinely suggest to glue cheese onto a pizza to stop it sliding off. These things aren't people. Stop anthropomorphizing a computer program. They'll make all the mistakes humans can make, as well as a new class of mistakes because they're just fancy next word prediction with no lived experience.

> No human being is going to genuinely suggest to glue cheese onto a pizza to stop it sliding off.

That mistake was made more than two years ago by whatever experimental version of Google's AI overview (which has strict performance requirement- i.e. needs to answer within a split-second) was up at the time. In other words, it was a very small and primitive system specialised in spitting answers as quickly as possible without a second thought. I hope you realise that basing your assessment of what LLMs can or can't do on that example is not much better than suggesting to put glue on pizza. A mistake that, if we were to adopt your reasoning, would in turn set a hard limit to the analytic skills of all humanity.

Re: LLMs aren't remotely like compilers or power tools

#20

Post by inigyou »

rjrje wrote: post #14
suprjami wrote: post #12

You had me up until you equated LLM usage to unreliable human delegation. No human being is going to genuinely suggest to glue cheese onto a pizza to stop it sliding off. These things aren't people. Stop anthropomorphizing a computer program. They'll make all the mistakes humans can make, as well as a new class of mistakes because they're just fancy next word prediction with no lived experience.

Indeed. Just to add - I’m working on something very novel and LLM’s are absolutely useless at them. It’s actually comical to see the outputs - when you work on novel stuff you quickly see what these models actually are.

Apparently if you show one the recently found counterexample to the Jacobian conjecture, it spins in circles double-checking it because it knows the conjecture is true.
Post reply on HN