Hacker News Reader

Comment by masterj | Hacker News Reader

Calculators have been available forever, but have not eliminated math education. Even algebra systems that can solve equations, do integrals and derivations have been available forever, but people understand that if they don't learn how it actually works they are robbing themselves. By the same token, if you need to do this stuff professionally, you are relying on computers to do it for you.

> Write a binary tree in C? Check. Implement radix sort in Python? check. An A* implementation? check.

You can look up any of these and find dozens of implementations to crib from if that's what you want.

Computers can now do more, but I'm not (yet) sure it's all that different.

View on HN

I agree with you, but just to steelman the other side: how do you know when you are robbing yourself and when you are just being pragmatic?

When I change the spark plugs in my car, am I robbing myself if I'm not learning the intricacies of electrode design, materials science, combustion efficiency, etc.? Or am I just trying to be practical enough to get a needed job done?

To the OPs point, I think you are robbing yourself if the "black box" approach doesn't allow you to get the job done. In other words, in the edge cases alluded to, you may need to understand what's going on under the hood to implement it appropriately.

> how do you know when you are robbing yourself and when you are just being pragmatic?

I don't know why we're pretending that individuals have suddenly lost all agency and self-perception? It's pretty clear when you understand something or don't, and it's always been a choice of whether you dive deeper or copy some existing thing that you don't understand.

We know that if we cheat on our math homework, or copy from our friend, or copy something online, that's going to bite us. LLMs make getting an answer easier, but we've always had this option.

I don’t know why you are ignoring the concept of opportunity cost to make an argument.

Did you drive to work today? Did you learn everything about the tensile strength of nylon seatbelts before you buckled up? How about tarmacadem design before you exited your driveway? Or PID controls theory before you turned on the HVAC?

The point I’m making is that some people disagree about how much you need to know. Some people are ok with knowing just enough to get a job done because it makes them more productive overall. So the question still stands: How do you gauge when learning is enough? To my point above, I think it comes down to whether you can get the job done. Leaning beyond that may be admirable in your mind, but not exactly what you’re being paid for, and I think some experts would consider it a poor use of time/resources.

Do you care if your subordinate wrote a good report using a dictionary to “cheat” instead of memorizing a good vocabulary? Or that they referenced an industry standard for an equation instead of knowing it by heart? I know I don’t.

I said it’s up to the individual to make their own choices. I don’t know who or what you’re arguing against, but I don’t think I have much to do with it. Peace

I was simply asking you to put a finer framework on how individuals should decide. Like I said multiple times, IMO it comes down to what is needed to get the job done, but I’m open to other thoughts. Saying it’s up to the individual isn’t really saying much, other than a veiled version of “I don’t know but I feel the compulsion to comment anyway.”

> I was simply asking you to put a finer framework on how individuals should decide

I'm not in the business of prescribing philosophies on how others should live their lives?

But surely that doesn’t preclude you from describing how you live your own?

I think you have a good point, but I think the paradigm shift here is that people are chasing careers and money using LLM tools in a way that wasn't possible with calculators, and enforced differently within well-paying, white collar engineering professions.

For example, there's actual liability (legal and financial) involved in building a bridge that subsequently falls apart - not so with small bits of code block. Similarly there's a level of academic rigor involved in the certification process for structural/mechanical/etc. engineers that doesn't (and maybe can't?) exist within software engineering.

>certification process for structural/mechanical/etc. engineers that doesn't (and maybe can't?) exist within software engineering

NCEES has a PE license related to controls software. The difficulty is most engineering work falls under an industrial exemption. It seems like the way to enforce that type of liability would be to remove the industrial exemption.

I'm not really sure what problem you're trying to point out here. There are legal standards and liability for engineering, and if someone violates them using an LLM they are held just as liable as they would be had they done the work themselves.

But the same is true for code? You are held to the same standards as if you had written it yourself, whatever that may be. Frequently that is nothing.

What change do you want to see here?

>What change do you want to see here?

Licensing, liability, and enforceable standards for safety critical code that interfaces with the public would be a good start.

I think all of those are great, but don't think that has much to do with AI tbh. How you get to the outcome, and the standards that it meets, should be all that matters

You don’t think the black box nature of much of AI has any impact on its impact in safety critical applications?

Just look at the diverse and haphazard way AI has been used in autonomous driving. I would argue it’s a misplacement of the “move fast and break things” (in some cases at least) that has no place in public-facing safety critical applications.

It brings up some difficult questions regarding adequacy of testing at the very least when the underpinnings are not very interpretable.

Ah, my understanding was that this discussion is about AI as dev-time tool, where the output is code, which gets reviewed : merged / deployed like any other bit of code whether written internally or installed via a library.

Using LLMs or other ML as components in systems themselves is a whole other thing, and I agree with you wholeheartedly.

That’s a real distinction, but auto-generated code has special complications in safety critical code review as well. It’s also not relegated to ML/AI, just that adds yet another complication to good verification and validation.

There are still schools where you can learn to shoe a horse...

Reply on news.ycombinator.com