Hacker News Reader

Comment by armchairhacker | Hacker News Reader

ChatGPT didn't end the knowledge economy, because it doesn't possess reliability, niche knowledge, and experience which is hard to put into words. Reliability, niche knowledge, and experience (pick two) are what get people get high-paying, stable, knowledge-based jobs, like legacy system engineer, sysadmin, and <thing> advisor.

Surface-level knowledge (which ChatGPT is good at) was already accessible by anyone who can do basic research, albeit slower. It wasn't a gated skill, so I can imagine some people achieved decent jobs from it, but most managers probably either did it themselves or assigned it to someone with a different job.

View on HN

If I could add to this, in recent hiring I have dealt with people using chatgpt in written communication and in interviews.

I interview remotely. The long pauses that get longer as the questions go from superficial to "you need to have done this for a while to know" are frustrating and obvious but also a great indicator if what is going on.

I always ask "what is happening during the long pauses". No one yet has said " I'm thinking". That's probably because I have always said going in "if you need time to think, or don't know, we can just talk about your thought process; noone knows everything and we all learn as we go". So a non-threatening out exists.

Only happened a few times.

Last time I had a senior dev on the call typing in the questions to chatgpt as we went and he was getting the same answers as the candidate with points in the same order. Some of these were wrong in very specific ways.

Back when I was marking assignments in university that was by far the most reliable way to catch cheating: two assignments with the same quirky error in the same spot. It didn’t even have to be a shared code/logic error; a misspelled comment or variable name was enough to mentally pop up a red flag and go back to find the other assignment with the same error in it.

On some of the assignments we gave out we would provide a set of, say, 10 initial test cases that the students could use to verify their work with the expectation that they would write more of their own. As a marking team we would also generate our own more detailed test suite; partly to make our job easier, partly to try to identify trends of places where students might not be understanding the course material. That ended up being another interesting way to identify plagiarism: if only 2 or 3 assignments failed a given unit test we would go and look to see if the implementations on those assignments was identical…

This is our experience as well and is an immediate rejection if the person is not honest about it. If they say "i need to check this quickly on a search engine or chatgpt" that's fine. But if they make it sound like it is their own answers, that's dishonesty and you don't want to work with people you can't trust when you are given the choice.

You caught only those who are just bad at lying.

Potentially yes, but that still reduce the population of liars in the end. If you have a recipe to catch people good at lying I'm all ears.

Onsite interviews.

You are 100% correct about AI today, however, I feel people are way too fast at making a final assessment of what AI can and cannot do as if it was set in stone. People seem so sure that the problems AI are having now are just intrinsic and just won't be overcome.

But think about what computers were like when they first hit the market. What about the first year of public internet? Or the first generation of cellphones. These technologies also had major issues and shortcomings that would, if not improved, severely limit their usefulness.

Nobody knows what AI will be like in 5 to 10 years with the burst of investment and focus that this past year has ensured.

> But think about what computers were like when they first hit the market. What about the first year of public internet? Or the first generation of cellphones. These technologies also had major issues and shortcomings that would, if not improved, severely limit their usefulness.

The question is whether AI's limitations are more like too I/O limited, too slow, not enough storage, or more like fundamental things like GIGO and (virtually) all software having bugs and people having to handle that.

We know what the internet has developed so far, and it's not all positive.

We don't know what and how LLM will impact the future.

> We know what the internet has developed so far, and it's not all positive.

A backup communication system in case the Soviets took out the phone system? What's non-positive about that?

The Soviets never took out the phone system :(

https://en.wikipedia.org/wiki/ARPANET#Inspiration

You aren't wrong, but we can't deny that internet changed the world, changed industries, and changed society. It's not all positive, but it's become a fundamental building block of mankind at this point.

> You are 100% correct about AI today, however, I feel people are way too fast at making a final assessment of what AI can and cannot do as if it was set in stone.

OpenAI is valued at $100b or something and given the hype I would have thought that it would have solved basic knowledge economy things like spreadsheet analysis, something that millions of well-compensated people are currently employed to do.

Why is there no "AI Excel" that can take a 1GB CSV and answer questions that are one step more complex than Pivot Tables?

As a business user, I don't care about AI that pretends to 'talk' like a human and 'draws' pictures. I need it to be able to do the work of a securities analyst.

> Why is there no "AI Excel" that can take a 1GB CSV and answer questions that are one step more complex than Pivot Tables?

Not sure what file size limit is but chatgpt can do this.

The code interpreter probably has access to sqlite for this purpose, as well.

You can't actually use startup investment numbers as a metric anymore because just like the stock market it's just become a avenue for financial gambling with little regard for the underlying industry.

I've thought a lot about exactly these limits to AI. The reliability problem will be solved sooner or later. Same for tacit skills that somebody trains into AI (manage people, do laundry). But if your job is specifically about being human ("nostalgic jobs", judge, athlete, surrogate mother...), or some niche tightly guarded ability (how to breed some special spice, how many oil barrels ship thru Suez Channel), you are safe - until some unforseeable rearrangement of the economy, at least.

> until some unforseeable rearrangement of the economy

I mean, telling the majority of the population that hey, you don't have secret knowledge, now you are mostly worthless is a completely foreseeable rearrangement of the economy, at least in the sense the rearrangement is coming. The question of how the new arrangement will look is unpredictable and likely spans anywhere from UBI to guillotines.

What counts as "niche knowledge"? Less than 5% of the workforce knowing it? 3%?

Software engineers collectively are 2.54% of the total USA workforce; the only time I've used YACC was at university, and given what it does I have to assume that (despite the large number of languages out there) YACC knowledge is relatively niche… and it's something ChatGPT "knows".

It might not be reliable enough (yay, I still get to be paid for my existing skills!), but I think it absolutely does have niche knowledge.

I'm talking more niche: topics where the amount of people in the world who know range from no more than 10,000 to under a dozen. Topics like, how to navigate a legacy system or bureaucracy (which may not be 100% public), dense academic subjects (especially recent ones), edge-cases in libraries and operating systems (ChatGPT is likely to miss anything unintuitive unless it's well-known). Maybe some topics with a wider audience are also protected from ChatGPT, but those definitely are.

You can (or at least could before ChatGPT) definitely get a job just by knowing yacc, aria, Kubernetes, etc. which are certainly niche subjects. But you get a very high-paying job knowing how to write highly-optimized code for a specific type of hardware, how to implement a cutting-edge algorithm in <insert language>, how to integrate a badly-documented library filled with edge cases, or how to satisfy legal/government requirements. These are jobs which rely primarily on knowledge, and a tiny bit of putting pieces of knowledge together.

Hasn't yet. Look at velocity, not position. ML is blowing past every bench mark way faster than we were 10 or even 5 years ago.

Is the knowledge economy really defined by knowledge though? Or is it simply that labor needs have shifted to information management

Reply on news.ycombinator.com