STEM shock: Unemployment for US computer engineering grads more than twice that of art history

Conventional wisdom long held that majoring in a STEM field was a surefire path to job security and high earnings. But new data from the Federal Reserve Bank of New York challenges that narrative, revealing that recent graduates with degrees in computer engineering face significantly higher unemployment than their counterparts in the humanities.

According to the Fed’s analysis of 2023 Census data, the unemployment rate for computer engineering majors stood at 7.5%, compared to just 3% for art history majors. The finding has surprised many education and labor market analysts, especially given the persistent demand for tech talent in the US economy.

AI is a long way from that. I taught programming at a Big Ten university and it’s pretty easy to spot an AI answer to a programming problem versus a “real” programmer. I seriously doubt that 50% of Microsoft’s code is written by AI. That might be a goal, but it’s still a long way off. AI, right now, is good at answering “fixed” questions (e.g., How many people are in the House of Representatives?), but not so hot on those that require “reasoning” to answer (e,g,. Which is better: Capitalism or Socialism?).

I work directly with LLM codegen with up to 25000 context tokens and I don’t see the good results your AI friend sees.

For automated unit test code, I’m seeing about a 40% success rate. It does a good job with Edge Cases, but try to get a good dependency-injection mock up and running and watch it fall on its face.

"AI is a long way from that. I taught programming at a Big Ten university and it’s pretty easy to spot an AI answer to a programming problem versus a “real” programmer. I seriously doubt that 50% of Microsoft’s code is written by AI. That might be a goal, but it’s still a long way off. AI, right now, is good at answering “fixed” questions (e.g., How many people are in the House of Representatives?), but not so hot on those that require “reasoning” to answer (e,g,. Which is better: Capitalism or Socialism?)."

My tech experiences with AI (in this case, Claude) are much more positive. For example, last night I was wrestling with getting a new server stood up, and my interaction with Claude was just as if I had a senior Dell technician sitting there beside me.

Also, recently I've had Claude successfully develop an application that I've wanted to create for years, but just didn't have time for. In this case the application was to import a large amount of historical price data for several fixed commodities, while detecting and ignoring data that varied over two sigma from the mean, and also ignoring data during times of economic stress (e.g. WW1 and WW2), and replacing discarded data with extrapolated data, to produce an inflation curve that can be used to analyze investments.

Claude generated the code very nicely, and I tested it rigorously. Along the way we did encounter problems and, as with the server config, it was like having a knowledgeable collaborator working with me.

I see AI today as being like a steam shovel was to ditch diggers. A lot of ditch diggers were put out of work, but they didn't miss ditch digging so much. AI is a force multiplier. However, today at least, it requires a human with whom to collaborate.

Agree. Your application is what I termed a “fixed” problem, where you can quantify the input parameters to the problem. AI is great for generating the boilerplate code and can save a huge amount of time doing what I call “grunt programming”. However, like I said, it can only go so far when the assistant needs to step aside and let the master finish the code.

"However, like I said, it can only go so far when the assistant needs to step aside and let the master finish the code."

I agree with that, although I do believe it will improve over time. I'm not certain it will bypass the "master". I have come up with a technique, however, that helps quite a bit. My technique is to embed subqueries and localized instructions within queries, and to indicate those subqueries and localized instructions with tags that are not unlike HTML tags. This allows my prompts to micromanage the flow. I've also defined a special query for pulling in code templates, which gives me even more granularity of control.

Of course, I may just be re-inventing the subroutine! :O

The "killer app" for AI will be, for me, when it can interact with me by voice using the voice of Mr. Spock. All other considerations pale in comparison.

I find it all to be quite ... fascinating!

Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.