10 reasons computer science degrees must change for the AI era

Does studying computer science still help in the age of AI? The name of the degree suggests it’s an obvious place to start. Want to make that box with the flashing lights behave? Want the chips and screens and keyboards to fall in line into an army that will conquer your problems? A CS degree can only help.

After all, if your goal is to master the computers that power AI, why not start with a science that has “computer” in the name? Universities have already built fine degree programs that teach many of the machine’s secrets, and over the past few decades many of the graduates of these programs have leveraged that knowledge to play a role in the computer revolution. It’s worked for them.

But many have also succeeded without a CS degree, instead coming from fields like physics, accounting, or even philosophy. Some don’t have any degree at all. 

And the gap between expectation and reality is only growing. AI and large language models (LLMs) are changing the rules about every field — and software development may be undergoing one of the biggest changes.

With AI increasingly part of coding and IT work in general, the question becomes: How can CS degree programs evolve to ensure long-term relevance and better prepare students for the work they are likely to be enlisted to undertake? The changes are so dramatic that degree programs must reform at least in some measure to suit.

Following is a set of reasons why the old computer science curriculum just isn’t able to prepare people for the modern realities of wrangling computers today.

All the old arguments against degrees still hold — perhaps more so

Computer science degree programs have long been questioned about their usefulness and relevancy, given how much of their curricula are mired in theory and snobbery. And all of the reservations about the necessity of a computer science degree still apply.

Learning about compilers is just as unnecessary for many students today as it has ever been. Practical skills are often pushed aside at institutions of higher learning in favor of theoretical and philosophical follies. Political games are just as dominant as ever at the schools. All the old reasons to skip the degree are not only still valid, but even more glaring. If AI is going to be an essential part of all computing work, clinging to the past and wallowing in old curriculum makes the degrees even more skippable.

AI is eating up the detailed work many courses pore over

A good part of any computer science degree is mastering several of the big languages by learning the details of punctuation and defining functions. And after students learn the big languages, they’re often taught some of the odder, more exotic languages because, well, it’s college. Now LLMs do all this for you.

In the past month, I’ve been writing code in several languages I’ve never studied before simply by asking the AI to write them for me. The AIs make mistakes from time to time, but they’ve been rare. I’ve been very productive in these languages without devoting any time to a CS course steeped in the details.

Spending time learning the rigor of programming languages may never go away, but the CS curricula needs to reflect the fact that much of the arcana about idioms, syntax, and punctuation will be handled by the machines. The cult of languages and the arguments over the right syntax can fade away and humans can skip those lessons.

Most CS courses are now obsolete

It’s not just the intro courses and the stress over where to put that semicolon. The obsolescence runs right up the stack. It’s not that the facts are now wrong; it’s just that most humans won’t need to deal with them.

Occasionally, some humans will need to dive in deep enough to straighten things out, but for the most part, humans will rely on AIs to do much of what’s taught about programming languages, algorithms, networks, data structures, and almost everything.

The AIs will handle most of the details that fill the brain of computer science graduates. Humans will still need a high-level view and a creative strategy, but that’s barely covered in the four years of a traditional CS degree. It may be very hard to teach the high level, strategic vision, but the students need to try.

The fastest moving areas of computing are too fast for a course

The business model of the university is aimed at being a trustworthy repository of the world’s foundational knowledge. It was designed around topics like classics or history — subjects that may progress through reinterpretation but rarely change.

Right now in the world of AI, the LLMs are changing from week to week. By the end of the semester, the sands have shifted so much that the syllabus for the last few weeks will need to be redesigned. Yet CS departments expect you to sign up for four years. Not only that, they want you to start planning to attend years in advance so you jump through their hoops and do their little admissions dance.

If the universities are going to have any utility at all, they need to shift and adapt quickly. Lab courses, hands-on assignments, and more evolving seminars are better than staid textbooks written by people who have already died. The curriculum needs to be more of a metacurriculum because the details will change faster than ever before. They need to be built for adaptability, or you may come out of your program years behind.

Will the jobs still be there when you’ve finished your degree?

For the past few decades, CS programs swelled to accommodate the demand for vast armies of tech-savvy workers. It’s still too early to know whether AI will destroy much of this need, but it’s certainly possible.

Many programmers say they are five, ten, or maybe even twenty times more productive using LLMs as their assistant. Does that mean that the world will need 1/5th, 1/10th, or 1/20th the headcount of programmers? Certainly new jobs often replace the old ones destroyed by technology, but the scale of change this time may be dramatically different.

Without evolving in the face of this uncertainty, what kinds of careers will computer science programs be preparing their students to enter? More practical collaborations with real internships could help students better glean where their efforts need to go to be ready for what’s available when they’ve completed their degree.

Savvy neophytes can now more easily outshine the academically credentialed

CS degree holders were in demand because they had spent years mastering the arcana of commanding these machines. Now the AIs know most of these messy details and this lets a human accomplish just as much with some vague handwavy phrases.

Wishful requests to the AI are just as effective as the hard-won knowledge that comes from years of wrestling with computers. So why study or memorize? A smart newbie can luck into creating the next big thing. Oh sure, good programmers have a bit of an advantage but it’s fading.

The industry has always been willing to hire people with degrees in physics or mathematics. Now people who’ve spent their formative years dabbling in anything can be just as effective as someone who’s spent a semester studying compilers. Perhaps more so. Yes, there will always be a need for the omniscient computer genius who can diagnose weird and inscrutable issues that the AIs can’t grok, but for everything else the boss may be just as happy with someone who doesn’t even know where to find the latch to the hood.

The need for the deeper thoughts of CS is becoming rare

There’s no doubt that some corners of CS deal with complex questions about the nature of knowledge, the importance of self-knowledge, and the limits of mathematical reasoning. These areas are sometimes fun and always deep. But very few employers want to pay for them. Yet they fill up the curriculum.

What companies want is their data arranged in nice tables. AIs don’t change this balance. They’re best at the useful tasks such as writing glue code to make the data line up nicely in tables. We can argue for hours about whether the modern LLMs can pass the correct version of the Turing Test. We can debate whether we’re any closer to achieving true artificial general intelligence. But the boss says do it on your own time down at the pub.

It would be a shame to eliminate such mathematical navel gazing from the computer science curriculum, but the degree needs to recognize that these can be more luxuries than the foundation for a career.

AI’s randomness undercuts standard algorithm courses

For their entire history, computers have been entirely predictable. Aside from an occasional alpha particle that flips a wayward bit or some weird I/O glitches, the machines always turn the same input into the same output. But LLMs are designed to be different. The algorithms incorporate some random number sources to respond differently each time.

All of this means that the standard algorithm courses with their rote devotion to a very mechanical process just don’t help us understand AIs and LLMs very much. They’re about as useful as a physics course about gravity to an astronaut floating in outer space. Yes, gravitational forces are still in effect, but they don’t make it any easier to bounce around the capsule.

CS programs need to open up the algorithm curriculum to reflect this. They need to ensure students are versed with the newer, less deterministic models of computation that are certain to remain a dominant part of the field for years to come.

CS is not much of a science anymore

From the beginning, many professors have bent over backwards to establish a very academic foundation for the discipline of CS, filling the textbooks with theorems and experiments. The success has even drawn some attention from “official” sciences, like physics. John J. Hopfield and Geoffrey Hinton, for instance, won the Nobel Prize in physics for developing the neural network.

All this doesn’t hide the fact that LLMs are more like Rube Goldberg machines than manifestations of  proper scientific theories. One sales engineer from a big LLM company told me, sotto voce, that their latest model responds differently to prompts with two spaces after a period than prompts with only one. Any prompt “engineer” knows just how flaky the models can be. The prompting strategies that work for version X of the LLM fail miserably with version X.1. If the “temperature” is a bit too high, the answers can careen off into craziness.

So, sure, the syllabus for a CS course can be filled with lots of official sounding science and the papers can be filled with lots of greek letters, but these days the real education comes from long hours of fiddling with fickle AIs. Why not incorporate that?  

Is a creative writing degree better?

Words are the interface for many modern LLMs and so it only follows that a mastery of words is a key skill for working with LLMs. A broad and sophisticated diction coupled with a savvy ability to deploy the right word are the foundation for success.

The best prompt “engineers” are out there searching for the best sentences that tickle the right part of the LLM’s matrix of weights. They are learning how to say the right things because each word nudges the answer in the right direction. As Mark Twain once said, “The difference between the almost right word and the right word is really a large matter. ’Tis the difference between the lightning bug and the lightning.”

While CS programs don’t need to go full Shakespeare, they must recognize that natural language has become and will continue to be a key interface for interacting with machines today. It would be a good first step to design creative natural language-centric coursework to ensure students are adept at it. The best place to look for that skill might be in the creative writing department, not the computer lab.