I think that's a distinction that the author's language made. Most programming problems have been solved. The head of the Zipf-curve of your work, no matter the domain, will be reusing old-hat knowledge. Every program, even HFT, needs proper data structures, algorithms, task scheduling, etc. Hopefully this stuff is mostly already provided by your platform, but a great majority of every developer's time is basically "bringing the platform up-to-snuff" with everything it really should have had to begin with.
Once that's done, the code that's actually business-domain specific (whether the domain is ERP, trading, games, graphics, AI, simulation, etc.) will likely be novel, because it's a codification/automation of a business process which is itself novel. (And business processes being novel is sort of the Anthropic argument of capitalism: a perfectly non-novel process means there's already a company doing it somewhere with lead-time on you, which means you have no advantages and won't survive in the market. So every new business that survives must have some sort of mutation, for better or worse.)
Did you know, that if you make a control loop running basically newton's method of approximation (more or less), it can be statically stable in continuous analysis but when you run it on real world discrete math with short-ish rounded floating point it can get squirrely and oscillatey (are those even words?) so that's why you ram the output thru a low pass filter to damp out the oscillations. Or at least that's one way to fix it.
Oh another funny one is picking a local maxima / minima that randomly happens to have a near zero derivative. From memory Newton doesn't like that very much. Another funny one is any time you have tiny rounded floats divided by other tiny rounded floats, who knows what you'll get as an answer. You can fix that by fuzzing the input first guess with some random noise. Makes your execution speed non-deterministic, but stuff is fast now a days isn't it, and its faster than doing double/quad floating point anyway, which may or may not help much.
Sure the google found some python or fortran or mathematica which on the surface is useless, but psuedocode is psuedocode and compilable fortran is a tighter syntax than any random human generated psuedocode so in some ways its better.
This is before you hit the "well, that library speaks to 16x2 LCD displays, I guess I'll have to start over from scratch because I'm using a 20x2 LCD..." Well, typically, no.
For HFT I agree, you're stuck behind layers and layers of patents and trade secrets. But embedded is not a problem to get google-type help.
I encourage anyone who wants to get an intuitive grasp of simulations to just write a simple little 2D game engine, and try to work up toward a simple two-player realtime LAN system with predictive physics for expected network input. Seeing sprites shudder and vibrate against their integer pixel positions, jump instead of smoothly curving when a correction is issued, etc., is a visceral learning experience.
Well, "chaotic" is what you'll read in textbooks, but that's just because they couldn't come up with "squirrely and oscillatey", I assume.
This depends a lot on the domain you're programming in. For example, if you're working on a proprietary application like high frequency trading or code for an embedded device, it's not likely that you'll find an open source library to do whatever you need, or get a lot of help from Google or Stack Overflow.
Also, a lot of open source libraries are not robust enough for commercial applications that have requirements like support of legacy East Asian character encodings (pre-Unicode). Not all programmers work on web apps where you get to control the environment (hardware, operating system, database, languages, etc.) you run on. Lots of software has to be built to run in the customer's existing environment.