nixCraft 🐧Nvidia CEO predicts the death of coding — Jensen Huang says AI will do the work, so kids don't need to learn. Coding is old news, so focus on farming. Ah yes glorious future for all of us as everyone is going to return to the farming. What makes you think machines can’t do farming? https://www.techradar.com/pro/nvidia-ceo-predicts-the-death-of-coding-jensen-huang-says-ai-will-do-the-work-so-kids-dont-need-to-learn
David Chisnall (*Now with 50% more sarcasm!*)@nixCraft He comes so close to saying something I agree with, before talking complete nonsense.
Our goal, as people who design programming environments, should be to empower everyone. We should be building languages that anyone can use to build things that directly help them. There are roughly a billion people who use Excel to write formulae that solve their problems. That totally dwarfs the installed base of any other language.
That said, the difficult bit of programming is not the syntax (though we can, and should, do anything we can so simplify that). It's not even understanding things like objects, arrays, and different data types. We can abstract those away for a lot of common tasks. The thing that makes programming difficult is that you must unambiguously express your intention.
LLMs do not help here. English-like programming languages went out of favour because natural languages are incredibly bad at avoiding ambiguity. Mathematics had a couple of thousand years of avoiding natural languages for formal reasoning before anyone tried using machines for computation. People who need formal reasoning in natural language, such as lawyers, end up with baroque and incomprehensible dialects and still end up needing other humans to mediate on disagreements over what the meaning really is.
You can absolutely use a probabilistic system such as an LLM to tell a computer what to do. You can even, with some feedback from a syntax and type checker, guarantee that it will generate one of an infinite set of valid programs. With a reasonably high probability, it may even solve the problem that you want it to. With a non-negligible probability, it will solve a totally different problem. With a moderately high probability, it will solve a problem that is somewhat like yours but that is closer to something in your training set.
If you do your SEC filings, tax returns, or stock control using something generated like this, will you be able to tell the difference between a correct result or one that pasted in last year's financial regulations or last year's customer data because they were in the training set more than this year's?