AI truthers are a funny bunch. By this I mean people who whine about LLMs being “stochastic parrots”, dumb machines that merely “predict the next token” and therefore won’t be able to do anything of import. It is amusing watching them flounder when yet another enormous challenge falls to the dumb token predictor, seeing their wailing and gnashing of teeth.
Surprisingly, software engineering social media site Hacker News is one of the more AI-sceptical places around. Sentiment ranges from broad denial that AI is currently useful or ever will be (with lots of commenters saying that yes, they have tried the latest frontier model, and it failed to do something basic and they had to use their big softwarey brain to fix its mistakes) to people saying that AI is very powerful but still fundamentally stupid or “it’ll never be able to do X”, where X is some thing which LLMs then turn out to be able to do within six months.
(Hacker News snark is perhaps best summarised by this infamous post about Dropbox: “you can already build such a system yourself quite trivially by getting an FTP account, mounting it locally with curlftpfs, and then using SVN or CVS on the mounted filesystem.”)
It perplexes me that it’s a crowd of such naysayers. Most software engineers I know are excited to play with fun new toys. This seems to be one of the reasons AI tools are diffusing into software engineering so much faster than other domains: because AI is the ultimate fun new toy.
Yet this front page post (comments) is by an engineer who is sad about AI. He doesn’t use AI at all, positioning himself as some kind of principled holdout (rather than a luddite1). He’s bewildered that his bosses, far from seeing him as a craftsman deserving of protection actually see him as an interchangeable economic unit, and are (horror) not willing to entertain him sabotaging his own productivity.
It was a shock to me that bosses were so keen to ditch people like me and replace us with an imitation. How naive I was[.]
If the “imitation” is just as good as the real thing, is it an imitation at all? Who says that every line of code has to be written by a human anyway? Metaprogramming isn’t new; we’ve had code that writes other code for decades. Now we have English that writes code. It doesn’t have to debase your craft.
Someone working on AI safety put it to me this way: a short timelines, soft takeoff world is the most interesting.
AI timelines can be long or short, and AI takeoff can be hard or soft. If timelines are long, then the world is boring, albeit much safer (anything interesting is years away from happening). If takeoff is hard, then everything changes out from under us in an instant and who knows, maybe we all die in short order.
But short timelines and soft takeoff means that we get to actually witness the magic of AGI unfolding before us, and to enjoy some of the amazing riches it produces, even as it threatens us in short order.
Feel scared, but don’t feel sad. We might as well enjoy the spoils.
When I was last hiring, I automatically disqualified candidates who had made no attempt to use AI.
We need to take the word "trivial" away from the Twitter/HN bros.