A lot of people have predictions about what field will be next to fall to AI. People didn't see art coming. Tesla FSD finally got good in the course of a few months. Code-writing is getting better every few months. What's next? And what will be the economic impact?
I think the order AI will move through the economy is from easy-to-audit fields to hard-to-audit fields.
Art was first because art is the easiest for the person paying for it to audit. It looks good or it does not. Audit complete.
Github Copilot was next because programmers are pretty good at auditing small snippets of code.
But things like AI-accountant? AI-contract lawyer? Those produce large outputs that must be meticulously audited by a human expert to be useful, and will continue to require a human expert until AI is obviously superhuman. The audit process is essentially the same amount of work as just doing it.
So I could imagine AI copilots for these fields, where the copilot is producing snippets or checking the human's work, the task won't be end-to-end automated like producing images now is.
All of the following tasks have the property of being easier to audit than to do, as well as being in a favorable direction for current trends.
This is the clearest one to me that's going to be completely solved within a few years. Book text input. Prompt describing the narrator's style. Listen to a few samples. Pay $10 and get an audiobook output in a few minutes. End-to-end solved.
This is a shame if you are a professional audiobook narrator! But great for audiobook lovers who want to listen to a book that has no audiobook.
I've used text-to-speech to create a few audiobooks already. It's not perfect — way too emotionally flat — but actually pretty listenable if the content itself is good. But soon the emotional regulation will be good enough for mass-market and that will be that.
By cafe music, I mean stuff like chill acoustic guitar, lofi beats, etc. Suno and Udio are getting really good really fast.
Mainstream music will be bottlenecked by lyrical abilities of LLMs which are still inferior to humans. It seems likely that generic pop or pop-country — both not very lyrically complex genres — will fall first.
There are a lot of call centers were humans just act out a flow chart. GPT-4o has demonstrated real-time conversation capability, and this will just keep getting better. The flow charts will be automated.
Dominos Pizza, for instance, has these call centers. The phone operator just basically does the "order online" for you. I suspect they'll be fully automated within 5 years.
Call centers with more open-ended mandates (not flow charts) are harder, but these are a minority of requests.
Dubbing movies, translating technical documentation, and books that don't have a lot of demand for a mega-high-quality human translation — all automated very soon. The starts of this are already happening. This AI dub of a speech by Argentinian president Javier Milei is a great example. It's essentially perfect, and sounds like Milei just speaks fluent English and gave the speech in English, but it was actually given in Spanish.
You'll still see human translations for more "high-literature" type books until AI is obviously superhuman — AI will still act as a copilot before that, though.
There's a whole cottage industry around editing out "uh, umm" from podcasts and transcripts.
A lot of local journalism is just listening to public meetings and reporting on what happening, and connecting those things to some narrative of previous reporting.
This is probably the most controversial, but I know several people who already use ChatGPT or Claude as a sort of therapy. Just someone to vent problems to and provide friendly responses. It's entirely conceivable to me that an LLM therapist will be better than a median human therapist very, very soon.
The main hurdle to overcome is the current lobotomized models have a "tone" already baked in — whereas you'd want the ability to make a more human-sounding model for a therapist. Large open-source base models lake Llama 3 are making inroads here.
AI actually has an advantage over humans here in that most people feel less judged when talking to a computer.
The following are just as hard to audit as they are to do, so will likely not be given over to AI until AI is obviously superhuman.
Accounting, banking, investing, etc. Very few will put their trust in AI when large amounts of money and legal liability are on the line until AI is superhuman.
Similar to finance. Large sums of money, legal liability, etc. Lawyers will use copilots, but not be replaced for a long time.
Things like implementing databases, network protocols, filesystems, financial code, etc will stay human the longest. Things where off-by-one errors are super subtle and easy to make. AI devs will probably do well for simple app development, though.
I have more predictions about what will and won't happen based on other trends, but these are my predictions based only on the auditability heuristic.
For example, even though reading a novel is a lot easier than writing one, writing a compelling novel is just really hard, so I don't think authors will be surpassed soon — writing good novels is probably AGI-complete.
I posted this to a group chat and learned my friend Milan had written a very similar article 4 years ago! I like his explicit formulation as cost of AI + cost of checking AI.