Earlier this week, I lamented the somewhat terrifying prospect of 300 million lost full-time jobs globally to the miracle of generative AI.
That rough estimate came from Goldman, whose Joseph Briggs said “one-fourth of current work tasks could be automated by AI in the US”+.
For context, global layoffs since October sum to around half a million+. That counts as the most acute bout of job cuts since 2009. The “full-potential” AI replacement figure floated by Goldman would thus constitute a veritable human employment apocalypse.
But don’t despair. AI is already spawning new, lucrative employment opportunities, and you don’t (necessarily) need to know anything about computer engineering.
If you’ve used any of the generative AI applications which took the world by storm in 2023, you’ve doubtlessly noticed that there’s an art to this (computer) science. Depending on what it is you’re trying to accomplish with generative AI, the quality of the results depends heavily on how you word your request. I imagine that’ll change over time as these programs optimize around themselves, learn from their own “mistakes” and thereby demand less in the way of precision from their human taskmasters. In the meantime, though, there’s a burgeoning jobs market for “AI whisperers.”
As Bloomberg detailed in a highly amusing article published Wednesday, one Google-backed startup is offering to pay as much as $335,000 a year for a “prompt engineer librarian,” while another company will shell out nearly a quarter of a million for someone who can “prompt and understand how to produce the best output” from AI programs.
“Given that the field of prompt-engineering is arguably less than two years old, this position is a bit hard to hire for!” the “librarian” listing exclaimed, adding that,
If you have existing projects that demonstrate prompt engineering on LLMs or image generation models, we’d love to see them. If you haven’t done much in the way of prompt engineering yet, you can best demonstrate your prompt engineering skills by spending some time experimenting with Claude or GPT3 and showing that you’ve managed to get complex behaviors from a series of well crafted prompts.
Before you apply for that job, note that the low-end of the salary range is just $175,000. And it’s in San Francisco, where $300,000 gets you a nice tent, a Kia and a place to park them, if you’re lucky.
In their study, Goldman noted that even if AI does replace scores of jobs around the world, it probably won’t be a dystopian future. “The good news is that worker displacement from automation has historically been offset by creation of new jobs, and the emergence of new occupations following technological innovations accounts for the vast majority of long-run employment growth,” the bank said.
Between “significant labor cost savings, new job creation and higher productivity for non-displaced workers,” Goldman said the proliferation of generative AI “raises the possibility of a productivity boom,” which could ultimately lift economic growth “substantially.”
The bank did concede that “the timing of such a boom is hard to predict,” which could be interpreted as an admission that in the interim between AI replacing humans across a variety of “vulnerable” industries and a glorious, high-productivity future, the outlook for humans could be tenuous depending on industry and position sought.
The linked Bloomberg article was keen to note that not everyone hiring in the nascent “AI whisperer” market is likely to offer $300,000 salaries to talk to ChatGPT all day. One partner at a law firm looking to “play around” with AI called such figures “ludicrous,” for example.
Additionally, this is already being likened to blockchain, crypto and NFTs, a comparison that looks more dubious every day. Apparently, references to blockchain are being scrubbed from pitch decks in favor of nods to AI. As one tech founder dryly put it, “That might be a sign of froth.”
Yes, it might be. But for now, it could be a lifeline for some left-for-dead liberal arts degrees. An engineer who spoke to Bloomberg said these “wordplay” roles can be filled by people with backgrounds in history and philosophy.
“Hello, I’m interested in the ‘prompt engineer’ position you have available.”
“Sure! We think it’s an exciting opportunity. Can I get your name?”
“Yes, first name is Ludwig. Last name Wittgenstein.”
As a side-note to the above, I recently used ChatGPT in order to obtain information on a software development problem. The particular problem to solve was “deep in the weeds” of an area for which there is little information available via the usual search websites. ChatGPT responded with a very detailed description (and code). Previously, on a similar problem, I had spent hours, literally, scouring the internet for similar information.
To put it simply, I was amazed. And I literally build AI programs — but not in the neural network area of AI like ChatGPT.
I’ve heard from an Oracle employee that Oracle management has forbidden the use of ChatGPT for software – for now – due to lack of clarification on how copyright law will work. (Yes, software can be copyrighted, although in this particular case the employee works on software products intended only for other Oracle employees.)
I had a similar experience with YouChat. I prompted it for a pretty strange function that was not likely to appear in existing examples on the web: a recursive function to replace every occurrence of the letter “a” in any string with it’s occurrence number (for example, the string “aaabbbaca” would be rendered as “123bbb4c5”.) I specified that it had to be written in the language of a not very widely known platform.
It did take several iterations of the prompt to get it to understand… you have to be pretty precise. However, on the fifth try, it not only generated a working function, formatted as I requested, but it decided on its own to optimize the code by passing the iterator as a second parameter of the function instead of trying to store it in a global scope… a pretty common programming technique on this platform, and not elegant or strictly correct, but definitely simpler to write and understand, and the technique most human programmers would have gone with. And then it specifically called out that it had done this, and reminded me to pass an extra parameter ‘1’ as the initial iterator value when calling the function. And then, the icing on the cake, it cooked up its own example: it concluded by explaining that, for example, functionName(“banana”,1) would return “b1n2n3”. This is exactly what I asked for. When I tried the code myself on the real platform I had requested it be written for, it was bug-free and worked as asked.
I was pretty flabbergasted. It was very much like asking something of a human programmer. Only, much faster.
The possibilities are exciting.
Kudos on the clever punchline. It made me chuckle.
I have both engineering and history degrees. The wordsmiths may inherit the earth.
While there are some very relevant applications I’m in the hype camp
At the very minimum, this will be as big as what AWS did for cloud computing, but on the optimistic end, this could very well be the biggest thing since the internet itself launched. I wouldn’t dismiss this as a fad. Not making a judgement on whether it’ll be good or bad for humanity, but this is much more substance than hype.
Now I’m curious: can you give us some examples of the prompts you use to generate the images at the top of some of your articles?
I’m sorry, I’ve already licensed that information to a prompt engineering firm for $20 million.
(I’m just joking.)
Please don’t say that kind of stuff until you implement little facebook-style “laugh” (etc.) response emoticons.
It’s all about how well the companies can fine-tune the models for their specific use cases. Once you start figuring out the prompt structure, you can start to scale the inputs and the world becomes your oyster. You have an immediate business case with the efficiency gains that you can sell to your customers eager to cut costs. All that vendor spend that was being cut will be reinvested into generative AI capabilities at the expense of people.
I had a somewhat similar experience as brc’s comment where I had a call yesterday with a vendor who showed me how they’ve integrated ChatGPT into their product in a way that created descriptions of the metadata within our system. That kind of documentation is another seemingly simple but hugely valuable use case for many companies who are sitting on years of legacy tech debt that can be very difficult to unpack. Suddenly, it becomes a lot easier to replace an engineer who leaves the company if you’ve got auto-generated documentation. It was far from perfect, but the potential is massive.
Blockchain and NFTs were a joke from the start, but ChatGPT cannot keep up with the demand to integrate with their product. The bubble is merely in its infancy stage right now.
I’ve been using chatGPT to formulate better prompts for another AI doing Stable Diffusion.
I’ve been using it to formulate witty replies to blog post comments.
You can find a fun example searching YT for “GPT-4 Challenges Tesla FSD”. Not a weird bake-off; rather, it’s one of the Tesla FSD beta drivers testing whether GPT-4 could help him come up with a list of waypoints for the next tricky test drive he wanted to push FSD through. Took a few prompt iterations, but he came away pretty impressed with GPT-4’s utility as the handy assistant.
By the way, forgetting GPT, the latest FSD (11.3.3) is pretty spectacularly successful, based on tester consensus. If you’ve not checked out the YT set of FSD Beta 11.3.3 vids, give one or two a look. Engineers (like me) will likely tell you that no one else is even attempting what FSD has already accomplished, all “Beta” snarkiness notwithstanding – but watch and form your own opinions.
The ML stuff has made deep improvements (AlphaFold!) but was less accessible.
GPT is a game changer because it’s applicable now (though it did take at least an invisible decade to get “here”).
Like the Mobile boom where a ton of things we already had got easier and more available: email, taking and sharing digital photos, video calls, online shopping, etc.
And new things made life easier: everyone has an updated map with realtime tracking (which then begat Uber/Lyft)
Considering that reach made Trillions $ (which is easier to calculate than “hours saved from a paper map”) I expect this wave will also make some people very wealthy.
(So paying for a talented prompt engineer might be smart like trying to rope in an early “Mr Beast”)