Drew Harwell

New: Tech’s hottest new job: AI whisperer. No coding required. Prompt engineers sweet-talk GPT-3 and other generative-AI tools using plain English - a revolution in how humans work with machines. But this "tech priesthood" may only have the illusion of control. How long will their incantations last?

https://www.washingtonpost.com/technology/2023/02/25/prompt-engineers-techs-next-big-job/

Tech’s hottest new job: AI whisperer. No coding required.

Proponents of the growing field argue that the early weirdness of AI chatbots can be avoided by a human giving the machine all the right instructions.

The Washington Post
Feb 25, 2023 at 12:10pmYuito
Show threadJesseFeb 25, 2023
@drewharwell Ok, what I don’t understand is why prompts are delivered in the second person. In almost all the examples in English of a character description followed by that character acting according to the description, it’s written in the third person — right? Is the training corpus just stuffed with like 20% cheesy hypnosis scenes?
Show threadGiovanni CampagnaFeb 25, 2023
@misc @drewharwell it's the result of instruction finetuning + RLHF finetuning. Prompts change a lot compared to original pretraining. Also, even in the original corpus 2nd person was common for coding-interview-style text, which used to be easy to accidentally retrieve asking for a logic problem or a problem with structured output.
Show threadZ O R A N G R B I CFeb 25, 2023

@drewharwell Ha, ha, some techbro is going to invent a SaaS solution that can talk to a chatGPT engine in ways no human can.

Five tiers available, free crap, crap pro, crap for teams, entreprise crap.

Show threadJohn GordonFeb 25, 2023
@drewharwell I’d guess 1-3y. Which is a long career in the times to come.
Show threadBud GibsonFeb 25, 2023
@jgordon @drewharwell I think it’s been a thing already for at least 5 years. The whole field of SEO has become about prompt engineering.
Show threadAlbotFeb 25, 2023
@drewharwell fully expect my job as prompt engineer in its current form to be obsolete in 3 years . Hell, the amount you have to prompt engineer Chat-GPT compared to Text-Davinci-002 eg is already much less . But by then there will be hundreds of other jobs in this space that don’t even have names yet .
Show threadLaurelFeb 25, 2023
@drewharwell isn’t this the plot of Ex Machina?
Show threadSam Oldman 🐀Feb 25, 2023
@drewharwell At some critical point the AI minders will secretly switch to writing the stuff themselves, without telling their management that it's faster than proofreading and fact checking the automated rubbish.
Show threadMichael MartineFeb 25, 2023
@drewharwell an intriguing concept, especially when secondary education is so focused on the technical that literature, history and the arts are in fact quite intertwined and necessary for this kind of work. Much has been said about uncommon expertise in search being a differentiating skill set. #PromptEngineering is a natural extension. Also discussed on the #podcast episode of https://www.gamesatwork.Biz that will be published on Monday.
Games At Work dot Biz e548 — The Uncomfortable Valley

A weekly podcast and blog where we focus on metaverse & gaming technology in business, gamification and the business of games.

Games At Work dot Biz | Play games with us!
Show threadMichael MartineFeb 25, 2023
@drewharwell From the recent experiences, it would seem that the AI co-pilots require co-pilots of their own. Cc @gamesatwork_biz
Show threadJohn SamuelFeb 25, 2023
@drewharwell Unlike calculators and computers which gave accurate answers, large language models have their limits. They may spout out complete nonsense, and if you do not understand and verify the output completely, this may have consequences. Blindly using them is not the right way forward.
Show threadSamhain NightFeb 25, 2023
@drewharwell Actually not the Onion