Nobody using “prompt engineer” as a “skill” in “ages.”
Yeah, feels like almost a year since “prompt engineer” faded. In decades gone by people claimed skills in crafting Google search prompts…
When I was “engineering prompts” a year ago, it was absolutely necessary to prompt in certain ways to get better results, but… I could just feel it at the time, everything I was doing was blatantly obvious and the models were learning it almost as fast as I was prompting them - what I had to do one week, I didn’t have to do the next week because the scaffolding incorporated it for future sessions.
The whole thing gets messy when you’ve got too much prompt for the context window, so there’s still some “engineering” to do there, keeping your prompt-info organized so it can be navigated efficiently, not blowing through your context window budget before you get all the relevant information loaded before a task execution…
I was explicitly taught google search skills in nursing school. It focused on how to look for results from major medical journals, the CDC / FDA, articles written by well known MDs/DOs specialists, and how to find manufacturers’ manuals / YouTube tutorials for obscure medical devices. I was explicitly taught that my future was probably going to include looking up the manufacturer’s YouTube channel to find out how to access my cancer patient’s obscure chest port by a 60y/o instructor who clearly voiced how insane that was compared to her work experience.
For better or for worse AI’s gonna end up in class somehow too, I’m just hoping it’s late enough down the line that it’s been standardized to super basic shit like fuzzy matching search results and advanced grammar / spellcheck. One of our sister hospitals is trialing a service that can generate a flowsheet column for review from listening to a narration of a head-to-toe and I don’t immediately hate that. Actually you know what would be a really cool use? Fuzzy matching name alerts in the unit patient list. Had two obscurely spelled patient names that sounded very similar the other day with a matching last initial that didn’t trigger name alert highlighting on the board. An llm might be able to catch the sound-alike with a little more flexibility than the current algorithm.
I worked with a research oriented M.D. in the 90s, he had a friend die suddenly in his mid-60s and he made a project of figuring out what really happened to his friend. He was adept at PubMed searches, etc. and worked the problem for 6 weeks, about 4 weeks in he had a solid case for the misdiagnosis that led to the maltreatment that made his friend first worse, then within 72 hours dead. Still, he didn’t find any basis for malpractice (you can always file a lawsuit even with less than zero basis, but that wasn’t his purpose.)
At the end of it, he was left wondering what the solution is, because no one person can carry sufficient knowledge to make accurate diagnosis in all cases all the time, to make the right referrals even. He died in 2020, I bet the LLM agents (AI) of the past year are what he was searching for 30 years ago - it will take time for adoption and trust, but in the end I think we’re going to get better benefit of all that knowledge captured (currently all but buried) in the research publications. As long as power players like big pharma don’t turn the whole thing into a pill referral system - they way they corrupted my last GP’s practice. Any conversation with that M.D. (who gets regular office visits from drug reps with samples and free lunch for everyone) turns into “there’s a pill for that…” I’ve since moved to a D.O. - he’s much less inclined to treat every hint of a complaint with an Rx.
Yeah, feels like almost a year since “prompt engineer” faded. In decades gone by people claimed skills in crafting Google search prompts…
When I was “engineering prompts” a year ago, it was absolutely necessary to prompt in certain ways to get better results, but… I could just feel it at the time, everything I was doing was blatantly obvious and the models were learning it almost as fast as I was prompting them - what I had to do one week, I didn’t have to do the next week because the scaffolding incorporated it for future sessions.
The whole thing gets messy when you’ve got too much prompt for the context window, so there’s still some “engineering” to do there, keeping your prompt-info organized so it can be navigated efficiently, not blowing through your context window budget before you get all the relevant information loaded before a task execution…
I was explicitly taught google search skills in nursing school. It focused on how to look for results from major medical journals, the CDC / FDA, articles written by well known MDs/DOs specialists, and how to find manufacturers’ manuals / YouTube tutorials for obscure medical devices. I was explicitly taught that my future was probably going to include looking up the manufacturer’s YouTube channel to find out how to access my cancer patient’s obscure chest port by a 60y/o instructor who clearly voiced how insane that was compared to her work experience.
For better or for worse AI’s gonna end up in class somehow too, I’m just hoping it’s late enough down the line that it’s been standardized to super basic shit like fuzzy matching search results and advanced grammar / spellcheck. One of our sister hospitals is trialing a service that can generate a flowsheet column for review from listening to a narration of a head-to-toe and I don’t immediately hate that. Actually you know what would be a really cool use? Fuzzy matching name alerts in the unit patient list. Had two obscurely spelled patient names that sounded very similar the other day with a matching last initial that didn’t trigger name alert highlighting on the board. An llm might be able to catch the sound-alike with a little more flexibility than the current algorithm.
I worked with a research oriented M.D. in the 90s, he had a friend die suddenly in his mid-60s and he made a project of figuring out what really happened to his friend. He was adept at PubMed searches, etc. and worked the problem for 6 weeks, about 4 weeks in he had a solid case for the misdiagnosis that led to the maltreatment that made his friend first worse, then within 72 hours dead. Still, he didn’t find any basis for malpractice (you can always file a lawsuit even with less than zero basis, but that wasn’t his purpose.)
At the end of it, he was left wondering what the solution is, because no one person can carry sufficient knowledge to make accurate diagnosis in all cases all the time, to make the right referrals even. He died in 2020, I bet the LLM agents (AI) of the past year are what he was searching for 30 years ago - it will take time for adoption and trust, but in the end I think we’re going to get better benefit of all that knowledge captured (currently all but buried) in the research publications. As long as power players like big pharma don’t turn the whole thing into a pill referral system - they way they corrupted my last GP’s practice. Any conversation with that M.D. (who gets regular office visits from drug reps with samples and free lunch for everyone) turns into “there’s a pill for that…” I’ve since moved to a D.O. - he’s much less inclined to treat every hint of a complaint with an Rx.