6 ms·

This is the (near term) future of white collar jobs: AI does most of the work, but real people will still be required to audit and verify basically everything.

This is the (near term) future of white collar jobs: AI does most of the work, but real people will still be required to audit and verify basically everything. I don't know how much net gain productivity there will be...in certain professions an error rate that high is unacceptable and dangerous.

__loam 2y ago
Sounds like an awesome way for an entire professional class to slowly lose their skills.
- dcchambers 2y ago
  Yup...that's my fear. That and the fact that we're no longer training and leveling up juniors. Companies want to hire just one senior employee and let them use AI as they would a team of junior employees. But since we're no longer training juniors, we're no longer creating seniors. What happens when there's no one skilled enough to replace those senior employees when they leave?
  - __loam 2y ago
    Companies continuing the pattern of offloading training costs while complaining that it's hard to find skilled workers. LLMs are just an excuse.
  - Workaccount2 2y ago
    This is only really an issue if AI plateaus right around now. I think the consensus though is still pretty strongly that these cutting edge models we have now are going to be laughably bad in a few years.
romeros 2y ago
Similar to the self checkout process at Walmart/HEB etc. One person will be there for any issues and will intervene minimally.
- throw46365 2y ago
  17-33% is a whole lot less minimal than the person guiding a row of supermarket checkouts. In a busy supermarket they probably intervene less than 5%. But then the supermarket checkout risk is a bounded one -- there's only so much that can be stolen using a supermarket handbasket. Legal cases? Why does anyone think this approach can work?
banannaise 2y ago
The whole AI workflow sounds incredibly backwards - you give all these creative tasks to computers, which are notoriously bad at creative tasks, and then the drudgery of verification and review goes to humans, which are notoriously bad at verification and review.
- pradn 2y ago
  These aren't creative tasks. They're research and summarization tasks. No one is asking a legal AI to write an opera.
- yoyohello13 2y ago
  Expectation: Me doing art and being creative while robots do the dishes and cook dinner. Reality: Me doing dishes and cooking dinner while robots do art.
  - Jensson 2y ago
    Robots replaced most painters a century ago, its called cameras. So what you are talking about has already happened many times over and started a long time ago, its just easier to replace art since it doesn't have a precise requirement. Or see social media replacing news etc, similar situation. It turns out that manipulating our emotions at scale isn't that hard. Edit: Look at old things that were made, craftsmen put in extra work to make them beautiful. Modern things machine took all that work and produce standard things, machines took most of the art making that lots of people used to engage with already. Instead today humans mostly do paperwork. Edit2: Point was that technology always made us move from art to ever more tedious work, only exception was probably farming.
    - mym1990 2y ago
      My understanding was actually that technology enabled more time for leisure activities, one of those being art. Maybe not in the past few decades, but certainly as we through the Renaissance period and onwards.
pmarreck 2y ago
On the other hand, marriage has an error rate significantly higher than that and is still considered perfectly reasonable
- Jensson 2y ago
  So why not ask ChatGPT who to marry and follow that advice instead of your gut, its just random afterall! No, people really care about reducing error rates, it doesn't matter that some things has higher error rates, if you can reduce the chance of a bad marriage or a bad contract people really want to do that.
throw46365 2y ago
Audit, verify, but also modify. I don't know why anyone thinks that an AI that just trots out commonplace solutions to legal problems is something that most people even want. GPTs might be able to do boilerplate basics, but since it cannot actually understand nuance, it's not going to help with really specific things. Which makes you wonder why a non-LLM system -- an expert system, a database, some sort of DSL even -- isn't better than this stuff. Why, if there is rigour and precedent, is there not a solution that doesn't start from a foundation of waffle and imprecision?
- freejazz 2y ago
  There are all sorts of legal treatises that solve this problem
- jcranmer 2y ago
  > Why, if there is rigour and precedent, is there not a solution that doesn't start from a foundation of waffle and imprecision? Law is a fact-specific discipline, laden with jargon, and where the answers can change on a pretty frequent basis. It's not that far off the mark to say that the answer to every legal question is "it depends." What you need is a system that can take a moderately free-form explanation from a user, tease out the legal aspects from that explanation to match against the legal database (which almost certainly isn't going to use the same terms the user used!) to arrive at the answer, and in the inevitable scenario where the user failed to provide enough information, also be able to recognize that more information is needed to answer the question, and query the user (in their own terms) to get them to provide that information. LLMs are decently good at handling the language translation aspects, definitely, far better than any previous AI solutions; so it's not hard to see why people try this stuff with LLMs.
eviks 2y ago
Interesting, in what profession is this acceptable?
- glenstein 2y ago
  Well as it relates to humans, typically at an orthodontist office you have the one licensed orthodontist and any number of dental assistants, and the license orthodontist can oversee the work of the assistants. The assistants do the majority of the work but aren't credentialed, but because they're overseen by the main orthodontist we regard the work as rising to a standard we associate with professional legitimacy. I think the net result in these cases is that there's a lot more professional capacity to meet the needs of people who have braces. I don't suspect AI is coming for our teeth anytime soon but there's precedent for it as an oversight structure that satisfies us of the legitimacy of certain types of work.
  - eviks 2y ago
    Would dental assistants pull the wrong tooth 1 out of 6 times without supervision?
  - kjkjadksj 2y ago
    The assistants do stuff like flossing the braces and such but the ortho comes in to actually install and set them up. Its not purely oversight but a skill gap and what is possible given an amount of training.
swatcoder 2y ago
Industries do explore new products on hype, but they don't generally become widespread or disruptive until they can actually demonstrate some value. Companies are obliged or incentivized to start introducing products so that customers can see what's acheivable, but it's not yet obvious where generative AI products will be able to really take root and where they'll just be written off until some big innovation is made. And whatever happens, it seems very unlikely that it will apply to white collar jobs in general any time soon. It will apply to some industries, but others almost certainly won't be able to get what they need from what seems possible far -- or will only find use for it in certain narrow segments of their work.
Muromec 2y ago
And of course people will just not do the review part because it’s boring, low paying job and you can get away with subpar results majority of time
joe_the_user 2y ago
Indeed, If an organization doesn't care if their pronouncements are bs, they would opt for AI and thing have to definitely be right, they'd opt for a human. Logically, the jobs of lawyers are safe but the job of marketers are already looking doubtful. The worst situation is when an organization can calculate that producing bs won't hard them but where there are people still harmed by the bs.

__loam 2y ago

Sounds like an awesome way for an entire professional class to slowly lose their skills.

dcchambers 2y ago
Yup...that's my fear. That and the fact that we're no longer training and leveling up juniors. Companies want to hire just one senior employee and let them use AI as they would a team of junior employees. But since we're no longer training juniors, we're no longer creating seniors. What happens when there's no one skilled enough to replace those senior employees when they leave?
- __loam 2y ago
  Companies continuing the pattern of offloading training costs while complaining that it's hard to find skilled workers. LLMs are just an excuse.
- Workaccount2 2y ago
  This is only really an issue if AI plateaus right around now. I think the consensus though is still pretty strongly that these cutting edge models we have now are going to be laughably bad in a few years.

romeros 2y ago

Similar to the self checkout process at Walmart/HEB etc. One person will be there for any issues and will intervene minimally.

throw46365 2y ago
17-33% is a whole lot less minimal than the person guiding a row of supermarket checkouts. In a busy supermarket they probably intervene less than 5%. But then the supermarket checkout risk is a bounded one -- there's only so much that can be stolen using a supermarket handbasket. Legal cases? Why does anyone think this approach can work?

banannaise 2y ago

The whole AI workflow sounds incredibly backwards - you give all these creative tasks to computers, which are notoriously bad at creative tasks, and then the drudgery of verification and review goes to humans, which are notoriously bad at verification and review.

pradn 2y ago
These aren't creative tasks. They're research and summarization tasks. No one is asking a legal AI to write an opera.
yoyohello13 2y ago
Expectation: Me doing art and being creative while robots do the dishes and cook dinner. Reality: Me doing dishes and cooking dinner while robots do art.
- Jensson 2y ago
  Robots replaced most painters a century ago, its called cameras. So what you are talking about has already happened many times over and started a long time ago, its just easier to replace art since it doesn't have a precise requirement. Or see social media replacing news etc, similar situation. It turns out that manipulating our emotions at scale isn't that hard. Edit: Look at old things that were made, craftsmen put in extra work to make them beautiful. Modern things machine took all that work and produce standard things, machines took most of the art making that lots of people used to engage with already. Instead today humans mostly do paperwork. Edit2: Point was that technology always made us move from art to ever more tedious work, only exception was probably farming.
  - mym1990 2y ago
    My understanding was actually that technology enabled more time for leisure activities, one of those being art. Maybe not in the past few decades, but certainly as we through the Renaissance period and onwards.

pmarreck 2y ago

On the other hand, marriage has an error rate significantly higher than that and is still considered perfectly reasonable

Jensson 2y ago
So why not ask ChatGPT who to marry and follow that advice instead of your gut, its just random afterall! No, people really care about reducing error rates, it doesn't matter that some things has higher error rates, if you can reduce the chance of a bad marriage or a bad contract people really want to do that.

throw46365 2y ago

Audit, verify, but also modify. I don't know why anyone thinks that an AI that just trots out commonplace solutions to legal problems is something that most people even want. GPTs might be able to do boilerplate basics, but since it cannot actually understand nuance, it's not going to help with really specific things. Which makes you wonder why a non-LLM system -- an expert system, a database, some sort of DSL even -- isn't better than this stuff. Why, if there is rigour and precedent, is there not a solution that doesn't start from a foundation of waffle and imprecision?

freejazz 2y ago
There are all sorts of legal treatises that solve this problem
jcranmer 2y ago
> Why, if there is rigour and precedent, is there not a solution that doesn't start from a foundation of waffle and imprecision? Law is a fact-specific discipline, laden with jargon, and where the answers can change on a pretty frequent basis. It's not that far off the mark to say that the answer to every legal question is "it depends." What you need is a system that can take a moderately free-form explanation from a user, tease out the legal aspects from that explanation to match against the legal database (which almost certainly isn't going to use the same terms the user used!) to arrive at the answer, and in the inevitable scenario where the user failed to provide enough information, also be able to recognize that more information is needed to answer the question, and query the user (in their own terms) to get them to provide that information. LLMs are decently good at handling the language translation aspects, definitely, far better than any previous AI solutions; so it's not hard to see why people try this stuff with LLMs.

eviks 2y ago

Interesting, in what profession is this acceptable?

glenstein 2y ago
Well as it relates to humans, typically at an orthodontist office you have the one licensed orthodontist and any number of dental assistants, and the license orthodontist can oversee the work of the assistants. The assistants do the majority of the work but aren't credentialed, but because they're overseen by the main orthodontist we regard the work as rising to a standard we associate with professional legitimacy. I think the net result in these cases is that there's a lot more professional capacity to meet the needs of people who have braces. I don't suspect AI is coming for our teeth anytime soon but there's precedent for it as an oversight structure that satisfies us of the legitimacy of certain types of work.
- eviks 2y ago
  Would dental assistants pull the wrong tooth 1 out of 6 times without supervision?
- kjkjadksj 2y ago
  The assistants do stuff like flossing the braces and such but the ortho comes in to actually install and set them up. Its not purely oversight but a skill gap and what is possible given an amount of training.

swatcoder 2y ago

Industries do explore new products on hype, but they don't generally become widespread or disruptive until they can actually demonstrate some value. Companies are obliged or incentivized to start introducing products so that customers can see what's acheivable, but it's not yet obvious where generative AI products will be able to really take root and where they'll just be written off until some big innovation is made. And whatever happens, it seems very unlikely that it will apply to white collar jobs in general any time soon. It will apply to some industries, but others almost certainly won't be able to get what they need from what seems possible far -- or will only find use for it in certain narrow segments of their work.

Muromec 2y ago

And of course people will just not do the review part because it’s boring, low paying job and you can get away with subpar results majority of time

joe_the_user 2y ago

Indeed, If an organization doesn't care if their pronouncements are bs, they would opt for AI and thing have to definitely be right, they'd opt for a human. Logically, the jobs of lawyers are safe but the job of marketers are already looking doubtful. The worst situation is when an organization can calculate that producing bs won't hard them but where there are people still harmed by the bs.