4 ms·

How does AI talk my boss and team into doing the obvious things they're not paying attention to? Having the right answer is a big help, but it's a fraction of

How does AI talk my boss and team into doing the obvious things they're not paying attention to? Having the right answer is a big help, but it's a fraction of the job.

windowsrookie 4y ago
In the future there won't be any team for the AI to talk to.
- RedCondor 4y ago
  Just like crypto.
- chongli 4y ago
  If you mean to imply that in the future it'll just be the boss sitting alone at his computer, interacting with the AI, then couldn't he still be more efficient by hiring a team of people to interact with the computer? After all, no matter how advanced the AI gets, the productivity of the company will still be limited by the speed at which the AI "operators" can interface with the machine. This is true regardless of the input choice: keyboard/mouse, voice input, or neural link.
  - Version467 4y ago
    Your argument defeats itself. If your AI is so good, that its productive output is bottlenecked by the speed of your supposed operators than you should just replace those with another instance of your ai, as well as the boss. Which is exactly the same scenario that your parent comment is presenting.
    - chongli 4y ago
      Ahhh yes, the fabled paperclip maximizer [1]. All skills become worth $0 in that scenario. It's a thought-terminating cliché at this point. [1] https://www.lesswrong.com/tag/squiggle-maximizer-formerly-paperclip-maximizer https://www.lesswrong.com/tag/squiggle-maximizer-formerly-pa...
      - Version467 4y ago
        Not sure how the paperclip-maximiser is relevant here. This still happens with friendly ai. If you accept the premise (all jobs replaced with ai, except for the boss, with no difference in productive output), then you can not solve the problem of all the lost jobs by rehiring them all as "operators". You said yourself that in this hypothetical scenario the bottleneck is human-ai communication (which is why you want to hire operators to increase productivity). But if human-ai communication speed is the bottleneck (and not ai capabilities, or compute, etc.) then you solve the bottleneck by replacing humans with ai, not by adding more humans. I don't think the original scenario is plausible (all jobs replaced except for the boss), so don't misunderstand my argument as defending that. I'm just saying that your conclusion (just hire operators) doesn't follow from the premise. It's not worth arguing this a lot further since I don't think the premise is plausible and we're talking about something that's only relevant if it were, unless I fatally misunderstood something you were saying?
        chongli 4y ago
        You said: just replace those with another instance of your ai, as well as the boss. I interpreted your statement as meaning to replace even the boss with AI, i.e. a fully-autonomous AI answerable to no one. That is just a paperclip-maximizer since it is no longer under human control. Perhaps you never meant for the boss to be replaced, in that case you may ignore my paperclip-maximizer statement.
        Version467 4y ago
        It is what I meant, I just don't see what the paperclip-maximiser has to do with it. As far as I understand it the primary idea behind that particular thought experiment is how a misaligned ai leads to agi ruin, even for simple goals. The scenario we talk about doesn't even contain misaligned ai. It contains friendly ai (the best case scenario), which still drops all current human economic value to zero. The contrived scenario has all jobs except for the boss replaced with ai. You propose hiring operators to increase the companies productivity. I say this doesn't make sense. Do you agree until this point? If so, what does the paperclip maximiser have to do with anything? If not, what did I misunderstand?
flangola7 4y ago
Your boss and team are also replaceable
Spivak 4y ago
Your boss and team will just ask the AI oracle, if it picks the right thing more often than the humans it's a win.
galangalalgol 4y ago
Chatgpt is way better at being persuasive than it is at being correct. It is probably already netter than you at convincing them of stuff. Probably not bad at spotting causes of organizational disfunction given symptoms, they seembto be the same ones over and over.
- xpe 4y ago
  > Chatgpt is way better at being persuasive than it is at being correct. I get the sentiment, but I doubt this is statistically true. Of course it makes mistakes, and certain kinds more than others, but ChatGPT’s recall is very good. > It is probably already netter than you at convincing them of stuff. I don’t think I agree, but I’m not clear what you mean. What do you mean? Care to unpack it? Here’s what I mean when I say I disagree: * ChatGPT, on the whole, doesn’t press an agenda. While it does have baked-in opposition to violence, hurting people, sabotage, bombs, etc, it does not (best I can tell) have a values system to weigh competing interests in a workplace. * Agency. This one may be obvious, but I think it needs to be stated. If you want ChatGPT to weigh in on a decision, you have to ask. You have to frame the question. An employee/contractor can come in with their own agenda and present a recommendation. * Much of the time (I’m not sure how often, this is an empirical question), when interacting with something like ChatGPT, people convince themselves. They frame the questions. They read answers. They stop at some point, perhaps when convinced, perhaps when the response seems incorrect, and/or perhaps when they run out of time or interest. Along the way, they can push back, asking follow-up questions or disagreeing. Of course, ChatGPT is presenting information and that information may be persuasive. But what “get’s in” to the human brain varies a lot based on what that person is receptive to. * Of course, a significant portion of Americans demonstrate limited critical thinking skills, for large portions of their waking life. One might debate this point, i.e. saying that sometimes just “going along with something” is a rational strategy to conserve energy, avoid conflict, fit in with a group, and so on. Whatever is going on, though, it goes appear that for many people, belief formation is not primarily driven by a quest for truth in the classical liberal sense. Such people often seek “advice” from ChatGPT not with the intent to question it, but rather to reduce the time and effort that thinking and researching requires.
  - 4y ago
pm215 4y ago
Your AI talks to their AI, so that the next time that they ask their AI, "hey, what should I be doing this morning/what's the most important change we should make to the codebase/etc?" it prompts them to do the thing you want them to be doing :-)
13years 4y ago
Your team won't be needed anymore. Your boss won't be needed anymore. Furthermore, whatever product or service your company provides, won't be needed anymore. This is the progression of AI. To be the provider of everything in the end. "Eventually you are just going to be in the way as a manager of AI tooling. AI will manage AI and learn from AI. No humans in the loop. No humans can keep up with the rate of advancement." - https://dakara.substack.com/p/ai-and-the-end-to-all-things https://dakara.substack.com/p/ai-and-the-end-to-all-things
Escapado 4y ago
I had this conversation with my girlfriend who is currently teaching herself web development without a formal programming background. She used ChatGPT to help her understand compiler errors or write small code snippets or react components and was terrified that it would write better code than her and seemingly understood things much better than she did. But here I sit, reflecting on my day-to-day: After the daily standup my PO calls me to ask if there is a chance that our external A/B testing tool can be coerced into tracking some arcane user flow in a slightly different way then the rest. After fiddling around with it for 30 minutes I conclude it can't but I have another idea on how they might get some of the data they want which I promptly put into words forwarding to the relevant people. I then get on to select a ticket from the backlog. Upon reading I notice the ticket is missing some info and the linked figma design is incomplete so I call someone from the design team to get some info and while we are at it we notice that the proposed solution is overly complicated and we settle on changing it - something that did not occur to anyone during refinement. Before I get to start coding the next teams meeting starts, this time to give us a brief of the strategy for the next quarter, and back to back the next refinement meeting starts. Half of the tickets have holes and missing edge cases in them and about 10% of them need to be outright discarded due to fundamental problems with our setup or architecture. We have lenghty discussions about minutiae of the user stories until finally the call comes to a close and I get some time to code. So back to my ticket. I open VSCode and I am greeted by our nextjs project: 850 components, 150 pages and 2200 unit tests and integrations with at least 12 different services in it - encoding the needs and wants of my corporate overlords and the history of it all. I get to identify where in our architecture the wanted feature fits best and think for a while of the exact composition of the components to fit our code style, best practices, testability requirements. I start scaffolding the first component when I see another call plop up. Someone from the testing team is confused why suddenly their locally initiated E2E tests against the dev stage fail. After 5 minutes of perusing logs it turns they didn't pull from main this morning and after doing that all was well. Back to coding. After scaffolding three more components I get to write a little more boilerplate and writing two or three utility functions to juggle around some data I proudly look at my code. Finishing touches and manual testing reveals that all is well. I write a storybook story for two of the components and think about which props should actually be controllable and mock out some of the involved api, write three little unit tests and one integration test to finally push my change to a newly minted feature branch. While the azure build and test pipeline slowly crawls towards completion I open our board, move the ticket to the code-review lane, put a small reminder into our dev teams channel and log my time in the time tracker. One hour of coding effort for ticket #31221. Good! While implementing my feature I noticed that a library we were using is out of date. A cursory search through the changelogs reveals that some of the breaking changes introduced in newer versions would break our app so I go on to write a technical ticket about that and put it in the backlog. Quick note to the PO that I got something for the next refinement and off I go to the next ...call. A bug in prod. Turns out one of our docker pods kicked the bucket. A quick restart and everything is fine again but we get into a lengthy discussion on whether we could finally migrate to some serverless architecture. I look at the clock. One more hour to go. Not really enough time to start a new ticket but looking at the board I find three tickets ready for code-review so that's what I do. One of them is from a less experienced dev so I try to be as thorough as I can and find a few things here and there which I promptly add as suggestions to the pull request. 10 minutes left. I sign off teams, close VSCode and close all tabs except for the time tracking one. I put in the last few entries and I am done for the day. I am not scared of AI taking my job just yet. Not as long as I have to defend technical choices in front of business people, implement the most outlandish user flows with obscure edge cases, have back and forth discussions with the design team on the difficulty or feasability of their longings, help QA people with their setups, debug devops problems, show off progress, write changelogs, estimate PDs for different stories and epics on different levels of granularity, onboard new developers, off board other developers, dance around in retrospectives, conjure up ideas and solutions for architectural re-orientations and so on. Actually writing some code is only <25% of my job anyway. And I am no staff level engineer or tech lead. Just a frontend engineer with some side knowledge. All the frontend devs on my team are like that. But maybe I am just working on weird projects. Who knows. Maybe one day I'll just orchestrate AI or have to re-orient but for now I am fine.

How does AI talk my boss and team into doing the obvious things they're not paying attention to? Having the right answer is a big help, but it's a fraction of the job.

windowsrookie 4y ago

In the future there won't be any team for the AI to talk to.

RedCondor 4y ago
Just like crypto.
chongli 4y ago
If you mean to imply that in the future it'll just be the boss sitting alone at his computer, interacting with the AI, then couldn't he still be more efficient by hiring a team of people to interact with the computer? After all, no matter how advanced the AI gets, the productivity of the company will still be limited by the speed at which the AI "operators" can interface with the machine. This is true regardless of the input choice: keyboard/mouse, voice input, or neural link.
- Version467 4y ago
  Your argument defeats itself. If your AI is so good, that its productive output is bottlenecked by the speed of your supposed operators than you should just replace those with another instance of your ai, as well as the boss. Which is exactly the same scenario that your parent comment is presenting.
  - chongli 4y ago
    Ahhh yes, the fabled paperclip maximizer [1]. All skills become worth $0 in that scenario. It's a thought-terminating cliché at this point. [1] https://www.lesswrong.com/tag/squiggle-maximizer-formerly-paperclip-maximizer https://www.lesswrong.com/tag/squiggle-maximizer-formerly-pa...
    - Version467 4y ago
      Not sure how the paperclip-maximiser is relevant here. This still happens with friendly ai. If you accept the premise (all jobs replaced with ai, except for the boss, with no difference in productive output), then you can not solve the problem of all the lost jobs by rehiring them all as "operators". You said yourself that in this hypothetical scenario the bottleneck is human-ai communication (which is why you want to hire operators to increase productivity). But if human-ai communication speed is the bottleneck (and not ai capabilities, or compute, etc.) then you solve the bottleneck by replacing humans with ai, not by adding more humans. I don't think the original scenario is plausible (all jobs replaced except for the boss), so don't misunderstand my argument as defending that. I'm just saying that your conclusion (just hire operators) doesn't follow from the premise. It's not worth arguing this a lot further since I don't think the premise is plausible and we're talking about something that's only relevant if it were, unless I fatally misunderstood something you were saying?
      - chongli 4y ago
        You said: just replace those with another instance of your ai, as well as the boss. I interpreted your statement as meaning to replace even the boss with AI, i.e. a fully-autonomous AI answerable to no one. That is just a paperclip-maximizer since it is no longer under human control. Perhaps you never meant for the boss to be replaced, in that case you may ignore my paperclip-maximizer statement.
        Version467 4y ago
        It is what I meant, I just don't see what the paperclip-maximiser has to do with it. As far as I understand it the primary idea behind that particular thought experiment is how a misaligned ai leads to agi ruin, even for simple goals. The scenario we talk about doesn't even contain misaligned ai. It contains friendly ai (the best case scenario), which still drops all current human economic value to zero. The contrived scenario has all jobs except for the boss replaced with ai. You propose hiring operators to increase the companies productivity. I say this doesn't make sense. Do you agree until this point? If so, what does the paperclip maximiser have to do with anything? If not, what did I misunderstand?

flangola7 4y ago

Your boss and team are also replaceable

Spivak 4y ago

Your boss and team will just ask the AI oracle, if it picks the right thing more often than the humans it's a win.

galangalalgol 4y ago

Chatgpt is way better at being persuasive than it is at being correct. It is probably already netter than you at convincing them of stuff. Probably not bad at spotting causes of organizational disfunction given symptoms, they seembto be the same ones over and over.

xpe 4y ago
> Chatgpt is way better at being persuasive than it is at being correct. I get the sentiment, but I doubt this is statistically true. Of course it makes mistakes, and certain kinds more than others, but ChatGPT’s recall is very good. > It is probably already netter than you at convincing them of stuff. I don’t think I agree, but I’m not clear what you mean. What do you mean? Care to unpack it? Here’s what I mean when I say I disagree: * ChatGPT, on the whole, doesn’t press an agenda. While it does have baked-in opposition to violence, hurting people, sabotage, bombs, etc, it does not (best I can tell) have a values system to weigh competing interests in a workplace. * Agency. This one may be obvious, but I think it needs to be stated. If you want ChatGPT to weigh in on a decision, you have to ask. You have to frame the question. An employee/contractor can come in with their own agenda and present a recommendation. * Much of the time (I’m not sure how often, this is an empirical question), when interacting with something like ChatGPT, people convince themselves. They frame the questions. They read answers. They stop at some point, perhaps when convinced, perhaps when the response seems incorrect, and/or perhaps when they run out of time or interest. Along the way, they can push back, asking follow-up questions or disagreeing. Of course, ChatGPT is presenting information and that information may be persuasive. But what “get’s in” to the human brain varies a lot based on what that person is receptive to. * Of course, a significant portion of Americans demonstrate limited critical thinking skills, for large portions of their waking life. One might debate this point, i.e. saying that sometimes just “going along with something” is a rational strategy to conserve energy, avoid conflict, fit in with a group, and so on. Whatever is going on, though, it goes appear that for many people, belief formation is not primarily driven by a quest for truth in the classical liberal sense. Such people often seek “advice” from ChatGPT not with the intent to question it, but rather to reduce the time and effort that thinking and researching requires.
- 4y ago

pm215 4y ago

Your AI talks to their AI, so that the next time that they ask their AI, "hey, what should I be doing this morning/what's the most important change we should make to the codebase/etc?" it prompts them to do the thing you want them to be doing :-)

13years 4y ago

Your team won't be needed anymore. Your boss won't be needed anymore. Furthermore, whatever product or service your company provides, won't be needed anymore. This is the progression of AI. To be the provider of everything in the end. "Eventually you are just going to be in the way as a manager of AI tooling. AI will manage AI and learn from AI. No humans in the loop. No humans can keep up with the rate of advancement." - https://dakara.substack.com/p/ai-and-the-end-to-all-things https://dakara.substack.com/p/ai-and-the-end-to-all-things

Escapado 4y ago

I had this conversation with my girlfriend who is currently teaching herself web development without a formal programming background. She used ChatGPT to help her understand compiler errors or write small code snippets or react components and was terrified that it would write better code than her and seemingly understood things much better than she did. But here I sit, reflecting on my day-to-day: After the daily standup my PO calls me to ask if there is a chance that our external A/B testing tool can be coerced into tracking some arcane user flow in a slightly different way then the rest. After fiddling around with it for 30 minutes I conclude it can't but I have another idea on how they might get some of the data they want which I promptly put into words forwarding to the relevant people. I then get on to select a ticket from the backlog. Upon reading I notice the ticket is missing some info and the linked figma design is incomplete so I call someone from the design team to get some info and while we are at it we notice that the proposed solution is overly complicated and we settle on changing it - something that did not occur to anyone during refinement. Before I get to start coding the next teams meeting starts, this time to give us a brief of the strategy for the next quarter, and back to back the next refinement meeting starts. Half of the tickets have holes and missing edge cases in them and about 10% of them need to be outright discarded due to fundamental problems with our setup or architecture. We have lenghty discussions about minutiae of the user stories until finally the call comes to a close and I get some time to code. So back to my ticket. I open VSCode and I am greeted by our nextjs project: 850 components, 150 pages and 2200 unit tests and integrations with at least 12 different services in it - encoding the needs and wants of my corporate overlords and the history of it all. I get to identify where in our architecture the wanted feature fits best and think for a while of the exact composition of the components to fit our code style, best practices, testability requirements. I start scaffolding the first component when I see another call plop up. Someone from the testing team is confused why suddenly their locally initiated E2E tests against the dev stage fail. After 5 minutes of perusing logs it turns they didn't pull from main this morning and after doing that all was well. Back to coding. After scaffolding three more components I get to write a little more boilerplate and writing two or three utility functions to juggle around some data I proudly look at my code. Finishing touches and manual testing reveals that all is well. I write a storybook story for two of the components and think about which props should actually be controllable and mock out some of the involved api, write three little unit tests and one integration test to finally push my change to a newly minted feature branch. While the azure build and test pipeline slowly crawls towards completion I open our board, move the ticket to the code-review lane, put a small reminder into our dev teams channel and log my time in the time tracker. One hour of coding effort for ticket #31221. Good! While implementing my feature I noticed that a library we were using is out of date. A cursory search through the changelogs reveals that some of the breaking changes introduced in newer versions would break our app so I go on to write a technical ticket about that and put it in the backlog. Quick note to the PO that I got something for the next refinement and off I go to the next ...call. A bug in prod. Turns out one of our docker pods kicked the bucket. A quick restart and everything is fine again but we get into a lengthy discussion on whether we could finally migrate to some serverless architecture. I look at the clock. One more hour to go. Not really enough time to start a new ticket but looking at the board I find three tickets ready for code-review so that's what I do. One of them is from a less experienced dev so I try to be as thorough as I can and find a few things here and there which I promptly add as suggestions to the pull request. 10 minutes left. I sign off teams, close VSCode and close all tabs except for the time tracking one. I put in the last few entries and I am done for the day. I am not scared of AI taking my job just yet. Not as long as I have to defend technical choices in front of business people, implement the most outlandish user flows with obscure edge cases, have back and forth discussions with the design team on the difficulty or feasability of their longings, help QA people with their setups, debug devops problems, show off progress, write changelogs, estimate PDs for different stories and epics on different levels of granularity, onboard new developers, off board other developers, dance around in retrospectives, conjure up ideas and solutions for architectural re-orientations and so on. Actually writing some code is only <25% of my job anyway. And I am no staff level engineer or tech lead. Just a frontend engineer with some side knowledge. All the frontend devs on my team are like that. But maybe I am just working on weird projects. Who knows. Maybe one day I'll just orchestrate AI or have to re-orient but for now I am fine.