How can A.I. and machine learning help with management?

[Accelerator]

So A.I., better and faster computers, and better machine learning. Better ads, better searches, machine generated procedures, protein folding, and everything else.

So what do you think is the future of management, coordination, and logistics when A.I. and computers are involved?

[sihnfahl]

So what do you think is the future of management, coordination, and logistics when A.I. and computers are involved?

No more than what a good data analysis does.

Walmart prioritizing foodstuffs like pop tarts, particularly strawberry, to areas in the predicted path of a hurricane, as an example.

Or getting optimum routes to deliver time-sensitive goods.

And good management isn't conducive to AI interaction. There's a personal component to management that AIs, at current, can't provide, and may never be able to because they lack relatability.

[NichG]

ML works well when you have a metric local to the task to optimize that actually measures what you really care about. Miles driven per delivery, percentage of stock wasted, etc.

With management applications, beware Goodhart's Law. Optimizing lines of code per day at the cost of bugs or employee churn or technical debt or just excessively wordy code could harm rather than help.

So step one is to figure out how to quantify what a manager should be accomplishing in a way that both directly ties to business success and generalizes across enough areas of business so you'd have the data to train it.

That said, I'd be tempted by an AI project that just learns when to cut off discussion in meetings and call for decision or moving on.

[Radar]

That said, I'd be tempted by an AI project that just learns when to cut off discussion in meetings and call for decision or moving on.

Shut up and take my money!

...is what I would say, if I had some venture capital burning my pockets.

I very much like the idea and it would be an interesting problem to solve. Namely, how to quantify discussion quality to predict the right moment? Is it actually possible to abstract something without understanding the whole conversation? I guess there would be some universal cues in the mimic and body language of the participants, but it would be difficult to prepare a learning set as we ourselves are not sure, when is the right moment to make up our minds on a given case.

[NichG]

Shut up and take my money!

...is what I would say, if I had some venture capital burning my pockets.

I very much like the idea and it would be an interesting problem to solve. Namely, how to quantify discussion quality to predict the right moment? Is it actually possible to abstract something without understanding the whole conversation? I guess there would be some universal cues in the mimic and body language of the participants, but it would be difficult to prepare a learning set as we ourselves are not sure, when is the right moment to make up our minds on a given case.

Pilot investigations would be to do something like take the meeting transcript and look at the dynamics of stuff like 'how well can you predict what will be said next, based on what has been said so far' or just generic embedding vectors (BERT embeddings, for example). If segments of discussion truly do tend to decay in utility and meaningfulness, you might find that the entropy of the conversation tends to decrease over time.

You could also do something like use eye tracking to collect participant attentiveness data and see if you can predict how attentive people will be based on conversation content, so you can generate a 'okay, lets move this to email' cue.

Those are both more hand-engineered metrics and so are likely to miss on really optimizing when you should actually cut things off though. I'm not sure where to get a reward signal or labels from meetings unless people actually go and annotate e.g. which parts of discussion corresponded to arguing in favor of the thing that was eventually decided. So if you can predict what the eventual decision will be after only 10% of the discussion length has passed...

[Radar]

Those are both more hand-engineered metrics and so are likely to miss on really optimizing when you should actually cut things off though. I'm not sure where to get a reward signal or labels from meetings unless people actually go and annotate e.g. which parts of discussion corresponded to arguing in favor of the thing that was eventually decided. So if you can predict what the eventual decision will be after only 10% of the discussion length has passed...

There is a problem there: even if the algorithm can predict, what the decision will be based on a cut out part of a conversation, the people might still need the following conversation to actually reach that decision as making them decide on the spot could give a different result. In other words, the algorithm could predict where the conversation is going and it is not the same as where current opinions of the people involved lie. There might be some inertia to the decision-making process.

So there needs to be a different metric available that can infer the changes in opinions of individual people in the meeting or how sure they are of their decision. This can be done for example by making people write down secretly and periodically this data: yes/no/abstain and how strongly they feel about it as a separate number. Single sliding scale might not be good enough as people might have a strong neutral stance just as well as a strong yes or no. This does affect the meeting a bit, but I do not thing it could be avoided.

Another interesting experiment would have people write down the exact moment, when they were convinced about their decision. Maybe there is some abstract cue in the eye movement or body language that can distinguish that or differentiate before and after from the same data.

There is an additional problem that for a commercial product the algorithm should require minimal data, so at best it should work from sound alone without the need of detailed visuals. Nevertheless, the same experiment ideas and conversation flow metrics will apply. Maybe there is something in the tone people use (very likely as it is connected with emotions), cadence, overall noise type and level etc.

I actually have a Ghostbusters moment here. Except I am sober right now, but this can be fixed in due time. I actually start thinking, if it is a marketable product.

[NichG]

There is a problem there: even if the algorithm can predict, what the decision will be based on a cut out part of a conversation, the people might still need the following conversation to actually reach that decision as making them decide on the spot could give a different result. In other words, the algorithm could predict where the conversation is going and it is not the same as where current opinions of the people involved lie. There might be some inertia to the decision-making process.

So there needs to be a different metric available that can infer the changes in opinions of individual people in the meeting or how sure they are of their decision. This can be done for example by making people write down secretly and periodically this data: yes/no/abstain and how strongly they feel about it as a separate number. Single sliding scale might not be good enough as people might have a strong neutral stance just as well as a strong yes or no. This does affect the meeting a bit, but I do not thing it could be avoided.

Another interesting experiment would have people write down the exact moment, when they were convinced about their decision. Maybe there is some abstract cue in the eye movement or body language that can distinguish that or differentiate before and after from the same data.

There is an additional problem that for a commercial product the algorithm should require minimal data, so at best it should work from sound alone without the need of detailed visuals. Nevertheless, the same experiment ideas and conversation flow metrics will apply. Maybe there is something in the tone people use (very likely as it is connected with emotions), cadence, overall noise type and level etc.

I actually have a Ghostbusters moment here. Except I am sober right now, but this can be fixed in due time. I actually start thinking, if it is a marketable product.

Sound only shouldn't be a problem. I'd even just transcribe to text and pass it through existing language models (since training one of those from scratch is a multi-million-dollar endeavor these days, and takes a noticeable fraction of humanity's entire text output to date as input).

As a rough guide, ML doesn't get off the ground with less than 5k to 10k examples of something. So if you need to have people annotate their meeting decisions before an initial product, it's necessary to cast a really wide net (e.g. enroll hundreds or thousands of participants) since otherwise it might be 5 or 10 years before you can actually get started. I don't think the machine learning aspects are all that difficult here - almost all of the work will be either the data collection and curation, or developing the product's software and hardware for market. The ML sounds like maybe a month of work for one engineer once the data is there. Depending where you have that work done, that'd be about $30k as contract work, less per month if you're hiring someone as a permanent employee but more total. I'm not sure about the software/hardware/interface development costs. The data collection is a big unknown for me to estimate though.

Some kind of staged release where the first users of v1 of the project also contribute the data that lets you train v2 of the product is probably the way to do it. So you have some imperfect metrics in the v1, but at least you aren't sitting still while you wait for the data to come in. Of course there are lots of privacy issues with anything that processes text, especially text at potentially sensitive meetings, so figuring out how to get customers to release the data to you for training is an important hurdle. I'd say something like federated learning could be used (you basically never send the raw data home, you just generate gradient updates on-site and send those), except that recent research shows that you can basically back out the training data from a trained language model if the model is big enough. So this would really only be privacy preserving if the model v2 is a fairly shallow or underparameterized network that is just doing post-processing on frozen language vectors. Anyhow, this issue is probably the one to really figure out how to solve if you actually wanted to do this.

[Radar]

Sound only shouldn't be a problem. I'd even just transcribe to text and pass it through existing language models (since training one of those from scratch is a multi-million-dollar endeavor these days, and takes a noticeable fraction of humanity's entire text output to date as input).

As a rough guide, ML doesn't get off the ground with less than 5k to 10k examples of something. So if you need to have people annotate their meeting decisions before an initial product, it's necessary to cast a really wide net (e.g. enroll hundreds or thousands of participants) since otherwise it might be 5 or 10 years before you can actually get started. I don't think the machine learning aspects are all that difficult here - almost all of the work will be either the data collection and curation, or developing the product's software and hardware for market. The ML sounds like maybe a month of work for one engineer once the data is there. Depending where you have that work done, that'd be about $30k as contract work, less per month if you're hiring someone as a permanent employee but more total. I'm not sure about the software/hardware/interface development costs. The data collection is a big unknown for me to estimate though.

Some kind of staged release where the first users of v1 of the project also contribute the data that lets you train v2 of the product is probably the way to do it. So you have some imperfect metrics in the v1, but at least you aren't sitting still while you wait for the data to come in. Of course there are lots of privacy issues with anything that processes text, especially text at potentially sensitive meetings, so figuring out how to get customers to release the data to you for training is an important hurdle. I'd say something like federated learning could be used (you basically never send the raw data home, you just generate gradient updates on-site and send those), except that recent research shows that you can basically back out the training data from a trained language model if the model is big enough. So this would really only be privacy preserving if the model v2 is a fairly shallow or underparameterized network that is just doing post-processing on frozen language vectors. Anyhow, this issue is probably the one to really figure out how to solve if you actually wanted to do this.

Even if you do not need the annotations (but I think they would help a lot as otherwise you have nothing to compare your predictions with), you still need the raw data and I doubt there are audio databases of meetings available anywhere. So you do need a lot of people to cooperate. The best option I guess would be to sell the pitch to a few multinational corporations, as they would be able to provide big and varied datasets easily. The downside might be that if they like the idea, they might want to develop it on their own.

Privacy would be indeed a problem - especially in the first phase of the program, where you try to isolate the key markers for determining, when to end a meeting. Actually, is it possible to extract those markers from a learned neural network, or are they impenetrable in that regard? The only other option I can think of is to try a few different neural networks at the same time: one full model for refernce and others focusing on specific aspects of the datasets (language entropy, noise level, voice tones, timing of when people say things etc.) or some simple combinations of those. If you have a large enough set of potential markers, you might even try to pick the best combination using genetic algorithms on your neural networks to pick the right input data. If such an approach would work, you could easily abstract the raw data on the spot without the risk of sending anything that could contain actual conversations.

Actually, there is yet another solution to privacy matters: encryption with public key for which you do not even have the private key. As long as the raw data is consistently using the same encryption pattern, it will not matter for the neural network to be honest. Well... there is a risk of making the learning process far more difficult as the data might resemble complete noise. So I am not sure about that.

[Bohandas]

It's definitely the future, just not the near future. As NichG observed, it probably won't do a good job habdling employees. Therefore, full implementation will have to wait until it's mainly managing other AIs.

When this point is reached all of the upper echelons of business will become redundant.

[Radar]

When this point is reached all of the upper echelons of business will become redundant.

Some would argue that we already reached that point.

[NichG]

I think that was Sihnfahl's point rather than mine.

I'm personally cynical that human managers actually do anything, but presuming there's some sort of direct metric to evaluate that I'd be willing to be convinced otherwise.

[Manga Shoggoth]

Besides, the correct phrase is "freed them up for other duties".

[TommyJo]

Some would argue that we already reached that point.

This is how much the human brain should be underestimated. To be sure that those attempts to create AI that is happening now can really replace at least one professional. At least one waiter. Or even a watchman. And you are talking about replacing people whose activities are completely focused on using the brain.

[Radar]

This is how much the human brain should be underestimated. To be sure that those attempts to create AI that is happening now can really replace at least one professional. At least one waiter. Or even a watchman. And you are talking about replacing people whose activities are completely focused on using the brain.

I guess you did not get the joke I was making there. I was quoting:

"When this point is reached all of the upper echelons of business will become redundant."

And the upper management is often depicted like this. So whether AI's are there or not, the point was that there are already a lot of redundant people in the management ladder. At least in popular opinion.

[kalkyrie]

...So what do you think is the future of management, coordination, and logistics when A.I. and computers are involved?

Logistics and coordination have already happened, to some degree.
The most up-to-date retailers had ML to predict sales at the product/store level a good 5+ years ago. If not 20.
Route planning is also a developed area- the Traveling Salesman Problem is now 90 years old.

There are some new techniques coming up of course. There are several companies focused around using camera data to estimate product quantities remotely (usually at the shelf-edge). IoT companies can detect equipment failures before they fail, etc.

Management is obviously harder. I can see two routes with this.
You could optimise the 'finer details' of it. Think Gantt charts, failure points of previous similar projects, etc. However the personal touch of good managers is not going to be replaced by an AI soon. A significant part of a good manager is helping your direct reports and pushing back against unreasonable requests.

Alternatively you could ignore the entire human side of management, and effectively have your employees reporting into an algorithm.
Think Amazon with their warehouse employees, or Uber with their taxi drivers.
'Pick X items an hour or be fired'. 'Pick up within Y minutes, or be effectively blacklisted'.
I'll leave it up to the reader to figure out the issues with that ;)

[NichG]

Logistics and coordination have already happened, to some degree.
The most up-to-date retailers had ML to predict sales at the product/store level a good 5+ years ago. If not 20.
Route planning is also a developed area- the Traveling Salesman Problem is now 90 years old.

There are some new techniques coming up of course. There are several companies focused around using camera data to estimate product quantities remotely (usually at the shelf-edge). IoT companies can detect equipment failures before they fail, etc.

Management is obviously harder. I can see two routes with this.
You could optimise the 'finer details' of it. Think Gantt charts, failure points of previous similar projects, etc. However the personal touch of good managers is not going to be replaced by an AI soon. A significant part of a good manager is helping your direct reports and pushing back against unreasonable requests.

Alternatively you could ignore the entire human side of management, and effectively have your employees reporting into an algorithm.
Think Amazon with their warehouse employees, or Uber with their taxi drivers.
'Pick X items an hour or be fired'. 'Pick up within Y minutes, or be effectively blacklisted'.
I'll leave it up to the reader to figure out the issues with that ;)

That said, if you automated the 'pushing back against unreasonable requests' thing it could be quite amusing, if not necessarily productive for business. In order for the CEO to set policy or demand projects, or for upper-level project leaders to add criteria, first they have to pass a totally automated screening system that basically just checks to make sure they're phrasing the requirements in a way that indicates they actually understand what it is they're asking for and are being clear and explicit about it. If not, the AI tells the upper level source of requests to go back to the drawing board and try again.

[Telok]

<ding>
"New message"
"Email analysis shows you are about to implement a 'hawiian shirt monday' policy. Currently 13% of indexed companies have a '[object] [clothing] [day]' policy. Analysis shows 8% greater valued employee turn over, 11% greater PTSD insurance claims, 23% greater 'HR is a bunch of bloody idiots', and 3% greater rates of management killed by employees. On average these companies exhibit a 4% decline in stock value over 5 years after the implementation of such policies.
Thank you for using BadManagerAI.com and have a nice day."

[Accelerator]

That said, if you automated the 'pushing back against unreasonable requests' thing it could be quite amusing, if not necessarily productive for business. In order for the CEO to set policy or demand projects, or for upper-level project leaders to add criteria, first they have to pass a totally automated screening system that basically just checks to make sure they're phrasing the requirements in a way that indicates they actually understand what it is they're asking for and are being clear and explicit about it. If not, the AI tells the upper level source of requests to go back to the drawing board and try again.

Like that yellow rubber duck thingy for IT?

[NichG]

Like that yellow rubber duck thingy for IT?

What is this thing? I mean, I know what a yellow rubber ducky is, but how is it used for IT?

[Radar]

What is this thing? I mean, I know what a yellow rubber ducky is, but how is it used for IT?

It is a problem-solving method. You take a rubber duck and try to explain your problem to it. In the process of explaining the issue you very often find a solution.

[NichG]

It is a problem-solving method. You take a rubber duck and try to explain your problem to it. In the process of explaining the issue you very often find a solution.

Ah okay, yes, that makes sense. Actually, AI sounding boards could be a really cool application since as you say, most of the benefit is by forcing you to express yourself. The AI would just have to have some better-than-chance rate of picking particular things to ask questions about or nod. Basically, fancier non-psychotherapeutic ELIZA.

[Tarmor]

I guess you did not get the joke I was making there. I was quoting:

"When this point is reached all of the upper echelons of business will become redundant."

And the upper management is often depicted like this. So whether AI's are there or not, the point was that there are already a lot of redundant people in the management ladder. At least in popular opinion.

When I want to complain about work, there's always Dilbert with exactly what I was thinking of. Although in this cases, I believe that the current state of the business (in regards at least one of the two managers/directors of the company) I work for has achieved that result with little more than two levels of management.

I firmly believe that an AI Manager would mean shorter meetings (if they occurred at all) and a lot less of the same questions being repeated over and over by the same people, either because they've forgotten or hope for a different answer. Hopefully, more things being done after something is raised the first time too.

[TommyJo]

When I want to complain about work, there's always Dilbert with exactly what I was thinking of. Although in this cases, I believe that the current state of the business (in regards at least one of the two managers/directors of the company) I work for has achieved that result with little more than two levels of management.

I firmly believe that an AI Manager would mean shorter meetings (if they occurred at all) and a lot less of the same questions being repeated over and over by the same people, either because they've forgotten or hope for a different answer. Hopefully, more things being done after something is raised the first time too.

Even the autopilot doesn't work very well. It's too early to think about AI being able to be creative.

[NichG]

Creativity is a much easier target for AI than autopilot. In autopilot if you make a significant mistake once in 1000 hours of operation you'll likely kill or injure a large portion of your users. In creativity applications, if you make certain kinds of mistakes with regularity it can be considered a personal style.

[Radar]

Creativity is a much easier target for AI than autopilot. In autopilot if you make a significant mistake once in 1000 hours of operation you'll likely kill or injure a large portion of your users. In creativity applications, if you make certain kinds of mistakes with regularity it can be considered a personal style.

Even more importantly, typical creative activity is iterative: you make something, evaluate and revise. So mistakes are part of the process. And you can actually train a neural network to solve for example specific engineering problems and the results are often very unorthodox. You can also teach a neural network to improvise jazz or other types of music. The example I linked worked on a simple laptop and used a generic neural network setup, so with a bit more resources and effort you could really make something interesting.

[NichG]

A recent example of something interesting you can do with more compute: https://openai.com/blog/dall-e/

The primary new feature here compared to caption-to-image stuff from a few years back is that this one seems to be able to compose attributes pretty well, including for example rendering somewhat arbitrary text in unusual contexts on request.