Comment by danjl
14 hours ago
Harsh take: AI should replace most middle management. It is the easiest part of an organization to replace. The people making things should mostly communicate about company strategy, cross-team issues, and job requirements with an AI. There should be a handful of high-level strategy on top of the AI. The AI should have access to all the documents for the company. The middle management should be put in a spaceship along with HR and sent off to another planet so the people who build things can just get stuff done. This will never happen.
I don't agree with this.
But I also found the article really unsatisfying. The idea that middle management should spend enormous amounts of time building relationships because other middle managers got vibes that one day it might be useful is insane. I think the article represents the worst of big, slow tech bureaucracies.
Replacing middle management with AI would not work, but using AI to avoid managers needing to have all these meetings would probably work really well. The idea that there's some AI system that has access to all the documents/email/task management systems at the company is a good one, and it could identify situations (like the one in the article) where two projects on opposite ends of the organization are colliding.
Instead of two middle managers needing to do 1:1s with no clear need for years because other middle managers got vibes that they should could be replaced by an AI system that uncovered situations like the ones mentioned in the article.
This wouldnt replace middle managers, but it might help them do their jobs better.
Adding AI to an organization that is somehow making process decisions based on "vibes" isn't going to solve problems, it's simply going to add yet another problem generator to a dysfunctional system.
I mostly agree. I'm curious to hear more details about why you think AI cannot replace middle management?
> The idea that middle management should spend enormous amounts of time building relationships because other middle managers got vibes that one day it might be useful is insane. I think the article represents the worst of big, slow tech bureaucracies.
If your org has anywhere north of 100 engineers across separate teams, intelligence gathering and relationship/trust-building is the only way to effectively do work that crosses the boundary of your team's area of responsibility. It's also the only way to protect your team from stepping headfirst into hot bullshit cooked up by clueless product managers, junior executives and other engineering teams who've unilaterally decided your area of responsibility is in their critical path.
> Instead of two middle managers needing to do 1:1s with no clear need for years because other middle managers got vibes
This isn't actually how this happens in practice. These 1:1s happen after their teams consistently have to share ownership over something or their work conflicts. It's more of a standup saying what your team is doing and what you're concerned about than a typical 1:1. You also calibrate the frequency as needed. For most of these it's a QBR but for some teams this will be monthly or even weekly. It's not "because vibes".
A key job of management is to figure out what's actually going on, as opposed to just what people tell you is going on.
LLMs are inherently gullible.
Human managers are inherently gullible; we've got no plausible path to unbias them. LLMs have at least one plausible path which is to train them to be a little bit cynical.
We're not going to call it "management" necessarily, but there is no question that LLMs are going to take over decision making from managers eventually. Why choose a monkey guessing what the evidence says you should do when you could have an optimised evidence-weighted statistical model making the bets? The only reason to use humans is there are still technical limits on how general the models are, limits that seem to be falling away at a pleasing rate.
> Human managers are inherently gullible; we've got no plausible path to unbias them. LLMs have at least one plausible path which is to train them to be a little bit cynical.
Firm disagree on claims 2 and 3 (paths to unbias each), though I agree humans (managers included) are inherently gullible.
There's a lot of research into human biases and how to overcome (or at least mitigate) those biases; and one can in principle always hire a "no man" to look for things which can go wrong. This is kinda what corporate lawyers (and, I hear, corporate economists) are there for.
AI, unfortunately, have a weakness which isn't present in meat-based intelligence, one which won't go away even if we get brain-uploads to copy meat-minds into silicon to make better AI: the very fact of being cheap enables us to find their weaknesses by spamming a bajillion variations at them to see what slips past their cognitive blind-spots.
Unfortunately, my take on the second paragraph is even more cynical than yours:
> We're not going to call it "management" necessarily, but there is no question that LLMs are going to take over decision making from managers eventually. Why choose a monkey guessing what the evidence says you should do when you could have an optimised evidence-weighted statistical model making the bets? The only reason to use humans is there are still technical limits on how general the models are, limits that seem to be falling away at a pleasing rate.
We're already seeing LLMs take over decision making from managers, not because they're good in the "optimised evidence-weighted statistical model" sense, but because they're good in the "hyper-persuasive to lazy primate brain" sense.
This also shows the limits of the "hire a no-man" strategy, as this is happening despite the list of people saying "aaaaa this is dangerous!" including many of the people developing these particular AI models, along with some Nobel laureates, various campaign groups marching around with placards, and a bestselling book.
Interesting that humans can't be trained to improve ("be less biased") but AI can. I would say this is a much more damning conclusion for the AI replacing ICs than managers.
Whats easier, training AI to make good bets (what does that mean in business? Make the most money? Worker quality of life? World a better place?) or training it to get code to compile?
If we can just simulate the business accurately enough we can solve having to interact with the market… which is also trying to solve interacting with us… We just need to do it more accurately…
Something tells me people will still be in the mix here.
> "optimised evidence-weighted statistical model"
Isn't such a model inevitably going to be lagging what is happening?
(Monkeys can see/smell/recognise the scat or track of a large cat very quickly and don't sit around to check the data)
Firm, but partial, disagreement.
People, especially in remote jobs, benefit from being organized into groups intentionally, with distinct rituals that enable them to operate effectively while they get to know each other better. Another person needs to design and oversee all that.
While you can provide templates for that structure that allow oversight to scale so that one person can oversee larger groups, that tends to be more effective in non-remote, and more predictable, work environments. Modern software development is very little of that.
I don't have much in-person experience with middle management in contexts outside of software development, and I suspect there are some opportunities to use AI to bring engineers closer to customers.
> with distinct rituals
If you do rituals you'll get ritual compliance, not people getting to know each other better.
AI Orchestrated Waffle Parties
I have a theory. How close does the following describe you?
* You're an engineer with 3-6 years of experience in a primarily IC role
* Maybe you've done some tech lead stuff, but you've never actively worked in engineering management.
* You feel that management (and HR for some reason?) is constantly in the way of you getting stuff done, and that your life would be easier if you could simply decline every meeting and only communicate through pull requests.
Humor me, please. I'll explain after.
Middle-management as we knew it at the turn of the 2010s is probably gone forever. You don't need to coordinate many many teams as you used to. Same as huge frontend team with dedicated support for graphql, etc. AI made most of that redundant.
By extension we're going to need a lot less middle managers as coordination problems decrease.
As for the point I think you're trying to make, the problem with middle management and other chokepoints in general (like PM teams) is that often they become an antipattern. They soak in all the information and then dole it out parsimoniously, so the typical experience as an IC is to be barely able to see the full picture
[dead]
For the sake of moving this along, that describes me perfectly. Please, continue.
Management advocates for AI because see ICs as commodities that just need to be coordinated to "do the thing." In this situation, remove the mid-managers, replace the ICs with AI, and use AI to enable them to coordinate the "workers." They forget that organizations exist to organize human output, which requires nuance, empathy, and communication.
ICs advocate for AI because they believe they are "doing the most valuable work." A rational AI would see that and let them do it. In this situation, remove the mid management, replace HR/marketing/sales/etc with AI and use AI to enable them to figure out what to build and they build twice as fast. They forget that the "rational" choice might not be what is best for them, their project, or their career.
Each one rebuts this with the way the system has failed them (managers feel that workers do everything BUT the work that moves the company forward, ICs feel like they can do everything BUT the work that moves the company forward)
I have 35+ years experience as a manager and engineer at large enterprise tech companies (what the kids now call FAANG, though some of the company names were different back then), and was a Founder, CEO and CTO at a $7M VC funded company and several other "differently-funded" startups.
Ha, excellent. I think you'll appreciate this then: https://imgflip.com/i/aswdth
I had a 50/50 chance of guessing which side of the bell curve you were on and I was wrong. :)
1 reply →
Plot twist!
Couple decades in with some leading but still remaining an IC officially.
The right work just doesn't get done without staff engineers and architect types having frequent conversations & meetings as well as constant code reviewing. Or long-term ICs effectively doing this role without the title and expectations/responsibilities (raises hand). You can identify these people because they ask questions relentlessly. Always well-considered ones, but even the ones that might make them look stupid in a meeting.
Coding is a small percentage of the work but also just as important. That's the sweet spot. The "non-stop meetings/socializing" people and the "headphones on & grind PRs" types are both two extremes of behavior that are boat-anchors in any organization and will bring productivity/customer-impact to a screeching halt if it goes unchecked for long enough.
And it's _always_ those stupid-seeming questions that uncover showstopping problems that would have bit you if left ignored.
Edit: Not to greenlight anything Palantir is doing, but in my opinion the FDE/FDSE model is probably everyone's near-future if your company is B2B. You can't be an "ignore meetings" type of person and do that.
It’s a trap.