Managers decided AI is worth 5x speedup; how do I explain to them how it really works?
Posted by chaitanyathengdi@reddit | ExperiencedDevs | View on Reddit | 282 comments
So I am starting this new project that is doing some work with agentic AI. It's fairly boring work, but it's not mechanical (e.g. generating reports or something). It's something that requires a bit of research per task.
A bunch of managers decided that they want to achieve 5x speedup in the work using agents. So for example, if it would take a team of say 5 people 2 months to complete the entire thing, the same people would use a bunch of prompt engineering to do the work in 12 days.
How they got the 5x figure, I don't know. Is it achievable? Don't know either. Is any speedup achievable? Maybe.
How do I tell them that they should not assume a target at the start and just go with what results the team can get? Forcing an arbitrary goal on people will lead to burnout and I want to communicate that.
brainhack3r@reddit
It has to play out unfortunately. There's too much market pressure right now. It's better to just engage in malicious compliance.
JaySocials671@reddit
So you’re saying prove that it doesn’t work by sabotaging the process?
brainhack3r@reddit
Sort of fair but I'm saying just do your job, not try to rescue the process by working extra to solve the problems that this can create.
JaySocials671@reddit
Why are you advocating malice?
_Heathcliff_@reddit
This is the way. Send your token usage through the roof. Let them see how comically expensive this is for lower quality results. Executives are like children in that they don’t believe the stove is hot, they have to touch it themselves.
BroBroMate@reddit
My company is measuring us on PR merge rate, they expect an average of 1/day, and also measuring us token burn - if you're spending under $200 per month, you get a please explain.
So if you want to play it like that, fine. I'm now top of the table in both PR merges and token spend.
Gru50m3@reddit
Holy shit, that is so fucking dumb. Do they even care what the PR is or do they just want the illusion of being productive?
bluetista1988@reddit
My experience with companies that use access to data as a form of micromanagement is that they don't care what you're doing just as long as you're doing a lot of it and quickly in a way that makes their lines go up. You can deliver absolutely nothing of value, but if they see their coveted "hockey-stick growth" (is that phrase used outside of Canada?) they will applaud you.
BroBroMate@reddit
They want to see that we're all delivering super fast now that we all have 10x superpowers (an executive literally said that in a meeting...)
My team lead had a wee chat with me based on my metrics of "only" delivering 1 PR a week, so fine, this feature branch is now 10 independently shippable branches, sure it won't deliver any value until the 10th one lands, but they're not measuring value added...
Goodhart's Law in action.
I told my lead I was doing so, he admitted he's doing the same.
yxhuvud@reddit
I'd recommend splitting larger PRs into smaller. They are easier to review anyhow.
BroBroMate@reddit
Who said anything about a large PR? You've assumed that 1 week = huge amount of code, why?
I've always practiced "break stories down into independently deliverable units that provide value".
And if anything, this current approach actually increases cognitive burden on reviewers because you need to read the changes in PR#5 in the context of the changes in PR#2.
PlushyGuitarstrings@reddit
What do you mean reviewers? You mean ai agents, right? More cognitive load is good! The more we spending tokens, the more we save! Doublethink was yesterday, today we Quadthink!
BroBroMate@reddit
And here I was doing Audi Quattrothink. It gets me around those awkward mental corners super quick!
dezsiszabi@reddit
The more you buy the more you save - Jensen Huang
yxhuvud@reddit
I assumed that because that is what i see in real life. Sure, the odd task can be a weeks job and then a small change, but all actual examples I've seen of people with the habit of regularly doing pull requests at that interval have had severe issues with being horrible at splitting things up. You may be the shining star that is an exception to this, but I'm still going to assume you are not.
BroBroMate@reddit
I spend a week on a PR when it needs one because I work in a complex domains where thinking and verifying behaviour before releasing is important. Because if we break prod, we lose customer trust, and that's easy to squander and very hard to earn.
I'm still doing that thinking and verifying, but now I'm merging the stuff that won't break things but delivers no value either as I go along.
Shit, I once took 3 weeks to deliver a PR that added 10 lines of code and removed 200, and significantly increased throughput and customer satisfaction. That was a good PR.
Let me leave you with a favourite quote from the father of nuclear physics:
Properly understanding the problem domain to ensure you don't do the wrong thing is greatly undervalued by people who do stupid stuff like measure merge rates as a proxy.
Move fast and break things can work in some markets, but definitely is a liability in others.
Just because you and I work in the same broad "typing spells to convince lightning trapped in fancy rocks to do things" field doesn't mean your subjective experience is objectively applicable to me, and likewise, mine isn't applicable to your work.
Please do me the service of keeping that in mind.
PlushyGuitarstrings@reddit
You are the kind of dev I like to work with. Cheers mate!
BroBroMate@reddit
Danke!
Whisky-Toad@reddit
Thankfully I have a CTO that knows if you make stupid rules like that we learn to just game it
BroBroMate@reddit
I wish my C-suite were that smart.
Suspicious_Stable_25@reddit
You should be doing this anyway. No one wants to review a 1000 line diff PR
Minimonium@reddit
It's impressive to be so up your own ass to instantly assume 1k+ diffs, instead of the more obvious scenario where they split micro PRs into one liners because the metric is the number of PR, not that they make sense individually.
BroBroMate@reddit
I already do, but my main focus has been primarily on breaking things down into independently shippable value.
Cheers for the condescension though!
Ysilla@reddit
AI brought back all the dumbest metrics we can imagine somehow, it's insane. I've seen merge counts being used so often recently, but even mentions of lines of codes.
Guitarzero123@reddit
In a meeting last week we were told that our AI spending was through the roof, but our PR merges had significantly dropped off...
Almost like laying off a bunch of devs reduced the amount of code we are able to output and review each week or something...
pheonixblade9@reddit
Huh, I'm in the several thousand per month spending range these days 🤣
brainhack3r@reddit
To be fair... AI can do a lot of things on auto-pilot now but most of these things you'd expect it to be able to do.
I love AI for doing this stuff.
It just can't replace the whole team which is what theyw ant.
chickadee-guy@reddit
Its also the only scenario where LLMs insane price point would be worth it, so the AI execs have to pitch full replacement
Just-Ad3485@reddit
They aren’t that expensive yet
A premium Claude seat is like 125 Canadian and you get a ton of usage (until anthropic changes things)
Minimonium@reddit
I have seen people burning through that plan in an hour, lol
guareber@reddit
I have it on good authority that change is coming in June for enterprise clients.
paradoxxxicall@reddit
It’s expensive for investors, who expect an even larger financial burden to eventually fall onto customers.
binkstagram@reddit
So the model is step 1: replace engineers then step 2: increase prices to something actually profitable now they are dependent on you
Tritondreyja@reddit
Yep
dbenc@reddit
no, more like they have someone else touch the stove until they find someone who tells them it's not hot
daddywookie@reddit
And everyone who said it was hot is now on the shit list for not being a “team player” or having “the right attitude”.
devfuckedup@reddit
huh , been wondering why people have been making a mess burning insane amounts of tokens. This exsplains it. Thanks.
tigerking615@reddit
It’s really easy to fire off a million tasks at once, not pay attention to any of them, and not do any of them well.
I’m feeling pretty comfortable on the $20/month Claude Pro plan which seems to be decent for all my needs. On a heavy usage month where I’m too lazy for session limits to reset I might end up spending $5 in extra usage.
nullpotato@reddit
At home I carefully craft prompts to be more efficient. At work they want us to use AI so tokens go brrr
normalmighty@reddit
Yeah, at work we're explicitly encouraged to try to hit our usage limits constantly.
oditogre@reddit
We don't have usage limits; we're on the AWS Bedrock thing so it's just a pay-as-you-go model.
There's a DataDog dashboard that shows who is using what models and what their usage is in dollars.
It's almost treated like a leaderboard - it's not "who is costing the company the most money", it's "who is leveraging AI the hardest, like we asked them to." It's insane, but whatever, if that's the metric, I can definitely pump those numbers. 🤷♂️
rum-n-ass@reddit
We don’t even have usage limits at my job, just expected to ball out on AI all day long
alerusey@reddit
i've just spent 700$ this month in the company account hahaha
corny_horse@reddit
I can't figure out if the Gastown folks are in this camp or unironically think its a good idea https://github.com/gastownhall/gastown
yxhuvud@reddit
I think they are serious (and insane). It's either that or a very cute art project.
SlaminSammons@reddit
Make sure to always use the most expensive model too. Doesn't matter what you are doing just make sure to charge the shit out of them.
Sunstorm84@reddit
And run at least 6 agents to verify each other all using the most expensive model
hdkaoskd@reddit
That's the key to 5×!
MeAndTheSatan@reddit
Love the last line !
QueSeraShoganai@reddit
That's a great analogy.
Ok-ChildHooOd@reddit
Second this. Let it play out. The blame if the quality is poor falls on the LLM and not you, which is way better.
turntablecheck12@reddit
It absolutely should, but it won't. I know the place I work is gung-ho on everyone using AI, while also explicitly stating that you're responsible for anything you commit.
empiricalis@reddit
lol I doubt it, the engineer will 100% be blamed and told that they aren’t prompting hard enough
donniedarko5555@reddit
Honestly do your "5x speedup" its totally achieveable to vibe out a PR and dump it without testing.
You'll have job security cleaning up the tech debt once the hype train dies
InterestingBoard67@reddit
this phrase will become quite common in the future to come.
prumf@reddit
Or no job because clients left 🤷
MaleficentCow8513@reddit
“We don’t need this project anymore because it has no user base”…. Yea because you told me to vibe code the whole thing and it has bugs and general sucks as a product
Jaded-Asparagus-2260@reddit
And management that demands a speedup of 5x by using "AI" is probably not the kind of people that talk to users, have market research, and invest in discovering how to actually improve customer value.
So many people confounding output with outcome.
circusboy@reddit
This is what I've been doing. Ill admit I've been having A LOT of cool things happening, but that token usage is through the roof! Fine, no junior devs to train? Ill use the fuck out of the tokens so you wish you didn't fire my team. I got to 633% as of EOD today, and i was off a full week at the beginning of the month. Fuck em!
big-papito@reddit
Yeah, just keep generating clunker code until crushing amounts of technical debt.
brainhack3r@reddit
Don't review it... make sure it passes functional tests. then just accept the PR.
big-papito@reddit
I mean, that makes sense, but I will hate my life when I have to build a feature on top of that, or debug it.
newmanP@reddit
You mean when the AI agent has to build a feature on top of it :)
big-papito@reddit
Yes, but at some point the agent will start running in circles with too much spaghetti context.
nullpotato@reddit
In my experience it just starts recreating things all over the place when this happens. Why are there 12 log creating helper functions? Because those prompts didn't look for an existing one first. The opposite of DRY, and nobody wants moist code.
ChemicalRascal@reddit
It's important to remember that the prompts don't look for code. The model was presumably trained on a lot of code that has helper functions in the same file as their use, and so it would be weighted towards "log stuff -> use helper function -> write a helper function at the top of the file".
Which, to be clear, isn't an excuse, but it demonstrates that these systems aren't intelligent. Lots of textbook examples or StackOverflow examples of that sort of code would have those helpers defined near the code body using them, possibly with a note saying the helper function would actually be in some common namespace; but the LLM doesn't actually read, it's just gonna associate all those words together and yada yada helper functions everywhere.
newmanP@reddit
Absolutely - if you don't have the influence to stop a bad idea, you just have to play out the full scenario for management. Ideally you do this in a way that demonstrates how bad the idea is while giving them the impression that you are bought in.
FeliusSeptimus@reddit
Change the color of the home page to a purple-ish diagonal gradient with a subtle random gridwork of animated lines
Sorry you're on your own buddy.
nullpotato@reddit
If the goal is more token usage, then have it review the PRs too. Hell have it review everyones PR so you get credit
systemnate@reddit
Let's say that you hypothetically could develop the code 5X faster using AI. That does not mean the entire project will get shipped 5X faster. Because there is more to completing a project than just writing the code.
scoopydidit@reddit
Managers will soon realise the bottleneck in projects is usually the stuff they enforce: meetings, approvals, workflows with managers in the loop, etc.
There is some bottlenecks that engineers also still face: designing and RFC aspects.
But the speed bottleneck has rarely been the code. I feel like that's usually pretty quick.
So now that we've a tool that sometimes makes the fast part of the process a little faster... management thinks everything else should be a lot faster. Not how it works I'm afraid. And sometimes this tool also makes things worse and slower. Lol.
RelevantJackWhite@reddit
my managers are quickly realizing that we have not spent enough time planning or validating, historically, and now it's burning us as we try to ship code faster
Opposite-Hat-4747@reddit
At least yours are realizing that
maverickzero_@reddit
This is the thing, any organic pitfalls become unmanageable when you hit them at 5x the rate by churning out code 5x as quickly.
Let alone any inorganic pitfalls coming from generated code.
itsmegoddamnit@reddit
“We’ll ship it like this and learn from it and pivot as necessary”
mikkolukas@reddit
😄
false_tautology@reddit
Send AIs to meetings and you'll sane 75% of your time!
chaitanyathengdi@reddit (OP)
Or 100 'cause I'm out
Business_Average1303@reddit
exactly
and AI usage should not be viewed as 5x features, but 5x between more quality, scope and features? definitely. And what you mention can also be improved with AI, it just needs to be understood as a tool not only for devs
HelenDeservedBetter@reddit
This is the part that I've been communicating to my manager. In my opinion, coding makes the easiest parts of my job easier, but that doesn't come close to even doubling my productivity.
mpanase@reddit
Think about all the shit unused features you had to implement. Are the people pushing for a lot of those feature the same ones expecting 5x speed out of this?
Now imagine they think the cost to adding these features is negligible.
RobertKerans@reddit
Writing the code is very rarely the bottleneck (I hesitate to say never, but I don't think it would be unreasonable)
SplendidPunkinButter@reddit
AI advocates like to say AI coding “shifts the burden to review.” It’s okay that a non-deterministic, hallucinating AI generated your code, because you’re going to carefully review it! So one would expect that with AI coding, reviews now take longer than they used to.
Nope. Your manager expects reviews to be faster too.
This does not end well or you care about the software not being crap/
scoopydidit@reddit
I mean the issue there is in the past... when a human generated code and another human reviewed, if there was a bug, the primary blame would be the person who wrote the code and the reviewer might get some questions about how they missed it.
Now... since there's no human generating, the human is expected to somehow thoroughly review this AI code and ensure it has zero bugs (never happens really, across hundreds of lines), and if there is a bug, the reviewer is now the primary blame.
boop809@reddit
Wait, people actually read the code during code reviews?
bluetista1988@reddit
The four greatest letters in the modern Latin alphabet:
l g t m
failsafe-author@reddit
It does shift the burden to review. And companies that don’t take the time to do it are going to crash and burn when they can’t manage the complexity of their apps.
It’s also why 10x isn’t achievable. Imagine telling people 5 years ago they could get 2x or 3x (which I think is achievable for a skilled developer). But now that’s not enough.
daishi55@reddit
Are you under the impression that humans are deterministic?
Buttleston@reddit
amateur. AI can do your reviews too! And write your tickets! And write documentation and release reports!
Curious_Owl197@reddit
And replace your managers and ceo too
shill_420@reddit
Hey, scheduled a meeting with you me and hr to talk about that attitude problem
Curious_Owl197@reddit
Just lemme change my profile pic to a middle finger real quick
trippypantsforlife@reddit
Now this I like
purleyboy@reddit
See Blitzy
dzendian@reddit
https://i.imgflip.com/2/1hhv9m.jpg
subma-fuckin-rine@reddit
heh yep. make huge code review no human could do quickly, ai reviews it and leaves comments, another ai address the comments, merge. what could go wrong...
Aleks_Zemz_1111@reddit
Management pulling a 5x speedup number out of thin air is exactly how factory managers treat the production line.
At my day job, they used to put one person on each Moeller machine. Then management decided one person could operate two machines at once. The 'efficiency gain' was just them forcing a guy to physically run back and forth between them. One guy twisted his knee doing it, and when he asked for compensation, the manager denied it, saying it was his own fault for running and ignoring health and safety.
They don't know where the 5x figure came from because it's a spreadsheet fantasy, not a real metric. If you try to talk to them about burnout, you will lose. They don't care. I've seen guys get replaced a week after dying of a heart attack. You are just a number to them.
If you accept this target, the AI isn't going to magically work 5 times faster. You will just be doing 5 times the donkey work, cleaning up the agent's hallucinations and fixing bad logic to hit their arbitrary deadline. You'll be the guy running between two machines.
Don't argue about burnout. Argue about liability. Tell them a forced 5x artificial speedup will introduce latent errors into the architecture that will cost them ten times more to fix down the line. Speak their language costs and risks.
NPPraxis@reddit
I’d say something like this:
“My daily work isn’t all writing code. It includes meetings, determining specifications, reading and reviewing written code to validate, documentation, negotiating specifications with vendors, project management tracking and support, coordinating with other developers, addressing build pipelines and security, etc.
If 50% of my time is this type of support and 50% is actually writing code - even in the most optimum case (let’s imagine writing code is 5x faster, which it’s not) - it would only free up 40% of my time. Which I can fill with more of that support.
Will I get more done? Absolutely. But even a 5x increase in coding speed will not result in a 2x increase in output, because so much of my job isn’t writing code. The people selling AI like to focus on the coding speed up, and not the rest.”
Unfortunately, if your manager is particularly craven, they’ll tell you to figure out how to get AI to do the communication part for you too. Don’t ask me how I know.
Savings-Giraffe-4007@reddit
Don't say anything.
Just abuse your token usage as much as possible, tun out of tokens and demand more tokens to your managee, when asked why just tell them anything below Claude 4.7 is too dumb to hit their required metrics and keeps making mistakes.
These morons only understand when the bill is higher than expected.
mpanase@reddit
Your managers are not ambitious enough.
Mine were talking about 10x to 40x.
They are scared shitless about having any public documentation of the product, because somebody could copy it using AI in a single night.
I'm just encouraging them to keep build agentic workflows and show us shit demos with the most generic examples ever. Let them sink.
In the meantime, I'm in the second stage of 2 interview processes.
CardinalM1@reddit
Here's the CEO of Optum Insight talking about 100X. Yes, 100!
https://www.fiercehealthcare.com/health-tech/unitedhealth-taps-sandeep-dadlani-lead-optum-insight-unit
mpanase@reddit
Yes, let slop it up in healthcare
Healthcare and aviation, please
Maybe even nuclear plants. Why not? 100x the nuclear power!
MoneyStatistician311@reddit
What makes you think is it any different in the place you are interviewing? Or that they will not do a total 180 and start requiring you to do the same shit?
mpanase@reddit
I now this is how it is in the current one, don't I?
ButWhatIfPotato@reddit
I genuinely worry that shit like this will one day make me snap. How many times do I need to explain to people that letting me do my job will benefit anybody and me not doing my job because you weaponized your ignorance to show how big your corporate pp is will benefit nobody.
yerfdog1935@reddit
My last supervisor had the gall to say he could do the entire sprint we were assigned (across 5 developers) in a single afternoon.
chaitanyathengdi@reddit (OP)
"Sure!"
ImpossibleEbb6862@reddit
Ask them to show you how.
Basic-Lobster3603@reddit
I heard this today. With agentic AI we did 100+ cards in 4 days.
6158675309@reddit
Someone should have assigned all the Jira tasks to him 😂
Gru50m3@reddit
Yeah, that's a put-up-or-shut-up moment.
TheTacoInquisition@reddit
Tell them to be less harsh about their contributions to the company, because if all it takes to beat them at their own business is AI copying the product, then they must be doing absolutely nothing useful.
mpanase@reddit
Oh, they are literally ask in every demo: "who is stopping our competitors from using this tool to copy all our products in a weekend?"
The answer: "Nothing. That's why we have to do it"
JohnDillermand2@reddit
Everyone focuses on competitors when you should be more concerned about your own customers.
knowwho@reddit
We were told to hit 2 PRs per day, and that 10 PRs per day should be possible by the end of this year.
SearchAtlantis@reddit
Literally fed Claude Code the ticket and which section of the code-base (final data warehouse layer) and it burned 40% of my daily tokens to give me an incorrect but adjacent answer.
FeliusSeptimus@reddit
Next time put it on that 8x Claude Opus model and let the company pay the bill!
MolassesCharacter664@reddit
sounds like they're setting up for unrealistic expectations without understanding the tech's limitations
mpanase@reddit
100%
Absolutely mental expectations.
It's the CTO and 2 product guys producing demo apps from scratch again and again. Not a single engineer involved, not even one try to use it with our existing products.
Other engineers have given up on explaining.
I never did. Hearing them speak made it clear to me that they need to crash and burn before they consider listening to anybody.
Last presentation of a vibe-coded demo involved the whole company, including a CEO (who was smart enough to say "exiting stuff" and shush) and the CMO who was essentially asking "oh, so I could create a product by myself?"
Most probable outcome IMHO: layoffs, vibe code an absolute shitshow, call a couple proven devs to handle the fire. I rather skip the shitshow, so I'm betting on being one of those called to sort it at a ridiculous daily rate.
Watchful1@reddit
The really funny thing here is that you just replied to an AI bot
mpanase@reddit
That's fine.
I don't discriminate.
I only judge the ideas.
linkardtankard@reddit
Let them create their own shit and get burned in the process, it’s the only way to break the irrational hype and inject some common sense into their brains
mpanase@reddit
I encourage them to do it.
I tell them it's not gonna work, but I encourage them to do it.
If they burn fast, they won't destroy too much in the process. If they burn slow... I suspect they will destroy the company in the process.
2053_Traveler@reddit
Demo apps were never the hard part. Step one: find some open source project a student or intern built. Step two: clone and follow instructions.
How can a technical person be so ignorant of demo vs enterprise app that actually works every time, stays up, can serve more than one request at a time, and handles surprises?
mpanase@reddit
That's the key.
They are not technical.
And engineers telling them this ruins their sales pitch.
They need to spend tens of thousands in token and fail miserably.
Future_Salad652@reddit
i had something like this with a 10x expectation once
mpanase@reddit
I ran an experiment and asked multiple AI engines what they though about their approach.
All of them told me that it was unreallistic and destined to fail.
Manager will be damn easy to replace, and actually improve in the process. Skynet, here we come, and it's the managers' fault.
Watchful1@reddit
This one is also an AI bot too
metaphorm@reddit
in my org, we've achieved about a 2x total throughput gain through utilizing a very built out custom agent harness with claude code. it's extremely hard to push it further than that in terms of throughput because there are bottlenecks that aren't strictly coding related.
MeAndTheSatan@reddit
What is the bill ? How many tokens are you burning
metaphorm@reddit
token spend is significant, but it's still a whole order of magnitude less than hiring twice as many engineers
MeAndTheSatan@reddit
What happens when AI isn't as subsidised as it is now ?
Impossible-Bake3866@reddit
You explain it by getting a new job
Ahhmyface@reddit
I work in an analysis heavy domain. Not a single person in the company could tell me what code needs to be written. They can give vague generalizations only. It's all in the data we can't find, and there's far too much to sift through. I can't use AI to sift through something if I don't know what I'm looking for yet.
Worse, it is complex enough that I need subject matter experts to look at it to explain intent to me, so I can explain it to AI. Then I need more subject matter experts to review the AI output.
This takes for fucking ever, but the gods still decree I should be 5x faster because I'm using AI to write the code. Not even engineering leaders seem to grasp that most of us write code like we write fucking English, and the problems we solve didn't get any easier.
rosstafarien@reddit
You can mentor devs in using AI tools to increase velocity, but you're going to be spending several weeks per dev, losing ~25% of the mentor's time. Some of those devs will turn out be good mentors. Others you'll want to let go.
You could see incremental team velocity improvements within a few weeks, but a 100% improvement could take a dedicated quarter and some luck to achieve. Once you've got CI/CD improvements in place to handle your faster devs on your code, you could go faster than that.
jimmytoan@reddit
The 5x number almost certainly came from a vendor demo or a Gartner-style projection, not a controlled study on your specific type of work. For agentic research tasks specifically, the bottleneck shifts from writing to verification - you spend the time you saved on generation double-checking what the agent produced for correctness and hallucinations. Real productivity gain on these tasks is usually 1.5-2x on a good day with high variance. Track actual token costs and time-on-verification from week one - that's the data you need when the arbitrary deadline becomes a conversation.
daishi55@reddit
I don’t get it. Are you against the idea of setting goals? Have you asked them how they came up with 5x number? Honestly even if they pulled it out of thin air I don’t see the issue. Did they say there will be consequences if not met? Or rewards if it is? It feels like you’re leaving out a lot of information and maybe just whining because you don’t like AI.
Free-Huckleberry-965@reddit
lol just quit now, you will never, never, explain cognitive burnout to upper management. They fully understand what they're doing, and they relish is ruining your mental health. Acknowledging that you know what they're doing will only make it worse on yourself
QueSeraShoganai@reddit
True. My CEO admitted to overworking IT intentionally to see what they were capable of.
ultraDross@reddit
I do believe they would admit that.
QueSeraShoganai@reddit
I don't really care what you believe. It happened. He spun it as how amazing IT did, and that he knew it was a tough year because they intentionally pushed the limits to test our capabilities. The tone-deaf statement happened.
chaitanyathengdi@reddit (OP)
Wow, that's dark
FitNerdDude@reddit
Hes not wrong. Harder you work. The more value they extract out of you. You get laid the same. They get rich
nullpotato@reddit
I mean, we are definitely getting fucked over
FitNerdDude@reddit
Yup
We're typically the smartest in the room.
Until it comes to our own best interests 😂😭😂
Imagine if SWEs came together and didn't log in for a week across the world.
abandonplanetearth@reddit
I genuinely feel bad for you and the ppl that upvoted you for never having worked with good managers.
Free-Huckleberry-965@reddit
I've had a good manager before. Who do you think told me this truth? She did.
abandonplanetearth@reddit
Oh. I thought I was reading your opinion.
My bad I guess 🙄
Free-Huckleberry-965@reddit
No, it's lived experience unfortunately. If you haven't encountered it yet, I consider you incredibly lucky.
Ok-Chair-7320@reddit
He should find another job (before) quit
FUSe@reddit
You can only work as fast as your slowest pipeline
authentic_developer@reddit
The hardest part of this situation isn't the 5x claim - it's that the number is unfalsifiable without a baseline. If your team has no tracked velocity right now, there's no denominator. You can't hit 5x or miss it. Which means you'll be evaluated on vibes, and that never goes well.
The move that's worked for me: don't push back on the number, propose a measurement framework first. Pick 3 representative tasks, do them with and without AI assistance, time both, then let the data set the expectation. Managers who believe in AI will trust a number they helped generate far more than a number you argued against. And if the pilot shows 1.8x on research-heavy work, that's now their number, not yours to defend.
The specific problem with research tasks is that AI speedup is extremely uneven by task type. Boilerplate and scaffolding — maybe 5x is real. Research where the model synthesizes confidently wrong information — verification cost often exceeds the time saved. Worth scoping the claim to task category before committing to anything.
_katarin@reddit
In my experience, testing ai's output is slower than doing unit tests on regular business logic.
alanbdee@reddit
Our team has not seen that kind of speed up. It has broadened what we're able to do. We can venture into new territory quicker. It has sped us up. Some things could be done by a factor of 5x or more. But when you still consider time to review it, user testing, etc. It's not that much faster overall.
mxldevs@reddit
Management is going to ask why this couldn't be replaced by AI too.
Curious_Owl197@reddit
Turn it around and ask why can't ai replace managers too? Chatgpt surely can spit out 100s of stupid questions faster than they can
painedHacker@reddit
I mean places are trying.. wasnt amazon saying they were aiming for 50 engineers per manager or something ?
Curious_Owl197@reddit
Now we only need half the number of ceos
Few-Impact3986@reddit
Just have AI do the communication and start sending emails with questions.
GregsWorld@reddit
"Nine women cannot make a baby in one month"
JorgJorgJorg@reddit
they will tell you to update your thinking and that old adages no longer apply. There is no reasoning with people who start by choosing the conclusion.
FlashBrightStar@reddit
Their thinking is "la la la I don't hear you" every time you warn them about potential issues. Those people are amazed by technology they don't understand because it's a magic box that does everything they do but better and faster. Because of this they think every position can use it the same way to "boost" productivity. It's basically a new version of 9 women delivering a baby in 1 month on steroids - chat, you are a super woman with high repro skills, make a baby, make no mistake, don't fuck up the dna history.
EbbFlow14@reddit
Our CEO somehow got Rust on his radar after someone told him about it. The greatest programming language, it's safe, fast,... You know the deal.
Anyhow, one day he spoke to me about rewriting the entire stack in Rust because it seemed a good idea to him. After all, Rust is the safest and fastest language.
I tried to give some counterarguments about cost and having to train engineers. Rust is a very hard language to learn, not impossible, but the on ramp for a developer is long.
Nope, he wouldn't listen, then AI came along and he lost interest in Rust, now everything is AI.
Sometimes I question why exactly I am doing this job. Every couple of years some hyped up thing spawns into existence and non technical people get boners after they read an article or something telling them how great it is. I hate it...
chaitanyathengdi@reddit (OP)
Worse, it's marketed to them that way. And they believe it.
SkepticCritic@reddit
I want to interject that it also uses the existing codebase as reference, so you're also speed running the Habsburg family line too.
mpanase@reddit
But what if I give all those 9 women a very ellaborate plan and 24/7 access to 9 doctors?
r0Lf@reddit
...and full access to opus?
FeliusSeptimus@reddit
Yeah, somebody needs to read The Mythical Man Month.
chaitanyathengdi@reddit (OP)
I wanted to use this in the question but it's already such a cliché
creaturefeature16@reddit
Turning an oven to 500 won't cook the turkey in half the time.
1000Ditto@reddit
all-in on the agentic workflows is like nasa throwing the turkey into the sun and hoping it's cooked in 10 seconds
creaturefeature16@reddit
Indeed. I wrote a pretty in depth article about Agentic Coding being a self made trap.
YeastyWingedGiglet@reddit
Love this article btw. Sent it to some coworkers and friends who all agreed!
creaturefeature16@reddit
Thank you so much! I've been getting a lot of great feedback on it. I didn't expect it to resonate as deeply as it seems to be, which is really awesome to see.
Scottz0rz@reddit
Actually, my AI orchestration layer agentic workflow kicked off 16 subagents each with their own air fryer convection oven skill to partition the turkey into equivalent parts which allows us to achieve maximum throughput with 16 ovens cooking 1/16th of a turkey which the orchestration advisor agentic superflow can hyperflux them together to recompose the turkey into a composite.
I made up half those words I think, but I think it would fool a VC / suit to prove you wrong.
insidious_concern@reddit
But agentic gestation...
azuredrg@reddit
However, Thragg making babies with Thraxans can pump out subpar Viltrumites at 100x the rate vs making babies with humans.
4dr14n31t0r@reddit
Came here to say exactly this
bbrizzi@reddit
When management/sales give a bullshit metric, the best clapback is to return a different, more correct metric, with technical justification.
ie : before AI, we completed on average 10 points of tickets per person per day, with AI we complete 14 points so it's a 40% increase. Also there a significant quality drop (+30% bug creation) and overhead to create tooling to support the new AI pipeline (new code review process, etc...)
This means the productivity increase with AI is +20%, not +400%
This may require to comply with the policy for now and come back with actual numbers in a couple of months.
mikkolukas@reddit
Seems like an excellent opportunity for manglement to walk in front and show how to do the job, as they clearly have more experience doing this /s
czlowiek4888@reddit
Just ask them what you supposed to do to achieve this result and then just do it.
thecrius@reddit
I read this recently: The Human Cost of Working at Machine Speed https://medium.com/@cvaleman/the-human-cost-of-working-at-machine-speed-de47c23b216c
Honestly, you will have to let it play out and hope that your management will realise it's not sustainable.
On your side you have the announced increases cost of agents so you can factor that as well it's a postmortem about the project.
Bemteb@reddit
Everything can be delivered 10x faster with AI, including burnout.
LeDYoM@reddit
"The mythical man month" is must read more than ever.
ThisIsProbablyFine1@reddit
Just keep spending. Don’t ask me questions.
rcls0053@reddit
Same idea as 9 women can give birth in one month.
LuckyWriter1292@reddit
If they won’t listen then find a new job..
HalcyonicStorm@reddit
https://koushikdasika.com/blog/coding-was-never-the-hard-part/
kagato87@reddit
Sounds like they bought the sales line. It's a real problem.
The only things where it approaches 5x for me are well-trodden methods that are new to me. It can cut through the seo garbage to some extent and at least answer "can I..." questions.
But most of my work is, as others are saying, planning and strategizing. It wants to jump straight to solutions using patterns that will explode or don't even match the spec it just finished drawing up.
Then I have to review its code, correct it (it's REALLY bad for sql anti patterns), then remind it that we need to performance test all the viable methods because we're doing this on a scale where milliseconds add up to hours.
wisconsinbrowntoen@reddit
I migrated hundreds of tests from one library to another. It took me a workday to run the agents and a week to review the changes. If I did it myself it would have taken 2-3 months to migrate all those files
kagato87@reddit
Yup. Simple, repetitive tasks, it's (usually) good at. Especially if you can just diff the old and new files to see what changed.
originalchronoguy@reddit
It is doable but it is $$$$$.
When I showed them 10 agents running concurrently and $50/hour billing accruing in real-time, they are starting to see everything is currently subsidized. That is why you have Claude Code Max resetting every 2-5 hours.
Want NO interruptions? It is gonna cost. And cost BIGLY!
It is actually pretty exciting but who has the budget? I can spend $2k a day with a large orchestrated setup. Feed the plans and it runs all night 12-hours.
it is gonna be interesting how this goes. But man, it is sure fun seeing how much it costs in real-time for unfettered access.
wisconsinbrowntoen@reddit
Future jobs:
"Hey I might not be as good or as fast as Claude 10 Turbo but I'm 1/10th the cost for the same result"
ryancoplen@reddit
I spent 6 hours today fighting with a new AI process that adds comments to CRs. The particular CR is spent the most time on was a bunch of markdown documentation. The AI tool kept flagging trivial issues and then I had to update the CR and wait another 5 minutes and it would come up with new super trivial issues.
I think I went through literally 8 or 9 revisions before I was able to actually publish the review for my team mates to actually review.
One or two of the comments were actually useful updates. But otherwise the AI review tool ended up costing me like 4 or 5 hours today. I’d call that negative productivity.
BlazingThunder30@reddit
If they are willing to listen to reason then start by showing them the research. You're not the first one to start using AI which means that there's been a fair amount of research in how much speedup you actually get and based on what circumstances.
feverzsj@reddit
Sadly, you can't do shit. See what happened to those poor Amazon seniors.
AdEquivalent1092@reddit
Data data data.. Track your work.. out of all that you do where does ai actually fit in.. what percent of your overall work os that Then add 5x to that percent.. Worked for us in setting right expectations
hw999@reddit
you say, "damn, that's a great idea boss", then proceed to burn through thousands of dollars in claude credits. that will teach them.
FarYam3061@reddit
If you can't explain it maybe you're wrong
magichronx@reddit
Ask them if they think 9 women can make a baby in 1 month
knowitallz@reddit
At most 50% increase in productivity for repetitive tasks. But thoughtful ones. Nope
SwaeTech@reddit
Well…I mean. They’re not wrong. But that’s assuming you weren’t already using AI as a copilot to begin with. If you explain that you’re already using AI to speed things up, they may back off a little. And if not, then a senior developer needs to put their foot down. If that also fails, then just consider it a wash, be maliciously compliant and seriously look for a new job.
N0y0ucreateusername@reddit
You thank your stars they only wanted 5x and try to figure out what to do with the other 80% of your day. We have ai now.
Banquet-Beer@reddit
Let It blow up in their faces.
alerusey@reddit
managers only understand when the deadlines are not met, they dont listen to reason
pigeon768@reddit
I'm a US citizen in the US. If you're in Europe or in India or are an H1B in the US, you will have to adjust.
chrisrrawr@reddit
10x it and pocket the difference of course.
shill_420@reddit
If they redefine “the work” to not include the inevitable forever march of production issues and support tickets then this technically works.
I don’t get the point but it works. It’s what my shop moved to.
anengineerandacat@reddit
Depends on the task really, we have been integrating agentic coding and research into out team for over a year on a 5+ million dollar project.
Currently seeing about a 11-30% increase in team velocity but it dips because not all processes and members are using the tools (dev team is, but delivery and preparedness for projects involves more than the dev team the SDLC involves various people of different technical skill levels and exposure to AI tools).
Across the year, we have accelerated the delivery by about 70% but now we are in this weird waterfall model waiting for QA to legit test everything which is slowly sucking back that efficiency gain.
5x IMHO does not seem achievable for an enterprise environment considering your stated conditions.
Do you have all the project requirements? Are all the stories groomed? Multi-sprint plan or Kanban board setup? CI/CD self service? Infrastructure self service?
All of that has to get out of the way and effectively let dev teams sling code like it's a personal project with a testing team involved.
Veera_MMAI@reddit
Totally get the pressure.
I’d frame it as “let’s benchmark first, then set realistic targets.”
ConsiderationSea1347@reddit
“You know how when you type a word into your phone and you know it is the right word and spellcheck still converts it into Ke$ha? That is what using AI to code is like.”
AdUnlucky9870@reddit
the funniest part is when they quietly revise it down to 2x after the first sprint and act like that was always the number
RPND@reddit
so you find how they're measuring the 5x,
then you game that metric.
Unironically, this shouldn't keep you from actually doing stuff that's worthwhile for your company
titogruul@reddit
It's ok. Once you figure out how to actually take advantage of this and integrate it into your workflows, they will hit you with "oh... It's coating us 50k, why don't you let off a little bit and try something cheaper".
phonyfakeorreal@reddit
My company also went “all in” on AI recently. Leadership is expecting 2-3x velocity based on AI-assisted speed on a few cherry picked examples of AI-conducive work, which of course isn’t representative of the work we’re usually doing. We achieved it… through Goodhart’s law. We’re inflating story points, we’ve stopped doing anything not strictly in-scope without a new ticket (even if it means leaving things in a broken state), rubber stamping PRs, all that good stuff. Unfortunately, I work at a company where executive leadership only cares about a chart going up, and you get fired if you dare question them. Everyone seems to have an unwritten understanding of this, so managers encourage it, directors turn a blind eye, and VPs package it nicely.
keptfrozen@reddit
Fast, Cheap, Good — they need to realize that they can’t have all three.
Atraac@reddit
I would just tell them do it themselves if agents are doing it anyway. They will take entire responsibility of maintaining and extending whatever they build. Hell, I'd encourage them to get even more agents. If one agent can give birth to a child in 9 months imagine what 9 agents will be able to do! Part of being senior in the team is learning to tell people no. This is that moment.
Pyran@reddit
That's a good way for them to say "Good idea. Good luck in your next job."
Do it, collect the paycheck for as long as you can, look for a new job in the meantime. Don't give them another reason to force you out.
LtRodFarva@reddit
You can cook a turkey at 325F in the oven for 3 hours. To a manager, this means you can also cook a turkey at 975F in the oven for 20 minutes. That’s it. That’s the logic.
chaitanyathengdi@reddit (OP)
Why not throw it in the sun? It'll get cooked in less than a second (and never be seen again, but who will emphasize that?)
Pyran@reddit
One second to cook, 8 months to put it into the oven. Poor ROI.
(For the record: that's the kind of answer I'd expect, not the one I'd give.)
Business_Average1303@reddit
The ones pushing 5x or more are simply very experienced who can actually get a lot of things faster vs without AI, because with the right prompts/knowledge beforehand, you can guide the LLM to achieve outputs much faster using the right tools and configurations which also take time
The problem is that without the experience it’s harder to get full advantage of the speed up for every team member, and don’t even mention how everyone uses differently without knowing impact of some little but relevant things. As an example some months ago someone showed me how he was using non-english language to interact with the models, and as we all know, they are probabilistic models, the data they were trained is mostly in one language and if you prompt it in a different one you can get probabilities skewed towards answers that might also be in that language narrowing down to maybe not the most appropriate answer, and just like this, there are Many many more other small details that adds up
If you know what a non functional requirement is, and you request this in a prompt vs just ask the LLM to give you a PRD, it may or may not give you that info in an answer, adding extra time in the process to get understanding between what you want vs how you want it, adding an extra back and forth of the document between business and technical roles which makes things much slower than the actual time it takes to correct because of death times between these back-n-forth rebounds
forbiddenknowledg3@reddit
Fucking annoying when I give an estimate (obviously at this point assuming we're going to use AI) and management cut it in 1/4. So 1 month = 1 week now.
unrebigulator@reddit
https://en.wikipedia.org/wiki/Goodhart%27s_law
"When a measure becomes a target, it ceases to be a good measure"
Sometimes AI gives me a 5x boost. Sometimes it doesn't. Sometimes it slows me down.
maverickzero_@reddit
I think you can be to the point: you DO know that a 5x speedup is not possible, and you are the domain expert.
All you can do is advise though, they're probably just going to do it anyway and fuck up the project then act confused.
sarhoshamiral@reddit
I once had a manager that put out impossible goals like that but I quickly learned that it was his style and he actually realized the goal isnt really achievable.
What happens if you dont get 5x gains? What happened in the past with similar demands?
zugzwangister@reddit
5x speed up for a new project seems doable.
With Claude code, I suspect a couple knowledgeable people could make that project in a month if it used to take 5 people two months.
The biggest question is rework not from poor code quality but from not understanding all the requirements.
Jolly-Leather6836@reddit
could you elaborate on what "you don't" means here
mxldevs@reddit
I don't think it's worth fighting.
If you don't meet the targets, they assume you're just a shitty dev and lay you off the next guy that will promise to finish it yesterday.
If you do meet the targets, they'll brag about how great managers they are and ask for 10x next time.
MercyEndures@reddit
What kind of research?
The puzzle may be to figure out why the AI fails on these research tasks, unblock it, and let them run in parallel.
bobbyturtle@reddit
Our CIO keeps talking about 100x…
kbielefe@reddit
First of all, targets are management's problem. Especially if they didn't get consensus from you before setting the target, just do your best and don't worry about it. Easier said than done, I know.
Your details are intentionally vague, but I could see it working under some combination of the following:
thephotoman@reddit
You can’t.
Managers are mostly overgrown toddlers. Most do not know what they’re talking about. This is especially true when talking about tooling. My boss occasionally references Windows and Office developments from the last 20 years, and I have to stop and tell him to explain it in terms of things I use, because that’s when I switched to Mac. (I switched to Python and LaTeX and even Linux earlier because of school, and I’m more comfortable with those tools now than I am with Word or Excel. Don’t make me PowerPoint. You wouldn’t like me when I PowerPoint. If something compels me to use PowerPoint, everybody is gonna have a bad day from the sheer VOLUME of infodumping with citations that’s gonna happen.)
lxpnh98_2@reddit
Just say 10 man-months.
zeke780@reddit
Don't listen to anyone here about quitting or fighting your bosses. Collect a check. Do whatever you are now. Upskill where needed and try to make ai work for you.
There is way too much money being dumped into AI for it to go quietly. Boards are heavily invested, CEOs are heavily invested, 401ks are heavily invested, it's literally propping up the economy right now.
Just keep doing what you are doing and reassess in 6 months, hopefully more realistic expectations exist by then
airemy_lin@reddit
Big tech and the rich are foaming at the mouth hoping this will do what bootcamps flooding the market did a few decades ago and finish the job but broad spectrum across all roles.
So yeah it’s not going away.
distractotron9000@reddit
Even without AI stakeholders seem to think parallelism perfectly scales infinitely. Two engineers two months? Just put 120 engineers on it and we can have it tomorrow!
Forsaken_Celery8197@reddit
Maybe you should give them a prompt preference snip they can add to their own AI. Something like:
My engineers are real people; when coming up with estimates, be conservative and assume work takes a little longer than the best case scenario most of the time.
call_Back_Function@reddit
Tell them that coding is between 10% and 15% of your time. The rest is understanding, coordinating, meeting and everything else. Then ask if ai makes writing documents faster he should be able to manager 5x the employees. That’s how it works right?
Dimencia@reddit
Just google the latest statistics. Last I heard, Google sent some marketing folks to do a presentation for our company and told us about how they measured up to 5% increased productivity with AI usage. Yknow, in the best case scenario that they cherry picked from their example data
another_dudeman@reddit
These fucks are seeing how much they can squeeze you. I wish you the best of luck.
caboosetp@reddit
You let it burn. Then in the post mortem, you recommend they read mythical man month. The same shit in there still applies to having AI as a pair codecoder.r
siscia@reddit
I can definitely see, right now, how the speed up to produce code is there.
As everyone will point out, that was never the bottleneck township anything.
Coordination, understanding the codebase, attempts that don't plan out, those are still there and those are still the bottleneck.
However, my suggestion would it be to explain to management that the speed up is for a single developer, not for a team. In a team setting those stuff above are the main contributors to a slow down.
So instead of aiming to move from 5 developer to develop in 2 month to 5 developer developing in 12 days, you should aim to move to 1 developer developing in 2 months. The single developer is faster, the team itself it is not.
And this speedup is much more realistic.
andupotorac@reddit
Im doing 50x more. So yes its doable.
__natty__@reddit
You want to be passive aggressive? Then tell him ai is so good he now may try to do some basic jira management. And show him some browser based ai. He knows it’s terrible so then you have option to show him dev perspective in a words he understands.
If you want to be polite then the answer is much harder and he needs to see in practice that even high usage and spending of tokens doesn’t guarantee 5x speed up. Based on the research it barely touches the same pace of 1x. You can try to explain difference between green field and typical easy problem vs brown field and old codebase changes across multiple repo.
slowd@reddit
I’m a believer, but you all need to dive in and be pros at pushing your model and harness to the limit. It’s not effortless and it’s not instant. They need to at least understand that it’s a new tool and it has a ramp up. Driving a car for many hours does not make a race driver — you need to specifically try to maximize what it can do for you.
andymaclean19@reddit
What you have to be careful with here is can the rest of the business keep up. With modern AI that 5x is very definitely achievable if you get the right AI and pay for enough tokens. You probably don’t even need to be all that clever about how you use it - you can just do the same process as you would coding by hand but when you decide what you want to write tell the AI to write it. It will do it faster than you can, get most things right and you’ll have 10x more tests than you would if doing it by hand.
What you can’t do as easily, though, is scale things like support, UX interviews and sessions for planning how it works, product roadmap planning to decide which direction to go in next, etc. and even things like code review by humans ends up being somewhat cursory to keep up. You spot when it is not being DRY enough, when it is making logic unsustainably complex, not using SOLID, etc but subtle bugs you need to rely on it.
support is the one I am personally worried about. In my product my team can add clouds at the rate of one per week, but if we get 20 clouds supported then they have unusual behaviour (the current one is StackIT timing out sometimes allocating an IP) then that still takes as long as it always did. But now there can be a lot more things like that all at once.
Fit-Notice-1248@reddit
My manager thinks if you just full send anything into AI, whatever it outputs is going to work. Apparently, you can just give it some business requirements (written by PO's who will change their mind next week) and if the LLM gives you back python and your codebase in Java, she doesn't know the difference or care. Also, if you use AI, you'll never have bugs or defects, all the code it outputs is perfect.
Things like this I don't even try to battle anymore. It's so unbelievably ridiculous that some people like my manager are just beyond reasoning. Whether it's like that for you idk. But you really do have to assess if you want to spend time being reasonable to people that are being unreasonable and are just too far gone.
tr14l@reddit
It's possible, but not unless you're willing to either invest in an expert that knows how to set it up or spend a lot of time and money on innovation. If you don't, it blows up in your face and you crank out garbage. You need lots of auto loaded tooling and context to keep AI on the rails, and you need process in place to give it what it needs to be successful. If you don't have all of that, it will fly off in no time
EdelinePenrose@reddit
your job is not convince them. your job is to record the decisions, experiments, expectations, and results.
ask for how they came up with the 5x improvement. try to get them to commit to an explanation. then do your best and record what happens.
the unfortunate truth is that you’re not dealing with reasonable expectations, so be prepared for unreasonable reactions.
arnorhs@reddit
To play devil's advocate, the "nice" thing about agentic coding is you can justify that the volume of code is not reviewable by humans, so you automate the review process, and when things are broken, you are not on the hook, you just always blame the model
jnwatson@reddit
"Well over half the of the time you spend working on a project (on the order of 70 percent) is spent thinking, and no tool, no matter how advanced, can think for you. Consequently, even if a tool did everything except the thinking for you -- if it wrote 100 percent of the code, wrote 100 percent of the documentation, did 100 percent of the testing, burned the CD-ROMs, put them in boxes, and mailed them to your customers -- the best you could hope for would be a 30 percent improvement in productivity."
-- Fred Brooks
lesinsectessontamis@reddit
Sorry fully agree but it would be 70 effort to do 100 of the work so a 43% increase 😅
Jolly-joe@reddit
To be frank, this is how it's going to be from now on.
Sitting down to write code for hours at a time is over. You'll write a prompt, hand it off to Claude, and if you wrote the prompt well enough, it will spit out something better than 90% of what human devs would write in a few minutes. This is basically how senior and staff engineers were using junior and mids before all these AI tools just it takes minutes instead of days.
I'm saying this as a staff engineer. Using AI has let me do what would normally be a quarter's worth of work in a week. There are people I work with who are getting left behind because they aren't buying into this future and it shows. My CTO reached out to give a kudos because of how productive I've been then launched into a "how can you get your teammates on your level" type talk.
It's a sink or swim moment.
philip_laureano@reddit
Ask them if they've mapped out all the blockers.
For example, you might have AI that creates code 5x faster, but that's only one step in the process.
You can't get a 5x multiplier if there's a 0.5x blocker that prevents the overall process from moving faster.
TooLateQ_Q@reddit
They saw the claude 5x subscription and thought it meant 5x productivity.
driftking428@reddit
Show them one of dozens of articles demonstrating that it doesn't actually do that.
CardboardJ@reddit
Typing code by hand went from 3 hours per day to like 30 minutes, so they're not wrong. They're not right either though.
UnderstandingDry1256@reddit
Realistically 2x to 3x is achievable.
Building is crazy fast, so the bottleneck is understanding customers and solving the right problems. Feedback loop and stuff.
If you’re talking about coders team who are said what they are supposed to do then yes - 5x speedup perfectly makes sense. Promote best practices, get rid of those who cannot keep up.
abandonplanetearth@reddit
There is no shortage of developers claiming 5x or more speeds with AI.
If you cannot do it, your manager will easily find someone else that claims they can.
That's the situation.
IrwinElGrande@reddit
Do screenshare sessions, show them that it's actually 20x /s
phoenix823@reddit
Don't. Let them figure it out on their own. People learn from their own mistakes, and smart people learn from other people's mistakes. These sound like the people who need to learn from their own mistakes.
Murky_Indication1885@reddit
Ship code you don’t read and then document before you ship.
Krom2040@reddit
Sure, just as soon as a magical oracle can descend from the heavens and give perfect, well-documented requirements along with bulletproof technical plans.
ExpletiveDeIeted@reddit
I’ve been pushing it pretty hard with Ai and I’d say I get no more than 1.5 to 2x out of it while still maintaining the same rigor of quality controls and standard practices.
If I went lure vibecode I’d prob do 5-10 but have zero clue what was built or how.
meltedmantis@reddit
Depends how well its spec d. My team is mostly coding pure ai now and frankly its pretty damn good. Thats a pretty tight timeline though. Success or failure is going depend largely on how well the build has been documented up front. If you don't have that... Well your fucked
Harlemdartagnan@reddit
who is the guy that talks to management???
do you have a team lead, a technical manager... something like that?
chaitanyathengdi@reddit (OP)
No, just me and another dev. Rest of the managers are part of the same meeting. I don't think I have enough rep with them yet.
Harlemdartagnan@reddit
so wait.... they just give you guys requirements???
There should be discussions about architechture infrastructure and security at the very least. who is the person that they talk to about that?
wxtrails@reddit
We just officially got the 10x expectation setting email 😂
gk_instakilogram@reddit
Just embrace the slop friend, or the slop will embrace you
Additional_Rub_7355@reddit
What AI makes faster is the easy stuf, but unfortunately I don't think you can explain them that.
chaitanyathengdi@reddit (OP)
They would get scared of being replaced by AI if I told them that
AcksYouaSyn@reddit
Truely, let the AI Agents do what they do. Upper management has communicated to you that AI slop is fine with them, so serve them the slop.
wayoverpaid@reddit
First determine if they got the number from somewhere. If it's merely aspirational, point out that setting a hard target means that it's no longer a measure. Goodhart's law is the kind of thing you can quote to a manager to bank up your claim.
You can further explain that the innate speed multiplier of using AI is unknown, and it could be faster or slower than 5x. If 5x was set too low, it's actually going to make it easy to sandbag. (This is unlikely, but it helps you avoid looking like you are dismissing the idea out of hand.)
On the other hand if 5x is too high, developers will compromise on something else. What if they compromise on reviewing the code?
"If we set 5x as a goal, and the prompt engineering doesn't deliver that without compromising on code quality, what happens? Do we let the speed slip or the quality slip? Is it better to have 3x at the same quality, or 5x with even more tech debt?"
Hopefully you can get management to switch from mandate to experiment mode, and declare "Try to use prompt engineering to go faster, and report on what does/doesn't go faster."
You still have an uphill battle here because some idiot on LinkedIn doubtless posted how they got so much done with AI, and it made them 5x faster, and that was generalized into a universal opinion. FWIW I've seen some serious speedups in some problem domains, and in others it's been net negative. How it will help you, only you can find out.
hitanthrope@reddit
Go to your companies org chart and find out who the boss of these managers are and send them an email, saying that these people had the opportunity to come up with a speed increase target completely untethered from any kind of reality, knowledge or experience and without consulting anybody and only came up with 5x. Explain that he has hired ridiculously unambitious people and that he replace all of them with you alone, as you are in a position to make far more aggressive empty promises.
olddev-jobhunt@reddit
Eh, there's never any reasoning with unreasonable people. If they want to blame you, they're going to.
But if there are some reasonable people in there, you might be able to make solid arguments. One is to be clear about what the tasks are. Start reporting the different tasks involved, and make it clear where LLMs are a good fit.
You may also have some luck pushing them to define a metric for it. You might be able to find a metric that lets you say you got a 5x improvement on that thing.
Alternately, give them what they want. Just have the LLM "research" and go with it. Obviously the quality will be abysmal but... fuck 'em, let 'em have it if they're so into it. If they're reasonable, they'll laugh at the terrible output. If they're jackasses, they'll say you're AIing wrong. But, I mean, if they want to be unreasonable enough then there's not much else you can do.
jbokwxguy@reddit
This is where you increase your story points by some equivalent factor to more closely reflect actual work. (FWIW every study I've seen says it's a 10-20% efficiency gain at most and those are near term gains)
And if management isn't consulting devs on timelines that's a bigger problem altogether.
stedmangraham@reddit
I dunno what you should do. Even if you can do some things faster, you’ll never get 2 months of work done in 12 days. That’s simply not how work works.
I would treat it like if your manager asked you to do any completely impossible thing. If you think they can be talked down, talk them down. If you think they need to learn it’s impossible, show them it’s impossible, and document EVERYTHING. If you think they will just blame you, then start looking for other work.
trakdtor@reddit
Take down prod and hold it hostage
samsounder@reddit
Speedup is definitely achievable, but AI isn't a panacea. It'll make some things easier, but it takes time to learn how to use it and how to integrate it into workflow.
You may be able to 5x speed, but that depends on the talent of the coder and their ability at directing the AI. Its quite unlikely.
I'd try and figure out metrics. How can you measure the baseline? How can you determine if you actually get the 5x increase? Irrational expectations fall apart under rational questions.
RelevantJackWhite@reddit
treat it as if they told you any other unrealistic timeline. push back and if they ignore you, so be it, if nobody can get there they'll figure it out soon enough.