That's pretty stupid. Most people who are incurring significant costs are just tokenmaxxing rather than being efficient with usage. You can get 99% of jobs and work done with Haiku/Luna in a collaberating working enviroment.
I feel like people who are later to the AI game just like to "oneshot" and sink a bunch of usage into generating garbage
There really is a skill to using it effectively. I've tried coaching some of the devs on my team. Some get it, some don't.
Our company has been tracking token usage and models used vs output (tickets, story points, PRs, deploys, etc...). A dev got chewed out, even after I warned him, because he spent over $2k in a single month almost exclusively on Opus while his actual productivity in terms of what he delivered was abysmal.
I get it but it goes against the grain for me. Isn't it ironic that we have to waste our precious and expensive human brain cycles to think about how to use AI cheaply so that it is not more expensive than us?
In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
> Isn't it ironic that we have to waste our precious and expensive human brain cycles to think about how to use AI cheaply so that it is not more expensive than us?
Nothing ironic about it, it's basic engineering efficiency optimization.
Just today I was calculating that if money was not a factor at all, we could simply use Mythos to handle all of our continuous security scanning needs, at a cost of about $15 million a year.
Well, my budget is far (far, far, far, far) below 15M a year, so that's not going to work. So I must find compromises to make it work within budget. The single most important engineering constraint is always budget. Everything would be easier with infinite money, but there is never infinite money.
It is ironic because choosing the right model is what the models themselves are good at. If there were no business incentives it'd be easy to have just a single frontend that routes to a suitable
model. Instead I have to develop the decision skills which is a complete waste of my time which I could better spend in the problem domain for tasks I'm better at than the AI.
> In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
It sounds like the parent is less talking about this, and more people burning tokens while not getting useful work done.
Attempting a serious but not-a-certified-whatever answer: "Points" do have meaning when properly used as a kind of moving-average tool for forecasting within a particular context.
Problems arise when people try to perma-peg them to particular tasks, or (worse) man-hours or (much worse) man-hours across teams. Even just encouraging the humans to answer in terms of hours/days taints the accuracy of the forecast by introducing a kind of bias.
My wife, despite loving all things French, just doesn't "do metric". She wants my height in feet and inches, my weight in pounds, boom done. So while I know my mass in kilograms, she needs the conversion done before she can even begin to have a reference point.
Upper management is the same way. They have forecasts that they need to make, deadlines and budget goals that they need to hit. They only deal in the units of hours and dollars (or local currency). Every software engineer is accountable for their work in those units only. The conversion needs to be done before the management chain has a reference point.
One easy way to do this is to have each engineer estimate the time it takes to fulfill a story after it's been pointed; then, upon completion, record their actual hours spent. Their estimated vs. actuals tend to stabilize over time, so even if they misestimate a task, you can arrive at a good guess at the time it will actually take.
> > Problems arise when people try to perma-peg [points to man-hours]
> Upper management [...] only deal in the units of hours
I feel you're mixing up different operations here. You can always express unfinished work as likely to require a certain number of team-sprints, which are convertible to theoretic man-hours. The key is that the conversation rate is only valid for a moment, and technically that moment was the prior sprint.
That's very different from management thinking (or worse, declaring) that points have a permanently fixed proportion to man-hours.
> [...] and dollars
If your management deals in international currency, then perhaps that would be a useful analogy to them: Points and Man-Hours are different sides of FOREX, and they fluctuate based on different conditions.
When Engineering predicts a group of tasks is 54 points, that's like a foreign company signing a long-term contract in €100 EUR instead of USD. You can estimate that you'll receive ~$112 USD in a year, but the actual dollars will likely be different because the exchange-rate will continue changing before that happens.
So, what you do is you recognize every ticket has a somewhat variable “actual effort”; and, if you’ve been honest in approximate effort pointing, you’ll know your team (or your own) velocity.
From there you can run Monte Carlo simulations - say a few hundred thousand, and get a pretty good estimate of actual time spent.
Assume that for various practical reasons, we've decided it's one of the best functions out there. How do we use it effectively, especially when it has noise, and drifts over time with unseen changes to the human_estimator and the hideously complex world_state?
A popular option is to run it multiple times with different person/task combinations, putting a projected number on to each task. Afterwards, the tasks finished in sampling period ("sprint") become a quantifiable total for that period ("velocity").
Do the same process again with the next set of tasks, and you can figure out which ones are likely to fit if the velocity doesn't change much. If you know the velocity will change due to losing staff or vacation days... well, we apply a multiplier and hope for the best.
Trying to "fix" the meaning of points is maladaptive, because they reflect many changing things which are outside our control and can't be independently measured.
Honesty isn't the issue, in my experience. The issue is unknowability. You usually can't know what clients actually want until you've given them a prototype and gotten their feedback. You often can't know the rough performance impact until you've done the first implementation. Sometimes you just have to work through a list of candidate solutions S1, S2, etc... and you can't know which one will eventually work. All of which means that forecasting beyond a couple of weeks is wasted effort, except for rote factory-line work.
I thought you weren't supposed to use it to estimate man-hours? Or is it ok if we don't allow engineers to provide time estimates, but permit simulations to do so?
Claude, vibe code me an entire startup, the actual product doesn't matter, but it should all be based on the incredible pun "turn 'sorry' points into story points."
/goal get accepted into Y Combinator, you have an unlimited token budget, be bold.
EDIT: no, do not just make a product that gives away your unlimited token budget to users for free!
there's a manifold to what "effective" means. The problem is once you get into the vibe flow, it's really difficult to eject yourself into the other realms of vscode or IDE or whatever it is you normal do because the vibing provides no anchor to what you're doing.
Even if these models are smart enough to reorient themselves, they get entirely stuck in a desert and now you're asking someone to just pull up stakes and digg them out even thought they only watched them get there and the UI provides so much speed that no human can comprehend how they got there in the first place.
It's like asking a pilot to take over in an emergency situation when they're not tasked with any of the every day requirements of the job. The orgs are relying on borrowed time of experienced professionals, and that's going to erode away and what replaces it is mostly people who understand how to navigate context but not use any of the _classic_ tools.
It's a real conundrum and won't be easily surfaced but for a decade.
I’m trying really hard to keep my skills up but it doesn’t feel productive when I’m using it to write code. It feels like I’m slowing down the AI to the point that it’s not as effective as just letting it go. But I don’t get all the learning that comes from that time along the way.
Have you found ways to stay sharp while using it? Or are you relying on other projects outside of work to keep your skills fresh?
I'm primarily using local models. I turn on thinking visible and try to review at speed what it's doing.
Other than that, no. I drag myself out to poke around occasionally because at times the local models get to far into context and refuse to do simple tasks.
> There really is a skill to using it effectively. I've tried coaching some of the devs on my team. Some get it, some don't.
Very much this. There is a vast range of effectiveness and combined with so many models and pricing tiers, it can be tricky for some. You really need to treat it like partly a programming language, but also partly as a management delegation exercise (do I delegate this task to the intern (cheapest model) or to the principal engineer (frontier model) based on complexity).
I do coworking sessions with most in my team to see how they are using it to understand and coach for effectiveness. Just using the most expensive model for everything isn't going to cut it anymore in the post-tokenmaxxing age.
It's funny, the thing that makes effective prompt also makes effective documentation/communication.
It's bizzare to see people that made clown issues (not enough detail etc.) suddenly start writing detailed prompts just because it is AI that will do the task and not the human on the other side.
The guys that made clown issues are still using AI to make clown issues, they just look like plausible specs now that claude wrote it more thoroughly. At least pre claude I could tell that they had missed something earlier on, ask a question to clarify, and get a real answer they had to type themselves. Now most of my stories are rehashed after I raise PRs or even after these same guys approve the PR and then realize they forgot a requirement.
You mean that now rather than taking 30 seconds to write a ticket, because we all know what we're talking about, we must take 10 minutes to give all the background information to the AI?
If someone burns through two grand in api tokens over a month and still can't move tickets across the board, the bottleneck definitely isn't a lack of compute.
I'm not exactly one-shotting software, but because of the pace I'm expected to keep, I place a lot more trust in the bot that I'm actually comfortable with. I need the job for the time being, so I just keep hitting the button and letting it do its thing until the tests are green.
I'm just working in DevOps though, so it's writing IaC, not application code (save the odd Lambda function or python script). Still, even when I spend an entire day conversing with Claude and watching "bot go brrrrr", I'm one of the lowest users in our company. I have no idea what the devs who regularly hit their limits are doing.
No large companies I know allows chinese models, and OpenRouter is abysmal for enterprise usage. Don't confuse personal/yolo startup usage with corporate use.
I feel like it's only within the past few months that opus got to the point where guiding the model is faster than doing things myself. I tried out sonnet recently and it was not a net positive to my work. I feel like anything that I'd trust haiku to handle isn't worth doing in the first place.
For context, I'm doing a range of tasks, everything from one-shotting adhoc scripts to having 4 hour 10M+ token conversations debugging things.
I’m a different person and definitely not a genius but my experience today goes even beyond theirs.
I had Opus trying to simplify a query for me which was slow - it ran for maybe 30 minutes, including writing and running tests, and came up with a refactor across 9 files with a couple hundred lines changed. I was looking through the output before moving onto the next step, and noticed something a little fishy- I said “why does it do x, isn’t that a more complex y?”
Opus thought for another 20-30 seconds then output “Actually that would make the majority of the diff irrelevant, if we do that change it is just these 4 lines in this single file instead.
So then I had it do that. 5-10 minutes of writing and testing and that was done.
So my company spent $25 in tokens and I spent probably an hour in total for a 4 line change that, in the days before Claude, I probably could have found the correct file and thought through the problem, understood the solution, and written the 4 lines of code myself. Probably in the same amount of time.
So basically there was no benefit at all for my time, an extra cost to the company of $25, and now I understand our codebase a little bit less instead of more if I had done all the work.
As good as Claude is at building greenfield projects it still struggles a lot at complex ones
This is a recurring issue for us. Very small team. We move fast and pretty loose.
Dev + AI spend 3-4 hours on a project plan, there's a "wait a minute" moment, and finally they spend another hour dialing it back to a solution that could have been built, tested, and deployed in 2 hours.
Example: Someone was setting up a dev environment with multiple DB migrations from different branches - AI planned this wild 8 phase solution with a pretty fancy cutover event.
In review I essentially said... "Wait, isn't this a dev environment? It doesn't need 0 downtime, why not just destroy and recreate the DB" and it turned into a <1000LOC script.
Technically the original plan would have worked, it would have been more robust, but it would have taken a good deal more time to implement.
Some of this falls on the devs to know what fits our team well, what's realistic, what's obviously overengineered, etc... But some of it feels like AI just defaults to the most complex version of a thing. I catch it SUPER frequently. (And unfortunately some devs think that more complexity means it's a better solution)
I feel like I’m going crazy, using all of the best models, spending time to have excellent prompts, configuring tools and skills… and still getting overly complex solutions with mediocre results.
Like it’s still impressive how far we’ve come, and undoubtably cool technology. It’s made a bunch of personal projects possible that I never would’ve started.
But for a business I’m struggling to see the ROI. Sometimes there’s a big benefit and sometimes it’s net negative. Not saying we won’t get there but I’m trying to stay grounded in the reality of today rather than the hopes of where the technology could get to
It's brilliant tooling, and made the mundane/tedious parts of my work a lot easier.
But, I fear that the "facade of complexity" makes the output SEEM better. Someone might read a 10 page Codex generated plan with fancy diagrams / charts and have a feeling that because it is so complex, it must be good!
Same deal with text output in general. Oh, it uses a lot of big words and there's a LOT here, it must have done a lot of work to get to that point.
I find, in reality, that it takes much more effort to get to the simplest solution.
As they say, any old fella can build a bridge that stands, but it takes an engineer to build a bridge that barely stands...
It's not so much that I can code fast, it's that it takes a significant amount of time to tell the model what exactly I want the script to accomplish, and at that point I might as well write the script myself. And often, figuring out what I want the script to do is that hardest part, so it doesn't really matter whether writing the script takes 10% or 20% of the total time.
> I feel like anything that I'd trust haiku to handle isn't worth doing in the first place.
The trick is making the workload manageable by the cheapest models, or costs will destroy you. We seek to make all repetitive tasks be effectively done by cursor composer model which is the cheapest.
Doing recurring tasks with anything more expensive than that will burn the budget in no time.
Or does a "collaborating work environment" mean that everything is basically spoonfed to them? Or do you only ever use ghost suggestions?
I genuinely cannot even fathom. Just how do you even get into a state where tasks are so clear and cookie cutter? These things are abhorrent. Not only are they not useful, it's an outright form of psychological torture to try and use them. They almost fight you.
Luna doesn't even respond to steers properly! You try steering it and it immediately gets distracted and then just stops.
I can imagine coercing Sonnet into doing some of my tasks okay, but Haiku? Especially 4.5? Really?
I think you might be overestimating the sort of projects most of us have worked on throughout our careers -- we haven't been doing much groundbreaking work. LLMs can easily and successfully write most code.
It's possible it's my role distorting my perception, cause technically I don't write software, I work an SRE role. None of my items come pre-chewed or paced, it's all good luck and god bless.
I'm desperately trying to classify and standardize my work items and delegate them to less capable models, because my usage is clearly unsustainable and this same sentiment as above keeps being pushed on me too. But all my tasks are genuinely fairly arbitrary, so there's no real way around the agent actually being able to reason about business and technical context proper. It's not even that they're hard, it's just that they're dynamic.
I can get Luna to do things like walk our observability stack and perform a healthcheck, then defer to a stronger model if anything looks super off, but if I'm being entirely honest, this could basically be just a script. Which Opus 5.5 will immediately write for itself if it doesn't yet exist, run that, and then off it goes depending. But Luna will never actually do an investigation proper. Heck, it can't even read our dashboards most of the time, tripping up on Grafana minutia.
It feels like that surgeon vs surgeon comparison, where you're made to decide based on their surgery success rate, and the better succeeding surgeon simply reward hacks the number by only operating on less dicey cases. Except there's no objective way to make this classification here, so jackasses like the above get to play with my insecurities with full obnoxious confidence, while I'm left desperately trying to slim my usage and failing to do so between two moments of crippling self doubt and blockers.
For the last few weeks I've been using Luna with Medium Reasoning for routine debugging, installing/updating dependencies, test creation/fixes, CLI/config miscellaneous problem solving, etc and it's been solid. I can let it churn away for a half hour and it barely moves the needle on remaining usage on my Plus plan. Previously I had been using Sol Low/Medium and would frequently hit the 5 hour limit doing those kind of tedious automation and routine problem solving tasks.
I wish we had such a platform (and was properly adopted). Maybe then the necessary context would be properly organized, and these lesser models could be effective here as well, especially if combined with harnessing integrations too. I still have a hard time accepting that Haiku/Luna tier models can be effective there even then, but I'll just have to take your word for it I suppose.
in my company there are a few who keep sharing screenshots of reaching limits on 3 separate subscriptions, 2 of them their personal on top of the company subscription
Wonder when subscription-hopping attacks will become more often (jumping from a personal model to injecting instructions into the business account and exfiling data)
I think they mean the new Haiku, which is mildly above Luna now . If you have a plan written by a smarter model (so the hard parts are solved) they can be great at implementation.
It's impressive for its total size if you need local inference.
Though it's significantly slower in Token/s and also thinks a lot more without matching the same intelligence (xhigh qwen27b scores lower than haiku's medium setting, and haiku-med is $0.05 per task compared to Qwen27B's $1.01 on AA's comparison)
It seems like more a backup if you need to work offline, imo, unless time doesn't matter and/or your electricity is free. Or you want independence from the labs (fair enough).
I feel like people who are later to the AI game just like to "oneshot" and sink a bunch of usage into generating garbage