Browse Comments — Raw (as collected)
Close reading of the corpus at each pipeline stage: raw → clean → relevant → coded.
118
comments matched
· page 5 of 6
Because we still in exploration stage, and people still learning on how to use it. Eventually token consumption will drop and cost stabilize
This is what happens when companies don’t truly understand what they are getting.
Ai is good , if utilised in the right way, but if your just installing AI , just because it’s AI …. Then your doomed to fail!
Microsoft didn't replace Claude with humans. They replaced it with GitHub Copilot CLI, which is also AI. Uber's adoption jumped from 32% to 84% in months, and 70% of their code now originates from AI. The MIT study you cited only looks at computer vision tasks like property appraisers, not coding. So which company in your post actually shifted work back to people?
Skynet and AI arrive, and you wonder what could go wrong?
We need to find a better way of AI usage. Here is some interested information I am researching on https://www.linkedin.com/pulse/two-ways-ai-agents-work-why-difference-matters-more-than-matsaylo-ttz0c/
...But at least it don't got any emotions. It don't need any benefits package. It don't need any additional desk or office space. The only time it takes a vacation is when the system it's running on crashes, or during some kind of scheduled maintenance or upgrade -- otherwise it's working 24/7/365.25 without even half of a complaint. And, best of all, you can yell and swear at it all day, and it won't even mind.
Seems we made a wrong turn somewhere.
Really!? Shocking!
I’m an engineer. I have never worked for $500 to $2000 per month.
Spot on, Saurabh. You’ve diagnosed the 'Token Trap' suffocating enterprise AI rollout.
This is the modern equivalent of writing a SQL cursor in a loop—instead of just slowing a database, it burns corporate treasuries. When an unsupervised agent runs unoptimized prompts over a massive codebase, token bloat compounds exponentially.
The failure isn't the technology, or the brilliant engineers prematurely laid off for 'AI efficiency.' This is a C-suite miscalculation. Replacing highly skilled human intuition with bloated, pay-per-token cloud autocomplete is a massive misallocation of capital.
As founders building physical AI systems, this data is a mandate to reject cloud API dependency. The future isn’t paying per-token for generalized cloud reasoning; it’s running purpose-built inference locally at the physical edge (like Apple Silicon) to permanently drop variable token costs to zero.
AI was never cheap, unsupervised labor—it is highly leveraged infrastructure. Treat it like an intern and the API bill will bankrupt you. Architect it as a localized power tool for human domain experts, and it changes the world.
$500 to $2,000 per engineer per month is not a lot of money, and at least in my case -- totally worth it.
I would also like to know which models of Claude people used at Microsoft. I predict people used Claude Opus too much instead of Claude Sonnet.
Prashant K. Sahni The only way AI does anything 5-10 times faster than a human is if it has already been done so many times, you shouldn't be doing it in the first place, you should just be using an existing API.
Yes, organizations purchased AI licenses across teams without fully evaluating where they truly fit or add value. As a result, many employees have access to AI tools but barely use them. Some use them only like a Google search engine, while others use them without proper context or prompting. Because of this, AI budgets are getting exhausted very quickly.
Now organizations are beginning to realize this challenge and are asking teams to follow specific guidelines for using AI more efficiently and with lower token consumption.
The honest version of the Microsoft / Uber story isn't "AI is too expensive." It's "we deployed it without unit economics in place." Those are different problems. The first says stop. The second says instrument, budget, gate by ROI. Most orgs that "discovered the economics were never stress-tested" don't stress-test any tool until the bill arrives - this reads as a procurement maturity story dressed up as an AI story.
This is such an important reality check especially when studies are already showing AI is economically viable in only a fraction of roles despite the massive hype around replacing humans. I think a lot of companies underestimated that scaling AI also means scaling infrastructure, oversight, context management and decision accountability.
This is only true if you let technical leaders that don’t understand AI implement your systems 😂🤣
It should cost as much as the energy it needs. If it’s more than a human being then it cost a lot more.
Why PAY for it? Things can be free , don't be a fool. Wake up and 🧠 use it. There are so many AI out there are totally free without paying API or usage or subscription, be smart don't be a sheep 🐑 be a human and use your intelligence.
The unit economics angle is spot-on. We've seen the same pattern play out internally, where teams spun up Claude for everything in Q1, then got real about where it actually moved the needle by Q2. The infrastructure costs looked fine on a spreadsheet until you ran it at scale for six months straight. Totally agree that it's a tool multiplier for good engineering, not a replacement for thinking about the problem first.