I currently only have two small subscriptions, so I haven't yet tried to get my employer to pay for it. Still, it won't be easy, haven't heard of colleagues who get to use their research funds and grants to pay for AI tokens. But eventually, that's the way we're heading. Just policy hasn't caught up yet I feel.
So, how about you?
If it's out of pocket, how do you justify paying for it for work use?
If it's your employer who pays, how did you bring it up?
If you're the employer, what's your policy ?
dayjob provides a subscription to GitHub copilot & anthropic.
I have a personal subscription to Cursor that I use for some side projects.
I have been tinkering with kimi & glm52 using Venice as a provider
also, I'm running a business that routes tokens to providers, and so we have something like unlimited access to open weight models that we route
ppq.ai when I need it.
I should check that one again, I've been using the same two platforms, more out of convenience and inertia, rather than aiming for the best model available for a specific task.
I pay for it myself because it is useful and fast. I just see it as something that saves enough time that it makes my life better.
Is that mostly writing, in your case, or do you use them for other stuff? Do you see clear benefit of using frontier models for writing (and associated research) purposes or do the older models work sufficiently well for your tasks?
I almost entirely use ai for research. It does a pretty good job of crawling niche topics and helping me find loose ends on which I can pull.
Also its handy for quickly getting up to speed on obscure historical events.
Occasionally, I will use it to write some little program to automate some kind of tedious research task or for data analysis.
When I use AI for privacy research, it repeats common myths. Expected, but disappointing.
I find that researching concepts where you rely on AI for explication of the concept is not very helpful, but if you used it to find privacy projects, especially if you took a very hands on approach to guiding it to various repos or developers and then asking for similar or forked projects, it might do a good job.
Which AI do you pay for?
Claude and OpenAI
I've seen a lot of people mentioning Claude.
He's pretty good, isn't he?
I have access to a version of ChatGPT Pro through my university, but I find that it's my least used subscription because it doesn't enable API access. Thus, its usage is limited to chat sessions rather than programmatic work
I pay for a Cursor Pro and Claude Pro subscription as well. I use it mostly for work and I justify the personal expense as an investment in the quality and speed of my work output. I could probably ask my department to pay for it, but they might object and say if we offer this to you then everyone is going to want one too
I put a few thousand sats on PPQ just to test out some different models from time to time.
I think I'm almost ready to start paying for some small subscriptions like yourself as I have a number of ideas that I'd like to have a good go on.
I think the models will get better and competition will continue to bring the price down.
#1536067
I do wonder though what will happen when the infinite money supply dries up and models have to actually start becoming profitable to survive.
My time is way more valuable than AI time.
Most costs for me are sunk into poor quality results, and then my main cost is my own time spent more than token spend. Also me getting it wrong is annoying, like I did yesterday when bot review questioned my decision and I was clearly wrong. Setback of 5 hours of think/design/work time, truly wasteful, and shame on me. I've declared those hours non-billable and thus lost more than all the bot monthly subscription costs combined at ultra levels would set me back for.
That said, it can also happen the other way around and then the bots waste your time, which is why I only have 1 minimum subscription left (would have another one if they weren't forcing me to buy a whole year before evaluation.) The most part I PAYG on ppq and it helps thinking things through before I send it off for processing (but apparently not enough.)
I got multiple offers from clients to pay for my tokens but I honestly don't want to be locked in like that, and I like having my bespoke tooling that is fit-for-purpose for me personally.
Billable hours -- that probably makes a big difference between you and me.
As a researcher on a fixed wage without productivity incentives, other than my personal drive and/or passion, my employer is not really pushing me to be more productive. If/when I eventually become a civil servant, the carrot that is the permanent academic post will disappear as well.
Still, I want to be the best at what I am doing now, advance the field, publish impactful work, etc, so incorporating AI in my workflow is non-negotiable. So, that's what rationalizes me to pay out-of-pocket, but I do hope that at some point I can use some of my grants to pay for it. Especially when I get back to heavy development work, it was painful when I was reaching my limits more often than not.
It's a different situation, but if you're talking about fixed cost subscriptions, this is often easier to manage when you have fixed income than when you don't; less risk. So I'd stick with whatever is affordable.
For my example above, I was getting a bit restless. So I said "okay Fable... propose a fix." (in expansive, precise instructions.) It burned through my 5h allocation + $19 in extra credits. I get the result, read it, it feels wrong. Fire up GLM and GPT concurrently to do an assessment. Both find the same error I thought I found, and extras. No clean solution. $30 down the drain plus a 4.5h pause on my subscription, no progress and thus no billables because no results. No offset to this cost. I just have to bear it, find another solution. Plan a refactor instead. Which is even more expensive and carries heightened risk, but it has to be done. So I'm doing that now.
Still, the bots found bugs that otherwise would be on other people to fix. It is still better this way. The quality of the final result will have to make up for it and at least I'll be able to sleep well, not worried about what bugs I may have introduced. Bottom line, I'll not make a loss on the token cost, but just worked to get rather inefficient, yet hopefully really high quality, results
I'm using free A.I now, no funds to pay for one
I pay for it if they let me or ask, but we don't use a company plan because it's more effort, has weird admin overhead usually, and creating one-off credit cards that are merchant restricted is better overall. For myself, it's more than worth it to have two maxxed subs, so I assume the same for others.
So basically one day there were a few handymen and a truck load of expensive electronics at the front desks. They were directed into the basement and spend a few days there and we were told to connect our shit to
companyname.internal/api/v1and that's itIf I was still at megacorp I'd probably still have exactly the same subs trying to build myself an exit from megacorp.