So you’ve become a model trainer. Not sure this can be a long-term job.
2 Likes
This seems like it would be bad for business, unless you have 4x more work in the pipeline than you used to have.
2 Likes
Fair point, and I did think about it.
For me, the market decides this, not me. If I bill 12 hours for what now takes 3, I’m competing against people who bill 3 and they win. Less business is bad but no business is worse. It’s like billing a client for the hours you’d have spent finding the answer in a book, when you found it on a forum in ten minutes.
Then, and this is really your own point from earlier in the thread: the hours were always just a proxy. You said it better than I did: you’re charging for access to experience, not for time elapsed. So if the same value now fits into fewer hours, the rate goes up, or the work moves to fixed price and the client buys the outcome instead of the clock. What I’m arguing against is billing 12 hours for 3 hours of work, not against being paid what the work is worth.
And the hours that come back aren’t dead time. More projects, or business development which as a freelancer is the thing I never had enough time for.
5 Likes
Probably not. I’ve found most projects are limited by budget and time, not features. Usually features get cut to make the deadline. If you can spend that extra time that freed up on polish, you get a better product.
Wouldn’t that be the case if the providers were allowed to train on the interaction? As far as they tell on the paid subscriptions they deliberately don’t use your interaction for training
Good example of how productivity works in an economic sense. Of course you could still charge for 4x the time it really takes you (because that’s how the “value” of the outcome was determined previously), but someone else probably won’t. And then that someone else will likely be the one who gets 4x the work in the pipeline, while your own pipeline will stall.
Now in the real world, things won’t be that linear. But typically those who embrace ways to increase productivity will come out ahead. It’s the nature of competition, and the nature of increasing productivity. I don’t think we should moan the working hours “lost” to better tools, but instead embrace the fact that those hours are now available for creating more (or better) stuff.
1 Like
Unless you skip validation and testing, I would question that 4x speedup…
Validation and testing is a bona fide sweet spot for frontier AI.
Professionals using it will know that it excels at TDD and unit test coverage. It’s incredibly detail oriented and thorough.
Extremely useful for single devs or small dev teams that don’t have access to dedicated QA teams. Especially considering that devs are notoriously bad at testing and validating their own code/apps.
The absolute killer is to build automation capability into your app and inform the AI how to use it and measure for success/failure during test runs.
Done correctly, not only is there a multiplier on productivity/time improvements but also on thoroughness and robustness of testing.
For some external validation of this regarding testing, look up Linus Torvalds recent interviews on the usage of AI on the Linux source for finding bugs and issues.
3 Likes
Does anybody have a good reference to how much IP is practically exposed when using Claude or Codex with (personal) paid subscriptions? I read they are not using your code to train their models.
PS. Assuming:
Claude: Privacy>“Help improve our AI models” off
Codex: Account>Settings>Data controls>“Improve the model for everyone” off.
Yeah, if they say they’re not using your code to train then that’s absolutely 100% true, these companies are so trustworthy and would never lie about anything or make any misleading statements
6 Likes