> Maybe because the models keep getting better and the tools are slowly getting good enough to replace people?
I am not sure we are there yet.
In my experience even the SoTA models are faaaaar away from replacing humans, maybe making them a bit faster.
The correlation between job cuts and ai growth is real. But it's related more to ai cost than ai performance. Especially in Oracle's highly leveraged case.
This conversation always gets lost in the weeds. You don't need a SOTA model to replace an entire engineer role from the time he gets into the office until he leaves.
If you have a team of 10 and make them all a little faster, you can do the same amount of work with 9. Run this out over the entire industry and it's hundreds of thousands of roles that are redundant.
Do you have any research, empirical data, or other hard evidence that shows that a model (especially a non-SOTA model) makes engineers even "a little faster"? I'm aware of anecdotes, but nothing more.
The cleanest positive study is Cui et al., because it used randomized rollout in real companies. Developers at Microsoft, Accenture, and a Fortune 100 electronics manufacturer were randomly given access to an AI coding assistant. Pooled across 4,867 developers, the authors estimate a 26.08% increase in completed tasks among users. They also report increases in commits and builds.
While that research was published in 2026, it analyzed the 2023 period. Very much NOT sota models!
https://pubsonline.informs.org/doi/10.1287/mnsc.2025.00535
Thank you!
> I am not sure we are there yet.
When talking about bullshit jobs like, say, taking a bill received by paper and manually extracting data from it (company name, invoice number, bank account details) to enter into an accounting program, AI is already good enough according to the pareto principle.
[dead]