So, the new models—o4, 4.1, and the whole family—are here.
On paper, it’s a revolution: expanded context, enhanced capabilities… Impressive. Truly. Except…
In practice, I sometimes feel like I’ve taken a leap… backward. Getting answers that strongly remind me of the “creative interpretations” of our good old predecessors like GPT-3 or DaVinci.
We were promised a Formula 1 car, and we ended up with a 2CV doing poetic zigzags.
And that’s when you think back, teary-eyed, to good old GPT-3.5. At least it didn’t think it was Rimbaud when you asked it for code. It did the job.
Predictably, sure, almost boringly so, because it worked so well for basic tasks. Ah, the sweet monotony of efficiency!
Hence this iconoclastic idea:
what if OpenAI gave us an “Efficient Mode (Legacy)” switch? Just so we could choose between the “intellectual adventure full of surprises” of the new models and the “tool that just does what it’s asked” from the previous generation.
You know, just so we can finish our work before retirement.
Any other nostalgics out there for AI that (simply) worked?