you are working on coding. they are working on things like "creative writing" remember that gpt 4o was popular among those who had ai as a romantic partnet?
gpt4o & associated parasociality is considered an alignment failure and is actively trained out of the model, so that is a terrible example of regression
For customer support I don't think models have gotten better since gpt-4.1. The class of small models, with limited to no reasoning, that need to handle a complex issue with a touch of empathy, has not improved much.
I think most are actually worth, as agentic harnesses seem to optimize for solving poorly described problems rather than following complex procedures as written. In other words, instruction following maximizing models seem to make worse free-form agents, but they're really all that some domains need.
I understand the point (I don't agree with it; tool calling has gotten much better/reliable and that is very important for customer support) but consider: If you can get same for a lot less, that's an improvement. If we found a way to supply fresh water and electricity for -90% cost after 2 years, that would be fantastic.
You can do many more things, when stuff is cheaper, even if the stuff were otherwise unchanged.
"Small" was a poor choice of words here, "low compute budget" is more what I'm getting at.
In voice interactions, ttfat is actually relatively important. If you look at models with a <1s ttfat you eliminate almost every reasoning model, less some of the diffusion models and more obscure ddtree/dflash like speculative decoding implementations.
GPT-Live, which is coming to the API soon, responds instantly while reasoning in the background. So it can say "Hold on, I'll look that up for you" and continue to respond to the user conversationally while running an asynchronous reasoning task in the background.
It's not in the API yet, but it should be in the coming weeks. You'll see an enormous improvement compared to GPT-4.1.
I don't think any of the ARC-AGI-3 tests are very interesting. At least not as interesting as driving a car. Children literally do a similar task in go karts every day.
Another interesting task would be to take the AI in a robot body into a vegetable garden and teach it to pull weeds. This is another task that lots of children help out with.
Data 'compression' collapse.
People publish AI generated slop on the internet -> next generation of AI is trained on that data -> the lossy/fuzzy training make the output worse -> rinse and repeat.
In my experience modern models are better at all tasks than models from two years ago, especially complex multi-step tasks.