undefined | Better HN

0 pointscachius6d ago0 comments

You can go local now with qwen 3.5 9B Q4 powering hermes agent at 35 to 50 tok/s with 99 percent tool call success rate on a used RTX 3060 for the price of two months of ChatGPT Pro and never bother. https://xcancel.com/sudoingX/status/2033020823846674546#m

This is the worst local AI will ever be. It only gets better from here. https://xcancel.com/sudoingX/status/2033959603944493192#m

0 comments

croes6d ago

Nope, if nobody trains the models on new data you have at some point an outdated model.

Imagine Qwen 3.5 created in the 1990s and then use it for today web or desktop development.

And is the problem solved that training AI with AI code makes the AI worse? If not the "it only gets better" claim is questionable.

autoexec6d ago

> Nope, if nobody trains the models on new data you have at some point an outdated model.

As people train the models on new data they'll be increasingly training on AI output including hallucinations and slop. More garbage in means even more garbage out and the cycle will continue as "updated" models decline in quality.

j / k navigate · click thread line to collapse