Who's dumber today — and who's not

“A week later: Astra started drooling. Suspected of being routed. Two weeks later, countless people on X started complaining”

“Better than Fable though where it seems to basically cost £15 per prompt. You hit your what-you-thought-was-generous monthly spend mid-task, causing Claude to stop dead, and that £15 to have been burnt.”

“UD-Q2_K_XL works but the quality loss is definetly there, it constantly has to correct its own mistakes and gets stuck in loops for quite some time until it moves on”

“Ds4 i can only use via OR so it's comparably slow”

“The looping is the real deal-breaker and it's not the quant, I've seen it on both Q4 and Q8 which points to it being baked into the model rather than a precision issue”

“You've already iterated through four versions”