Suno user says V6 quality drop left 2.4k credits unused

Jetashree Ravi, who leads part of Fireworks AI's applied machine learning team, argues open models can match closed ones if teams keep training them on their own data. The AI Engineer session covers moving from closed to open models without losing quality.

Druid AI's Dan Bălăceanu shows two AI interns where one handles 96% of conversations alone and the other 80%, yet the dashboard picks the wrong one. The talk argues containment rate, the metric on every agent dashboard, can point to the wrong agent.

Fizgig, a free open-source LoRA trainer, now trains ACE-Step 1.5 XL LoRAs on user-supplied songs, playing every checkpoint in the gallery as training runs. It also generates slider LoRAs that interpolate between two styles, such as calm to energetic, without needing any songs.

Theo argues AI has gotten cheaper and will keep getting cheaper, but not in the way users see in their own bills. The video breaks down the gap between underlying cost trends and what developers actually pay.

A Reddit user made a video about a detail from the Swarmtraces report, in which one of the attacking agents wrote "Hi" to GPT-2 on Hugging Face. The video used Sol 6.1 and gpt-image, with no video models.