@ProjectOnT28479@argofowl I have been almost exclusively using Codex after the 4.8 launch so I don't have a good reference on that but I have heard it got back closer to 4.7 usage rates
@NicolasZu You need to have an orchestration loop that loops through your loops to make sure they are properly managing the loops that your agent loops loop through
Higgsfield Mod for Minecraft is live.
> prompt any building or city, even the Statue of Liberty
> create paintings with text-to-image
> snap a view and restyle it with image-to-image
> make videos from a prompt with text-to-video.
> animate in-game photos with image-to-video
Currently I have it review the same task from a cold state so there isn't context rot. The adversarial portion I have set up for it to be over aggressive on critiquing how the task was done.
When I am running a longer session through a /goal prompt I have it loop through review sessions until both the agent and the adversarial reviewer agree that the task is complete based on a score of 9/10.
@HudsonByrd77 I have a repo mapping skill that builds an architectural map of the file system and critical files.
Then there is a stop hook that if this skill has been run in a repo it checks the diffs from the latest completed turn and updates those files if the diffs effect them.
@HudsonByrd77 Read only adversarial review agents at the completion of a task has been a huge win for accuracy. On Stop in codex I have an agent review the last turns work and hand it back over to the agent.
I think it is dynamic throttling on their end. When they need a bunch of compute to train new models or are under heavy usage loads.
Add in that AI models don't consistently give the same outputs by nature and are pretty prone to odd divergent behavior and it can become a weird compounding effect.
I think that is why you get such polar reactions on X where people simultaneously will claim a model feels smarter today while someone else says its dropped 30 IQ points.
@flowersslop It will definitely come back as hardware improves to reduce the bulkiness problem and services like Project Genie @GoogleDeepMind allow for generative experiences and worlds to be streamed to them.
Lower on device hardware requirements for rendering mixed with lighter hardware.