@sama Creative writing!
It has nothing on proper writers when it comes to unique fictions or interesting storytelling, I'd really like to see that change!
@thsottiaux Longer context capacity. The ability to pick subagents to be a different model type or reasoning length. Pro models in codex. Not exactly a codex thing, but finally getting that creative writing model Sam posted about ages ago! Great stuff by the way, codex has been amazing!
@BlinkDL_AI Amazing! Is the light blue the validation set? If so, how large?
Very much looking forward to seeing how it scales and hoping for the open source release of the training code! :)
@sam_paech I think we may have gotten to the point where the LLM judge is the limiting factor. I significantly prefer 5.2-thinking's creative writing (short and long-form) over opus-4.5-thinking's outputs.
@BLUECOW009 As a PhD student with much of my work in the machine learning field, I'm really looking forward to hearing more details about your work! Hope to hear more soon!
@iruletheworldmo@VictorTaelin Yep, pretty sure you're right here. Arc AGI was never meant for this purpose. It would however be interesting to see how well this architecture works when trained for a more suitable benchmark though. Given that it still has some interesting ideas going for it.
@TheyCallMeNov@bbcdoctorwho He never left, he's still playing the part of the 14th doctor with Donna, just not on the show. (Watch the 3 60th anniversary specials for context). Don't think we need to bring him back to "save the show", it's been great. Especially season 2.
@blovereviews @skipsterdev @sama Nope. Definitely the fault of the company. We don't have the same regulations on AI as most of Europe because we're not part of the union anymore. They did it continent wide rather than union wide.
@sama I also have a strong inclination that it is massively under thinking many of the problems I send its way. It would be much better to overthink slightly too often than to underthink too often.
@sama O1 really seems to be worse at keeping track of the information provided in past messages than o1-preview was. It keeps forgetting things asked of it in the past, as if they should no longer apply. However no indication is provided that the past instructions should now be ignored
@TheLone08157926 @jceeclay @RadioTimes They're not over 7 days, if you read the article they're over 2 days. The Saturday and the Sunday. Nothing from Monday onwards is counted.
@thefootdaddy @WilliamWhoDW When I visited that website it said everything from the release until Sunday and did not include anything from Monday onwards, so definitely not 7 day consolidated. Also, it doesn't include those who chose to watch it on Disney Plus instead.
@Sith_Elvis@HarrisonChase6@RadioTimes I have a feeling not many people are simply just not watching it live specifically. For one thing, I've been watching it on Disney.