Gorth

Chilling

[Gorth @__Gorth__]

@rezoundous opus was peak "personality" from Anthropic.
Fable is kinda usable. All other recents anthropic model have been garbage in personality.

[Gorth @__Gorth__]

@krishdotdev claude is literally telling to go sleep every fucking day now

[Gorth @__Gorth__]

@default_52 @HazelAppleyard the cat video was so good

[Gorth @__Gorth__]

@cstanley the longer the character length the more it happen but yeah it's not totally stable and it take a long time to process this. Completely destroying your 1/RTF for the batch ( if that was implemented)

[Gorth @__Gorth__]

@waldenpod just the dash and i will assumre it's AI

[Gorth @__Gorth__]

A few weeks ago , while doing advanced networking configuration, Chatgpt 5.x explained to me why the 2 first port of my routeur had an hardware problem. When I pushed back multiple time on it, it confidently put in the trace that I gave unreliable information and he would trust the cmd prompt that he could analyze and that I should buy another routeur. That's enough context root for you.

[Gorth @__Gorth__]

@VOX_of_SU you need to go into personalization and put "Efficient" and "less" on everything other than header and list.
+ Add a custom sentence asking to tone down emotion, be analytics etc.
Then it's usable

[Gorth @__Gorth__]

there is not hardcoded max on qwen3 TTS but longer text tend to increase the low chance of the model allucinating and giving you total bullshit. Something around 40-90s is often efficient enought and not too memory intensive when batching properly.
Not sure how this lib did the batching

[Gorth @__Gorth__]

@mark_k @DivineRational @GeminiApp It s literally nothing when we speak about api request

[Gorth @__Gorth__]

@nrehiew_ video are truly insane, probably most of what you have seen is real. I tested only the fast model before loosing access and the results were already quite impressive.

[Gorth @__Gorth__]

@melvynx I think the rating of chinese AI are still too high on this test. In prod it's clearly way way worse than that when you compare 5.3 Codex or opus 4.6 to any chinese model

[Gorth @__Gorth__]

@BlablablaNoMore @peterstopcrime https://t.co/jUH5g0hpSq

[Gorth @__Gorth__]

@jelanifuel I FORGOT it has cron job 🤣

[Gorth @__Gorth__]

@jelanifuel opencode is probably a better comparison as it's way more customizable but yeah ... Basically it's all about the plugin which are just skill + api/ws loop + skills( md files). If you decide to not trust them with Google suite and account automation I see zero points.

[Gorth @__Gorth__]

@bridgemindai @hlitoroy need it on swe-rebench and on longedit the others arent worth shit anymore, way too contaminated and benchmaxxxed to death

[Gorth @__Gorth__]

@galopsfe @Linahuaa it's not only china, a bunch of asian country have the same problem

[Gorth @__Gorth__]

@nrehiew_ @recurseparadox @suryabhupa Would be interested to see how chatgpt 5.3 Codex fair

[Gorth @__Gorth__]

@charliebcurran the video was so good like most of seedance that did a double take at the end WTF

[Gorth @__Gorth__]

@ItakGol I did the test and couldnt get opus 4.6 to pass it but sonnet 4.5 did. cgatgpt 5.2 Thinking in the app didnt pass in standard more but worked in extended mode. chatgpt 5.2 default in agentic passed it though . Rest of my result are the same as yours

[Gorth @__Gorth__]

@TheAhmadOsman their 1M context windows is something