Qya Hibali

[Qya Hibali @QHibali]

@TJeparskis I saw a thing calls "memories" running at the background. using 5.6terra

[Qya Hibali @QHibali]

@TokenGremlin old usage limits are not enough

[Qya Hibali @QHibali]

@thsottiaux shame?

[Qya Hibali @QHibali]

give up Codex updates, enjoy Claude Code, touch the grass

[Qya Hibali @QHibali]

@buildwithrajath possibly stay for image gen or computer use, but other providers are being good at those

[Qya Hibali @QHibali]

@mr_easonyang at least opus4.5, I've done amazing work using 4.5, but not seeing much improvement in 4.6. most of my workmates started from 4.6

[Qya Hibali @QHibali]

@thsottiaux seems like another trade-off

[Qya Hibali @QHibali]

@designwkarthick feels like, same. both run out my usage in less than one day. no 10x-like things happened

[Qya Hibali @QHibali]

@thsottiaux nah, I think it's probably due to the translation. On my side, the double usage reads like "5 hour limit back and each week has two 5hour limit"

[Qya Hibali @QHibali]

the new feature in Codex, Composer prediction, the "tab", is very important. It might be a key factor for the significant improvement of opus5.5.

As the accuracy of predicting what users would say to continue their dialogue increases, they could no longer need waiting for user instructions but solved them directly. Even us normal developers know to automate (using skill/hook/specific agents) some duplicate review model outcomes/continue dialogue, dealing with this thing from model/harness perspective could be much more influential.

this thing is a key to deal with the "proactive problem". cursor and claude have had such feature for a long time, this might explain how they made their recent gorgeous products. grok bot, which seems more rely on orchestration, or harness, and opus5.5, which seems tried to solve this problem model-wise.

If oai only just started caring about such features recently, I bet we could only get "better" (than opus5.5) oai models in their next generation models.

This might explain why people, like me, thought both models are powerful, but opus5.5 could do things better. they should be equivalent in most aspects, but opus5.5 wins in this and the visual quality/front-end design side.

Small features before started to show great influence today.

[Qya Hibali @QHibali]

@Endercloud513 @situ2oo1 至少可以加禁用词检查。还有些常见的怪字/词，也能找到比较��适的人话对应

[Qya Hibali @QHibali]

@Da7_Tech grok bot

[Qya Hibali @QHibali]

@TokenGremlin a better model after 2 months is a bad model

[Qya Hibali @QHibali]

codex, canceled, hypemaxxing

[Qya Hibali @QHibali]

remember they are the people who uses unlimited Astra/GPT7 Ultra Ultra Fast. they may be never ever thought their models are both stupid and slow than opus5.5

[Qya Hibali @QHibali]

@bot guess what will tibo provide at the day 4/28

[Qya Hibali @QHibali]

I hate this, but time chooses Anthropic