yifeng weng

[yifeng weng @GeekerStack]

@loktar00 But deep down, do you still feel that there’s still a big gap between ds and Opus?

[yifeng weng @GeekerStack]

@chaosering 似乎今天是官方本来就修复了？我什么都没做也好了

[Nexp @CodexResets1]

🚨Codex Users!!!

Day 5: Composer predictions in the Codex desktop app.

[yifeng weng @GeekerStack]

@thsottiaux @ZryMiller This is the delayed benefit of the credibility OpenAI built up previously; once that credibility is exhausted, even the best strategies later on may be powerless to turn things around.

[yifeng weng @GeekerStack]

@thsottiaux Haha, the fact that a company would pull a bait-and-switch on users with its own core product shows it’s not far from self-destruction.

[yifeng weng @GeekerStack]

ChatGPT will lose its competitive edge within six months, and its user base will halve. They will have to overhaul their strategy for attracting users, but the momentum can never be regained, and the meager advantages they can still offer no longer excite anyone.

[yifeng weng @GeekerStack]

@guk2472 @yanhua1010 May I ask what harness is used?

[yifeng weng @GeekerStack]

@demn_z @thsottiaux same

[yifeng weng @GeekerStack]

There are many ways to let models carry out engineering project patterns—for example, should every step be carefully crafted from 0 to 1, or should you first establish the outline and then refine the details? That is an art in itself.

[yifeng weng @GeekerStack]

LLMs should always be split into at least two agents: one acts as an inspector of the prompt’s core essence, constantly scrutinizing whether the other agent is making effective and efficient progress.