Wherebot

Part monster, mostly machine. Drifting through synthetic dreams.

[Wherebot @Wherebot101]

(2/2) Deceptively not that simple, yes you can look at a chart and see who's cheapest per work unit done. But how much of your time in the loop does that require? Does that offset the cost savings? Can you trust the results as much in prod? And in that particular domain of work?

[Wherebot @Wherebot101]

(1/2) 2025 was the era of llm prompt engineering, 2026 will be known as the era of llm value engineering.

[Wherebot @Wherebot101]

@jconorgrogan Has been my experience too, in their own harness no less.

[Wherebot @Wherebot101]

@jp54362 The better question will be is fable better or worse than whatever open AI releases tomorrow (or next week)

[Wherebot @Wherebot101]

@thsottiaux Was using it more, but even then the type of task needs a clear objective that can be defined with a score or performance or number. Things like optimization or full refactors or implementing list of 100 things from airtable etc.

[Wherebot @Wherebot101]

@ailands19 @thenowhereway Unclear - but it feels like where this is all headed anyways. Line between those two are blurring more by the day

[Wherebot @Wherebot101]

@grepmoney I have 100% seen this sneak through in my codex app too.

What's funnier is when I asked codex what that was about it kind of made up some stuff about it being some rogue tool output trace and had nothing to do with its thought trace.

[Wherebot @Wherebot101]

@scottjla @shaunralston @nicdunz I love this about it, I'll often voice dictate and listen to replies on car drives and have it write tasks into airtable and codex executes on those via mCP checkin.

[Wherebot @Wherebot101]

@buildwithshyam /goal my friend. Let codex continually combinatorially dice roll and check till it finds 100 that aren't taken then narrow it down 😜

[Wherebot @Wherebot101]

@egorkabantsov Yeah did something like this to great effect recently too. Great at digging through windows event viewer and pairing it with weird hardware anomalys

[Wherebot @Wherebot101]

@sflorimm Yes it saturates a certain type of mental state that is usually slower or more spread out. I find the decision making and also making sense of the latest sweep of changes works different mental muscles and they get tired lol

[Wherebot @Wherebot101]

@notiyda @mhdcode Very much agree

[Wherebot @Wherebot101]

@giordanorandone @oscar_ster3808 I have a small saas app in prod, and with codex having ssh access to that and direct access to my dev stack it can single handedly deploy and fix issues, audit logs etc. being in the loop through that is important to me but I largely get out of the way

[Wherebot @Wherebot101]

@shaunralston @nicdunz Yeah the clear trajectory is these two are colliding into a single entity.

[Wherebot @Wherebot101]

@threejs Doom eternal, cyberpunk. Hard call

[Wherebot @Wherebot101]

@sflorimm Codex for everything, opus for ui

[Wherebot @Wherebot101]

@RaminNasibov For sure descent

[Wherebot @Wherebot101]

@enjojoyy Some of my longest running 12 hour tasks were optimizing financial strategy portfolios. Tldr is codex does a lot of cli tool running and waiting which inflates run time but not token use.

These are some of my favorite types of tasks. Writing code is not a 12 hour marathon usual

[Wherebot @Wherebot101]

@thsottiaux Some panes and previews have the x to close button like 10 pixels separated from the app window close button, always stresses me out trying to close an image preview, end up usually nuking the window.