AI Interfaces Hide Numerous Subtle Tricks And Traps For Effective Use

I think my rule of 10 hours of using AI for real tasks is true, but a lot of people find the first hour hard and never learn what these systems can do before putting it in a small box (“kind of like Google”)

@emollick Could not empathize more with this post as a digital learning specialist at a middle and high school. Even just getting people to use thinking instead of flash-lite seems like a Herculean effort.

@emollick the only people who can really make Codex work well are software engineers!

@emollick teaching three people surfaces every assumption you never knew you baked in. brutal and correct

@emollick Teaching someone else is the fastest way to discover the hidden assumptions in your own workflow.

@emollick That's half the point of creating row-bot. Honestly its so easy, even my mom can use it.

Three things I explain to people learning to use LLMs:

- Question is as important as the answer, you'll get exactly what you're asking for, not objectively correct information - LLM will follow your instructions literally, most of the time. Define objective success criteria - LLM will not hesitate and will be wrong confidently, you will trap yourself if not accounting for this - ask it for adversarial self-review

@emollick 💯 That’s one of the reasons we started this series, Ethan. Would love to have you on sometime!

https://hugobowne.github.io/show-us-your-agent-skills/agent-skills/

@emollick Also the UI itself People who are not coders have no idea what git is So they don't know to search for the commit button Why not have it as a "save checkpoint" button or something similar so that so that more people actually use it or something like that Many other things as well

@emollick Doesn't help that many people seem to delight in when the LLM fails at their task. "See!! I told you they're dumb!".

@emollick what is an example of an intuitive service in this case?

@emollick Well, give them a CLI :-)

@emollick Don’t worry the ASI will build more intuitive views and use cases

@emollick the real skill is learning how to prompt, not just type questions

@emollick created this to help teams get over that first hour hump - https://chasingnext.com/teams/projects

meant to give people that aha moment by having them set up a project/notebook for something each individual works on. power of going just a little deeper pays off big!

@_coenen @emollick I disagree. I'm not a SWE. I'm an English teacher. I currently use Codex on brokerages, to make web-based lessons via Apps Script, to run simulations, etc. I'm not a gamer or influencer. My use cases directly impact my life. If you need anything, I'd say you need vision.

@emollick totally - I made a whole post about the next level, the second brain framework, and got very little traction. Most of these ideas haven't been really productized yet.

https://www.paolo-luciano.com/technical/an_ai_system_thats_useful_a_fundamental_shift_in_how_i_work/

@emollick A lot of people who I talk to "outside the (large) bubble" use AI like a natural language search bar. The chat interaction is viewed as a means to follow up, not a context loop jumping in and out of an ecosystem of tools. The "unintuitiveness" has analogs in human interaction.

@emollick every "ai is so easy to use" claim was made by someone who already knows the tricks. the onboarding gap is the whole industry's blind spot.

@emollick That's the good news — the barrier is a learnable habit, not a moat, so the floor rises for anyone who pushes past hour one. The divide won't be access to AI. It'll be who kept going. Hopefully we can all help spread the word on how easy it can be.