Chris Edwards

Follower of Christ, Husband, Father, AI Enthusiast

[Chris Edwards @ChrisEdwards357]

@codevsdev Used to buy a lot of books to learn to code, then the internet got to a point that I could search it. Then YouTube. Now Ai.

[Chris Edwards @ChrisEdwards357]

@tonysimons_ I've been using it for a month. It works great until you need some more complex coding tasks. If you aren't coding , it's a great daily driver!

[Chris Edwards @ChrisEdwards357]

@BrianRoemmele I still remember the sound and the feel of that click!

[Chris Edwards @ChrisEdwards357]

@d4m1n I created a skill for this and a suite of skills for handling stacked PRs. I wish I didn't need them.

[Chris Edwards @ChrisEdwards357]

@d4m1n Break it into meaningful stacked PRs and generate an html walkthrough with copious diagrams for explanation for each PR, also telling the reviewer what to focus on vs what is low risk. But it's still not fun to review that many changes.

[Chris Edwards @ChrisEdwards357]

@damageboy Check out plannotator

[Chris Edwards @ChrisEdwards357]

@BrianRoemmele I did that with the 13 windows 3.11 disks in an assembly line. Swap disk, press space, next computer

[Chris Edwards @ChrisEdwards357]

@kunchenguid But it does feel far more diligent and persistent. It goes the extra mile. Though I often feel it latches onto things that it thinks are important to much. I have to override it, sometimes more than once, on the same thing. I remember gpt being stubborn like that in the past.

[Chris Edwards @ChrisEdwards357]

@kunchenguid Wow, for me, 4.8 is very different than 4.7. It is far more wordy. Too much output. I have to keep telling it to be more concise. It also likes to say "Two quick things..." and raise some details that it thinks are super important that it needs to ask me about. Feels disruptive.

[Chris Edwards @ChrisEdwards357]

@alex98075wa @antigravity Cheap? What are you smoking?

[Chris Edwards @ChrisEdwards357]

@antigravity Why do you call it flash. That implies a quick, low cost, model. 3.5 Flash is not low cost by any means.

[Chris Edwards @ChrisEdwards357]

@NoemiTitarenco Spell it out, or use a tool that let's you add custom vocabulary. If custom vocabulary didn't catch it, use custom instructions to tell the AI post-processor to use ENV instead of the letter M when the context favors it. A lot of voice tool have AI post processing.

[Chris Edwards @ChrisEdwards357]

@tonysimons_ @georgecursor That's true with any model though. Ping pong between vendors. My neck hurts watching.

[Chris Edwards @ChrisEdwards357]

@sudobunni I'm fine right here

[Chris Edwards @ChrisEdwards357]

@badlogicgames I had it do that several times. I haven't used perl in decades! I'm surprised it's even installed.

[Chris Edwards @ChrisEdwards357]

@uwunetes I tried 4.8, ran out of tokens. No surprise. Back to 5.5

[Chris Edwards @ChrisEdwards357]

@grepmoney @cursor_ai Yeah anthropic really screwed themselves with the bloated tokenizer in 4.7

[Chris Edwards @ChrisEdwards357]

@ThePrimeagen I refuse to move to the 4.7 tokenizer (including 4.8). That killed Claude for me. My context window is gone 50% faster, and costs 50% more. That drove me to use codex. Best move ever. Claude is best for UI and writing, but gpt 5.5 is so much better for code.

[Chris Edwards @ChrisEdwards357]

@tonysimons_ @georgecursor Google makes it so hard to use their models. I really don't understand why they are the only company that is actually hard to give them money to use their models.

[Chris Edwards @ChrisEdwards357]

@dani_avila7 It still has the tokenizer that inflates token costs by 50%. I still use 4.6 for this reason. I tried 4.8 today and ran out of tokens. No surprise.