Follower of Christ, Husband, Father, AI Enthusiast
Ohio
Joined May 2008
- Tweets2.7K
- Following513
- Followers213
- Likes1.2K
Chris Edwards retweeted
It's absolutely wild how a single X post can erase $3 trillion from the US stock market.
Let the Claude Code UI clutter begin!
So I saw this unexpected clutter in my Claude Code UI this morning.
I hope plugins don't start cluttering my UI like this.
I uninstalled for now.
(Good plugin...but not worth my screen real-estate)
I tried to get my factory to design good abstractions for the code it writes. However I am finding that abstraction locks in the assumptions that motivated it. When that assumption changes, or if the AI got it wrong (it will), the abstraction is much harder to change.
It has been said, "A bad abstraction is worse than no abstraction at all."
Abstractions should be used only for things you know are not going to change. The AI does't have this foreknowledge, so it's poor at predicting useful abstractions that last.
This changes my thinking.
When intelligence is "good enough" in open/cheap models (which is just about here), there will be a huge drop in the demand for frontier models.
I already have MUCH less need for them.
I absolutely hate 5-hour limits. It destroys mental health by dragging work outside normal hours just to "use" what you paid for. This is not healthy for us humans.
I just love that Claude is erroring out with 500s left and right and the Anthropic status page says "No incidents reported today". Total BS.
I think agent memory is a great idea. But it makes it hard to do harness engineering if you work on a team. Does the agent work well because of the harness or your memories? You want a consistent harness, but everyone's memories differ. It's the modern "Works on my box!"
When code reviewing, I have seen Sonnet 5 on high use so many tokens it costs more than Opus, and finds fewer issues. I guess it just doesn't know when to stop.
If planning with a smart model, don't let it be too prescriptive and detailed about the implementation. Otherwise the agent that implements it may follow the details even when it finds they are wrong. No plan survives contact with the code. Leave flexibility for the implementer.
I'm really getting frustrated with the way Claude Code has been hiding all the tool calls it makes. It's so hard to tell if it's stuck or still calling tools. This grouping of all calls under one aggregate heading is not the way!
This really gets you to think. Very capable models are getting much cheaper. While the highest capability models keep getting more expensive. The ability to leverage low cost models will increasingly become a huge cost advantage. The days of "one model fits all" are numbered.
I'm using Luna medium as my daily driver in Hermes now. It is dirt-cheap! About the same price as DeepSeek V4 Flash. That's really impressive! Loving it so far.