Jonas Feroz

@thorstenball Looking forward to the next one!

In case you missed the note in last week's issue: No newsletter this week. I'm 99% offline until Tuesday.

@ibuildthecloud That's fair and I do — but I also find there is a surprising amount of ambiguity in asking such a simple question. I can't quantify it but the model sometimes thinks too narrow or too broad and I don't want to play language games to make it think differently.

@jonasferoz Why not just tell the LLM you'd like to plan or design. That's what I do. It works fine.

Another one! I told you I'll keep retweeting people that agree with me on this. #PlanModeMustGo

@yacineMTB Turns out most humans aren't great at programming either. That's not likely to change, while models are improving while we sleep

These AI models are very bad at programming. If you trust them, you are going to have a very bad time

@ibuildthecloud I do feel like some of the current blowback to plan mode is in part annoyance with 5.4's incessant chattiness

@ibuildthecloud I love plain prompts for brownfield but with a new project I think plan mode has value, particularly for ambitious noobs like me. Also plan mode produces fairly consistent output and even if it's slop I know I can iterate down. I don't actually enjoy using it but it works.

@ibuildthecloud Codex versions are less chatty. Some folks prefer 5.3-codex over 5.4 for that reason. Amp code also did a good job getting 5.4 to behave more like a codex model with their Oracle subagent

@ibuildthecloud hoping they release the codex version soon

wow, ask gpt5.4 for a design and it write a novel.

@mitsuhiko @1st1 Unexpected. Kind of a different format. Would be cool to see more like it.

Something different for the weekend. I talked with @1st1 about lat. He built a knowledge graph that connects specs with code and integrated it into pi and other harnesses. It's an interesting experiment.

@reach_vb Codex for Accessibility

Codex for Open Source Codex for Students Codex Ambassadors Codex Security What should we do next?

@thorstenball When the future is finally finished being built you should host a long form podcast where you discuss each joy & curiosity.

"Everything is changing!" Couldn't help myself. New Joy & Curiosity is out: registerspill.thorstenball.com/p/joy-and-curi…

@thorstenball My brain patched right over it

WHY DID NO ONE TELL ME! I left an older paragraph in there that I actually rewrote, so now I had two paragraphs in there! I guess it's bonus content: you now see the first and the second version of it.

@krzyzanowskim 5. People who are never going back to the ways things were before. The technology doesn’t even matter it’s the people that change everything. 1 - 4 can all still be true and in 6 months or whenever everything changes. That doesn’t mean change is pretty.

There are 4 types of people saying "AI will replace Software Engineering" in 6 months 1. they own a company that benefits from replacing engineers (ceo, stockholders etc). they need to justify the amount invested into them 2. Company that made big cuts and they want to attribute it to whatever and AI is good excuse 3. People who never worked as as/with Software Engineering and have no clue the work is not just typing 4. Burned out engineers who don't want to do that job anymore

@thorstenball That's winning at life

Big issue in our house: 8-year-old running out of books. Four library cards and she's running out. 30 books in 3 weeks. Including first two Harry Potter books.

@ibuildthecloud Remember when reading and writing were only for a select few? Access didn't kill its value. Maybe when software is no longer a medium for work and consumption, its value will plummet. I think we have a ways to go.

I'm calling bankruptcy on software. It has no value anymore. It's not special, everyone can do it now.

So many variables. The model, task, prompt, your expectations, experience, and time you’ve dedicated to learning them. Models are general but not oh shit amazing unless most of those align. What’s in your control is the time + model. Use the best model you can and put in the time with an open mind.

@thorstenball I’m just not that confident in the capabilities of AI. My personal experience, with small model which might be the problem, show that, Ai isn’t even capable of reading and understanding docs, then fixing the errors terraform plan spits out. I have the feeling driving myself is

Having the ohshit moment and realizing that AI is going to change software dramatically is a one-way door. Why haven't more walked through it? Some thoughts at the start of today's Joy & Curiosity. registerspill.thorstenball.com/p/joy-and-curi…

@thorstenball Our tools should evoke joy; otherwise, what the hell are we doing?

This is why I get up in the morning

@thorstenball Ship it! Also if the thread can be declassified could be enlightening

theme: switch I'm kinda shook. Amp's new experimental deep mode built this today. I didn't write a single line of code. I gave it a big prompt at the start with strategy for how I'd implement it and then nothing but handoff and 2-3 "continue" and small messages. Ship it?

@doodlestein Pay for agency or pay to give it up. It’s definitely not college that will set you free.

Hmm... one year of college, or 35 Claude Max and GPT Pro accounts? Which would create more value for a smart, ambitious young person with the drive and interest to actually use them to the fullest? It's not even close. And isn't the social life at schools kind of lame now anyway?

@ibuildthecloud Would be sick if you shared your thoughts on it at some point. It might not be your cup of tea but I feel like it's hard not to respect the audacity and earnestness of it.

Gas Town is a crazy long rant. Things like this typically boil down to one or maybe two good primitives. I've got to think more about this but worth a read if you want to get seriously confused. steve-yegge.medium.com/welcome-to-gas…

@mitsuhiko There has to be a sweet spot between agent psychosis and analysis paralysis. For shared systems, we should probably err on the latter; individually, it might be okay to go a little wild.

Weekend thoughts on Gas Town, Beads, slop AI browsers, and AI-generated PRs flooding overwhelmed maintainers. I don't think we're ready for our new powers we're wielding. lucumr.pocoo.org/2026/1/18/agen…