Eric Khun

am i the only one noticing opus is inferior to 4.5? tried codex 5.3 and it yields a lot better results than opus?

would you do it that way?

Vibecoding is fun until it's time to vibe debug . But why bother? just throw out everything and restart from a clean state.

Do you "--dangerously-skip-permissions"?

i'm watching claude code on the toilet instead of scrolling reels and it's not even close
Meta spent billions trying to get your attention.
Anthropic did it with a coding tool , how come?

Accidentally typed a Pokemon name into Google and a whole game appeared right in the search results. I have to search every of them to catch them all 🤣 Love these contextualized Easter eggs

Tried Claude Code's "Orchestrate Teams" to let AI agents collaborating. Agents stepped on each other's files, tasks went to wrong roles, tests all failing.
"What do you want me to do? Cleaning up the team and doing it myself, will be faster"
Claude being honest with itself 🤣

we can see which one is developers favorite 😂

That's not the future. That's today.

If you haven't tried a coding agent yet, you're missing out.
Give it access to your logs, metrics, and source code. Ask it "why is this service slow?"

Watch it pull data, correlate events, and point you to the exact line of code. Even better? let it do automatically 🤯

Hot take: SRE/DevOps is becoming about putting coding agents in the right places.
Incident → agent investigates (logs/metrics/source code) → points at root cause → suggests fixes → writes post-mortem.
LLMs are imperfect today, but give it a few years (or months?), they'll do that easily.

I believe this is a new way to think about how we build applications.

Coding agents don't just help developers—they enable entirely new kinds of products.

- Traditional: Product decides flow → user follows
- Agent-native: User states what they want → agent figures it out

I just wrote about it 👇

Debugging tip: Codex 5.2 (ultra high) is surprisingly good at finding bugs.
I spent 4-5 hours stuck on a silent error with no output. Claude Code couldn't see it.
Codex found it in 10 minutes.
It somehow traced the issue even with zero error output. Worth trying when you're stuck.

exactly what he said.
if LLm isn’t doing (at least) 80% of your work today, you’re probably not doing it right
x.com/karpathy/sta...

First time seeing an smart tv being smart. got this screen after wifi went off

Alex Honnold is rumored to be paid 6 figures (~$500k) to climb Taipei 101 live on Netflix.
• Jake Paul: $40M
• Tyson: $20M
• Chappelle: $24M
• Tom Brady: $25-30M
One guy free solos buildings where a slip = death. The others punch or joke.
How much should he have been paid?

github.com/badlogic/pi-...

What if your app used a coding agent to figure things out?
No prompts. Just README + skills. Your app become a list of README files, and figures things out.
The really cool part: non-devs can modify the skills themselves. 🤯
That's @badlogicgames's coding-agent project

Do you know what caused this?

my blog google search queries. Spot the odd one out

Most stressful part of watching Alex Honnold climb Taipei 101? People waving and trying to attract his attention through the windows

What's the future of engineering if anyone can build an app in an afternoon?
Building was always the easy part. The hard part is keeping it running for years.
The first 90% is a demo. The other 190% is engineering.

If you think Claude or ChatGPT chat app = their coding agents (Claude Code / Codex), you're missing out.
It calls tool chains, browses your files, and with skills? Mindblowing.
That's why clawdbot is blowing up right now.
I barely open the chat apps anymore.

Some people just have different brain wiring.

This is a golden age of building.
Every project you ship with LLMs right now is subsidized by VC money. The tokens powering your app, your code assistant, your AI features—all priced below cost.
How long will this last?

Even more true for any app you're using today. Making an app just perfectly fits your needs has never been that easy. And we're about to see some really great innovative apps coming up soon (well they're already here...!)

Built a Claude skill for drafting social posts.

It knows Buffer's 100M+ post analysis, my learnings, X's algorithm tips—personalized per network.

A LinkedIn post shouldn't sound like an X post.

What would you add?

I can see Claude Code becoming an OS just by itself. no more client or frontend needed

New study: people who use ChatGPT for writing show weaker brain connectivity than those who don't.

But is "reduced connectivity" cognitive debt... or just efficient resource allocation?

What's your experience with AI and deep thinking?

Anthropic published Claude's "constitution" - their AI safety playbook.

Priority: safety > ethics > helpfulness

But only for "mainline" models. DoD/Palantir versions play by different rules.

And who decides what's safe? Anthropic does.