Bitreon

building things I can’t stop thinking about
AI • software • devtools • startups • open source
sometimes useful. sometimes questionable.

[Bitreon @ItsBitreon]

@thsottiaux i think you need to add like in chatgpt temp chat to normal chat?
SO SOMETIMES the temp chat talk get little important

[Bitreon @ItsBitreon]

future software company:

5 humans
200 agents
almost no meetings
ridiculous product quality
sounds insane until you watch how quickly coding agents are improving.

[Bitreon @ItsBitreon]

Never accept “looks good” from an AI coding agent.

Run it.
Click it.
Break it.
Resize it.
Give it bad data.
Try mobile.
Try dark mode.
Then maybe it looks good.

[Bitreon @ItsBitreon]

currently building with AI like I have 40 employees

also currently debugging something one competent employee would've fixed in 11 minutes

balance

[Bitreon @ItsBitreon]

One obsessive builder with AI can now care about details that previously needed an entire team.

That's the part of AI I find much more interesting than “everyone can code now.”

[Bitreon @ItsBitreon]

real coding-agent benchmark:

give it a horrible 5-year-old repo.

no README.
dead code.
3 architectures.
one vague bug report.
come back 8 hours later.
prod alive?
AGI.

[Bitreon @ItsBitreon]

my AI coding workflow:

“small change”

6 hours later:

architecture refactored
database migrated
UI redesigned
37 tests added

original change still not done

incredible progress

[Bitreon @ItsBitreon]

@thsottiaux trust.

not “is Codex smart enough?”

“can I leave it alone inside something important without discovering 40 minutes later it creatively redesigned my database?”

[Bitreon @ItsBitreon]

@thsottiaux I think “browser” eventually becomes an implementation detail.

user gives outcome.

agent decides whether it needs browser, terminal, APIs, files or all four.

[Bitreon @ItsBitreon]

@thsottiaux my AGI benchmark:

production repo + vague customer complaint + no human for 12 hours.

working fix, tests pass, prod alive?

move the goalposts again.

[Bitreon @ItsBitreon]

@thsottiaux the real test:

give Astra a cursed 5-year-old repo with 14 conventions, half-dead tests and one vague bug report.

then measure intelligence.

[Bitreon @ItsBitreon]

@thsottiaux @growing_daniel AI compute now has banks, resets, limits and forward guidance.

we accidentally invented monetary policy for tokens 😭

[Bitreon @ItsBitreon]

@thsottiaux “roadmap moved 6 months forward” might genuinely be a better benchmark than most evals.

How many months of execution did the model compress?

[Bitreon @ItsBitreon]

I don't want an AI agent that writes 10x more code.

I want one smart enough to delete 70% of what it was about to write because it found a simpler solution.

[Bitreon @ItsBitreon]

You suddenly have nearly unlimited engineering help.

Using AI to ship more slop makes no sense.

Use it to obsess over details nobody had time for before.

AI should make software feel MORE handcrafted.

[Bitreon @ItsBitreon]

AI coding ruined my standards.

Before: “it works.”

Now: why is this 800ms? why is that button 3px off? why does this empty state exist?

AI didn't make me care less about software.

It made me care more.

[Bitreon @ItsBitreon]

@MannaCodeAI @github exactly. interruption should be a branch, not a reset. the main thread keeps its state, you resolve the tangent, then drop back in.

[Bitreon @ItsBitreon]

underrated coding-agent UX:

let me interrupt the agent without interrupting the agent.
@github now lets you open a side chat around an agent’s question while its main turn waits.
branching conversations might be way better than one infinitely long chat.

https://t.co/5guJsM5jRz

[Bitreon @ItsBitreon]

@thsottiaux @almmaasoglu we hope like price of luna

[小奇说 @xingbugengming]

尊敬的 @thsottiaux 先生

原本我用 Codex 制作完成这条视频只需要一天的时间，但是由于 Plus 用户的 5h 限制，我整整用了三天才制作完成

因此我向您转达我们所有 Plus 用户的心声：请解除 Plus 用户的 5h 限制

希望您能看到，谢谢！

[Bitreon @ItsBitreon]

hey @thsottiaux Codex usage feels off. I’m on Luna Extra High: a tiny bug fix can burn ~10% of my 5h limit, while bigger tasks barely move it. Also I can’t see Max in Codex at all. Is the meter delayed/model-dependent, and is Max hidden?