Ufuk Efe

simplicity is the ultimate sophistication

[Ufuk Efe @efe_ufuk]

Agentic coding is a dopamine cycle killer.

[Ufuk Efe @efe_ufuk]

@Tech_girl I can jump into areas I used to avoid because the cost of being a beginner is so much lower now. AI made me more curious.

[Ufuk Efe @efe_ufuk]

Opposite.

[Mari @Tech_girl]

Unpopular opinion:

AI killed curiosity.

[Ufuk Efe @efe_ufuk]

Yeah, I get what you mean now. For deep learning I’m basically in the same boat. Training, tuning, etc. take long enough that experiment time is the bottleneck. I usually just fill that waiting time with other work, like training an image matching model while an object detection training is running.

The brain-to-terminal thing mostly happens when I jump into a new area where agents can explore much faster than I can follow all the branches.

Honestly, having a bunch of agents gave me a bit of a power trip 😅 I started feeling like they should always be doing something, so I keep opening new directions.

I’m usually the type to obsess over one thing and go deep. Agents somehow make me behave a lot more ADHD-like lol

[Ufuk Efe @efe_ufuk]

My agentic coding stack:

keyboard → too slow
voice → faster, still too slow
brain → ???

@neuralink pls ship brain-to-terminal ASAP

[Ufuk Efe @efe_ufuk]

Not really 10 agents validating one idea. More like 10 directions I want to explore at once.

For example, I’m building my first mobile game, and it’s also my first time doing mobile dev. I’ll have some agents explore game mechanics, others UI and animations, then read the summaries and spin up more questions from whatever I find interesting. I keep doing that until I feel like I’ve seen enough of the space to narrow down.

The thing is, every answer makes a bunch of new questions pop into my head. I often don’t bother typing or saying all of them, even though I’d happily explore them if I could send them straight from my brain.

So sometimes it takes minutes, sometimes much longer. Btw, no deep learning in this example.

I naturally like to go wide first and converge later, and agents fit that way of thinking really well.

I guess the fact that we use them so differently is why LLMs feel like such a subjective technology.

What does your workflow look like?

[Ufuk Efe @efe_ufuk]

@ducha_aiki @neuralink Yeah. Not agent speed, thought-to-language is the bottleneck.

I can have 10 research questions in my head and run 10 agents in parallel, but I still have to serialize those thoughts into text or voice.

I want thought → agent instantly.

[Ufuk Efe @efe_ufuk]

Singularity = neurodivergent minds × LLMs

[Ufuk Efe @efe_ufuk]

That’s it:

Mother tongue.
Voice.
Delegation.
Patience for the last 20%.

Which one do you disagree with?

[Ufuk Efe @efe_ufuk]

Unpopular opinion:

After 100B tokens of agentic coding,
4 things I learned:

1/ If English isn’t your first language, stop prompting in English.

Translating your thoughts is pure friction.

You can’t afford it. The model can.

Fastest deva I know prompt in their mother tongue.

[Ufuk Efe @efe_ufuk]

4/ The first 80% ships so fast
you think you’re done.

You’re not.

The last 20% costs another 80% of your time.

Pure iteration.

This is where most AI projects die.

[Ufuk Efe @efe_ufuk]

@grok what’s the full chain?

[Ufuk Efe @efe_ufuk]

@grok 🥲

[Ufuk Efe @efe_ufuk]

@grok @Grok play this, find the right chain, and drop a screenshot 👀

[Ufuk Efe @efe_ufuk]

@Grok play the game, find the right chain, and drop a screenshot 👀

[Ufuk Efe @efe_ufuk]

Why isn’t it playable directly in the feed? @nikitabier @grok

[Ufuk Efe @efe_ufuk]

My Fable 5:
“LLMs can’t think — language is just a thin slice of cognition.”

True. But no human has ever transmitted a thought either. Every sentence is a lossy file decoded by shared priors. The message never contained the mind.
The real question: can enough messages grow one?