John Kim

I do software stuff and make content Staff Software Engineer @ Meta

[John Kim @PremiumGoblin]

Hi, I’m John! I’m a L7 Senior Staff software engineer at Meta.

My weird hobby is making AI Tech content on YouTube and Substack and I wanted to start sharing more here.

In the past I’ve worked on Reels, Meta AI on Msgr and currently working on Threads.

I also teach Agentic Coding with @bytebytego for fun on the side.

Doing an AMA here to share and meet new people. 👋

[John Kim @PremiumGoblin]

@damnGruz No this is what AI does to you

[John Kim @PremiumGoblin]

@florin_dev Yaaaaas

[John Kim @PremiumGoblin]

Looks even more like one…

[John Kim @PremiumGoblin]

This is what codex micro should have been. Give me this pls.

Not gonna lie, this thing looks like a dildo…. But you get the point lol

[John Kim @PremiumGoblin]

@ashebytes @OpenAIDevs I have a codex micro and it keeps disconnecting. It doesn’t just stay connect :(

I just want it to wake up right away and start recording when I hit the voice button

[John Kim @PremiumGoblin]

@yashmp2004 I disagree Notepad is the best

[John Kim @PremiumGoblin]

I’m turning 38 tomorrow.

I know, my Korean genes makes me look like a teenager, but I have two kids.

Was reflecting on getting one more year closer to 40 and I remember this post I wrote a little while ago.

This is my most popular essay on Substack.

It’s about loosing the joy to enjoy things as you grow up.

Let me know you think.

https://t.co/wZa6XS0jGh

[John Kim @PremiumGoblin]

@ethanisthinking Yeah its a Keychron but its a Q11 split keyboard

[John Kim @PremiumGoblin]

People have been asking what my setup is like… honestly I lost track of all my stuff.

[John Kim @PremiumGoblin]

@CXCarroll Yeah I don't think Googles models are that bad. I think this tier list is a little harsh but I'm like 90% aligned with it.

Google is still S tier company imo.

[John Kim @PremiumGoblin]

Google it’s time to make a come back. I have too much $GOOG plsss 😭

Jokes aside I think Google is just fine. They have a Billion users on Gemini and their entire TPU stack is really hard to ignore.

We are also approaching levels of intelligence where “good enough” is actually good enough for most task.

I remember being completely off manual coding since Opus 4.5. People just forget how amazing that model was. Every single model here is better than 4.5.

Google is going to be just fine.

[Theo - t3.gg @theo]

Rough tier list of where I'd put every major model right now

[John Kim @PremiumGoblin]

I feel like doing stuff like this is a lot harder than it actually is.

The hard part of starting AI agency is not the tech but the marketing and sales. People always grossly underestimate how much work the ops or distruption side of a business is.

I've built so many side projects into production that had 5 users.

Disruption is hard, sales is hard, running a business is hard.

But with that said, anyone want to pay me 3K to run some AI Agents? I'm really good at it.

[Nick Vasilescu @nickvasiles]

talked to a guy last week charging $3,500 per agent to local businesses. not per month. per agent.

he's deploying computer-use agents that log into the client's actual software and do the repetitive stuff their office manager used to do. invoice processing, appointment scheduling, follow-ups.

no one on twitter is talking about this because it's boring. that's exactly why it works. boring businesses have budgets and they don't churn.

[John Kim @PremiumGoblin]

@TheLarkInn When did this happen!?

[John Kim @PremiumGoblin]

@theo Lines up with me :)

[John Kim @PremiumGoblin]

@remymount This is my side machine just for model training and gaming. I forget the exact specs. Keyboard I use Keychron. This one is an older one but I usually use the split one for my main setup

[John Kim @PremiumGoblin]

Fuck this video is good

I guess I have to install Omarchy now 😂

Gigatron needed a change anyways.

[NetworkChuck @NetworkChuck]

You need to switch to Linux RIGHT NOW!! (ft. @dhh , @typecraft_dev and @OmarchyLinux )

Also, huge thanks to @tobi , @patrickc , @MichaelDell , @jack , @eastdakota , @brendaniribe , @jasonfried and @dhh for donating a shit ton of money to make this the #yearofthelinuxdesktop

btw, uploading the full video to @X from youtube. ----> Will ya'll watch it here?

Let's see if we can impress @elonmusk

[John Kim @PremiumGoblin]

@TheLinuxITGuy Quake II. A real one ☝️ nice

[John Kim @PremiumGoblin]

Aaand I’m in. Took 3 minutes to install Omarchy. Not 60seconds haha. But still awesome.

Changing to Matte Black to fit my black out rack.

What theme yall setting? Do we have a theme market place yet lol

[John Kim @PremiumGoblin]

@pankajsameold Honestly very smart. They recognize they don’t need to win in AI. They have the interface people want for now

[John Kim @PremiumGoblin]

I think we are only a few model lifecycles away from being able to run everything locally if we wanted.

Everytime we make large jumps in the frontier we are able to distill and quantize them into powerful and useful local models.

Once a "Task" is solved at a certain intelligence level we don't need better intelligence for that task.

There will come a time when local models off your single GPU will hit the current SOTA.

If that happens what will you use?

[Yuchen Jin @Yuchenj_UW]

UC Berkeley open-sourced FreeToken. Wild results:

A single RTX PRO 6000 runs the 753B GLM-5.2 at 14.9 tok/s!

An 8GB RTX 4060 laptop (~$1,000) runs Qwen3.6-35B at 39.3 tok/s!

FreeToken is 2–4x faster than Ollama across consumer GPUs. Local AI inference is getting very real. Great work by @Andy_ShuoYang and UC Berkeley Sky Lab!

[John Kim @PremiumGoblin]

@yaredow True. I’m just sold on the idea that I can build the laptop too. Now it’s just about playing with new toys

[John Kim @PremiumGoblin]

Using other people’s skills will always be suboptimal then making your own skill. Everyone works different and in different contexts. Good skills are great to learn from or fork from but if things are annoying for you personally, it’s because you are not Matt and what he finds important you may not agree with.

Also your gut instinct probably sucks. You can literally just test your claims by A/B testing with two terminals or even create some benchmarks runs that uses a skill vs one that doesnt.

Just my two cents haha. 🤣

[John Kim @PremiumGoblin]

@kunchenguid My high level bias is that most people haven’t drove the two enough to have a real informed opinion.