My life mission is to make Korea proud.
I turn every moment into an inspiring story to reach that goal.
Money doesn't inspire for long, but a great story does.
> Quit Intel 2yrs ago, burned the boats.
> Love fast iterating and building.
> All my life, I cared about what others thought. Now I'm learning not to.
> Not fluent in English, but made 20+ close US friends in 3 months.
> Love crashing through any kind of barrier.
My personal page: sungman.world/
3
14
1,467
Sung retweeted
1/ OpenRouter's cheapest price is not your lowest inference cost.
That price was set for everyone who shares the endpoint. A stack tuned to your own traffic can cost far less.
What that takes, and what we've measured:
Sung retweeted
I talked with two people working in ai infra. same premise but different approaches.
Both agree on the one premise: the inference engine will be commoditized. Every kernel trick, scheduling algorithm, speculative decoding variant becomes public knowledge within months. Neither of them is betting on the engine.
But they are betting on opposite directions.
The first is going below the engine. He believes there will never be one inference engine that fits for everything. Interactive, long horizon and other workloads each want something customized. So he is redesigning the hypervisor itself, the layer that enables us to run many engines on elastic, variable-sized GPU VMs. As he has a superpower rebuilding this, he could build his own moat at this layer and afford to tell customers rip us out any time.
Another is going above the engine. He did not set out to build an inference company. Instead he built general execution infrastructure, queues, scheduling, autoscaling, and inference simply became one workload on top. Then sandboxes. Then agents. His moat is not only technology but also gravity; workflow relies on the platform, tooling depends on it, developers stick with this. If a customer switches providers every month, they will keep switching forever. He builds the thing they do not want to leave.
I think both are right. If you own something few people on earth can rebuild, the tech is the moat. If what you own can be commoditized, you had better build the platform and the relationships above it.
We have seen this movie before. VMware created the hypervisor, and for years it was a wonderful business. Then KVM and Xen commoditized it, and the durable value migrated up, to vCenter, to enterprise contracts. gVisor and Firecracker were open sourced by the very companies that built them. The one pure-technology moat that held was AWS Nitro, and only because it is welded to custom silicon that no startup can copy.
Now the same cycle is beginning again, this time for agents. Nobody owns the VMware position of the agent era yet. And the workloads are diversifying fast. Until now most agent workloads were coding agents.
Going forward there will be even more coding agents, plus training workloads like RL rollouts and evals, plus non-coding agents for support and research, plus orchestrators managing other agents. Each wants a different shape of infrastructure. What happened to inference engines is now happening one layer below. The winning abstractions will be different, and they have not been built yet.
That is what makes this moment so interesting. We are early. The stack is still up for grabs.
Need to think about combining model families, seriously.
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.
This is a 17-place jump from Kimi-k2.6 (#18 -> #1).
In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.
The full model weights will be released by July 27.
Congrats to the @Kimi_Moonshot team on this major milestone!
50
we need prediction market for predicting reset timing @Kalshi
1
40
👀 expected Claude for students
We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states.
claude.com/solutions/teacher…
52
Many people are interested in model routing.
Next thing will be building dynamic harness for each projects.
Maximizing the performance by using optimized harness + model routing.
github.com/onlycastle/openfu…
1
12
One of the best way to achieve successful AX is
Do hackathon inside of the company.
I talked with marketers, engineers, executives, designers in IT company that wants to do AX. But, what I observed is 90% of them are failed. reason isn't the single root cause. political issue would be the most critical one.
Do not expect external AX specialist can resolve all AX problems in your org. they can't understand what's really happening in your org. 1 time lesson is also useless. employees won't use the AI after that session.
Your employees know their problem. but, have no time to use AI tools to resolve their problem. easy to keep their work momentum than break it. From this perspective, hackathon would be the great way. your employees know the problem, and then know how your company orgs work.
1
126
At ICML and craving some great Korean food?
Stop by this place for chimaek
(Korean fried chicken + beer).
It's the same spot Jensen Huang visited.
maps.app.goo.gl/ZH76ktw3RLW8…
#ICML2026
4
435
+6 hrs, Added simple asset, and grid water flow effect and scenarios.
This is the team play game to keep town. save the AI NPC. each player has each role.
All work is done with Fable5.
Construct the harness for the game development.
Just started to build Sunny Street v2. 6:21PM, Jul 9. (KST)
I only have 1 hour experience in Unity. curious what will be happen? follow me and then check updates
1
4
209
Grok 4.5 is strong: not the #1 smartness, but super fast.
I've been using Grok4.5 for the development workflow. and here is my summary.
- Similar UX with Claude Code: it made spec -> implementation_plan -> implementation (sub-agent driven / sequential).
- Super fast : I know Opus4.8, GPT5.5 are smarter, but we need to wait long time to get results and sometimes it's the bottleneck. Grok4.5 is super fast. we can do more iteration with the same time.
- Text output is easy to read. love the colors. (left: Codex, right: Grok4.5)
Worth to try it. work well with previous harness, skills.
1
9
1,093
















