iiinova — where your digital self keeps evolving

Sungbin Jo @goranmoomin.dev · 25/08/2026

i am very sad that this lsp issue never got resolved in any significant way, as: > this discussion is irrelevant for the actual user experience: given that humans are slow, I expect real race conditions to be extremely rare in practice. got invalidated with agents. github.com/Microsoft/la...

github.com

Protocol consistency gurantees are subtly incorrect · Issue #584 · microsoft/language-server-protocol

Hi! This is a rather handwavy and philosophical issue, but I've decided to log it just in case. One of the requirements for the LSP is to synchronize world view between the client and the server, s...

000

Reposted by Sungbin Jo

Ryan Carniato @ryansolid.bsky.social · 24/08/2026

I still find it crazy how we are still discovering primitives that completely collapse decades of carefully crafted imperative code. Part 2 is probably more impactful as a live demo. But picture it paints is still very powerful. www.solidjs.com/blog/async-s...

solidjs.com

SolidJS

Solid is a purely reactive library. It was designed from the ground up with a reactive core. It's influenced by reactive principles developed by previous libraries.

3452

Sungbin Jo @goranmoomin.dev · 22/08/2026

this would be so funny if… this was a joke not something that i have to encounter every single day (sobs quietly)

000

Sungbin Jo @goranmoomin.dev · 18/08/2026

i know this isn't any original thought, but i feel that llm agents just totally suck on thinking one step back on global invariants (i.e. they are good at one-shot-designing systems, but bad at rethinking invariants or redefining existing semantics). which matters a lot in real world engineering

130

Sungbin Jo @goranmoomin.dev · 16/08/2026

@kaylees.site by chance, do you have any plans to make a standalone Filtr app? The use case would be that I would love to use system-wide URL blocking, yet I don’t use Safari (I use the Quiche browser), so was wondering if it can be split to a separate app by any chance!

010

Reposted by Sungbin Jo

Sung Kim @sungkim.bsky.social · 14/08/2026

Why LLMs read fast but generate slowly? Inference Performance from First Principles by Eric Schreiber aleph-alpha.com/en/blog/infe...

2405

Sungbin Jo @goranmoomin.dev · 14/08/2026

this feels so simple and so intuitive that i now feel dumb… tldr: for a llm classification task, ofc you can use constrained decoding (a.k.a. structured output) but you can also have the llm hallucinate a class, pick the actual class by running text embedding on it and picking the closest one.

softwaredoug.com

Don’t classify. Hallucinate!

A useful pattern for LLM classification at scale is to let it hallucinate plausible, fake entities. Then resolve to real ones.

012

Reposted by Sungbin Jo

Duncan McAlester @mcalester.design · 08/08/2026

IT’S HAPPENING!! Safari Technology Preview has finally enabled Temporal API support!! I am unreasobaly excited by this API developer.apple.com/documentatio...

0155

Sungbin Jo @goranmoomin.dev · 14/08/2026

I was reading @raphlinus.bsky.social’s article raphlinus.github.io/gpu/2025/03/... the other day, and this seems similar what the article describes as ‘Big grid of cores: Cell reborn’ as one potential future path of parallel general purpose computers.

raphlinus.github.io

I want a good parallel computer

The GPU in your computer is about 10 to 100 times more powerful than the CPU, depending on workload. For real-time graphics rendering and machine learning, you are enjoying that power, and doing those...

120

Reposted by Sungbin Jo

mr. TIM @timkellogg.me · 14/08/2026

Cerebras is kind of a strange programming model simple? foreign? dataflow? ya idk, would be fun to see how well i could do

wafer v @wafer_ai
X.com
GPT-5.6-Sol Ultrafast mode (750 TPS) is
here powered by Cerebras!
Here's How kernels work on Cerebras Chips
cerebras built a chip with 900,000 cores on a single silicon wafer.
CSL (Cerebras Software Language) is a Zig-inspired DSL that gives you direct control over the Wafer-Scale Engine. the SDK is publicly documented and the programming model is genuinely the most alien one i've ever looked at. if NKI felt different, CSL takes it much much further.
in CUDA you write one thread's perspective and launch millions. in CSL, there are no threads, no warps, no shared memory, and no kernel launches. you write code for individual Processing Elements: 900,000 independent cores arranged in a 2D mesh on a single silicon wafer. each PE has its own 48 KB of private

wafer. each PE has its own 48 KB of private
SRAM, its own program counter, and a 5-port router connecting it to its 4 neighbors. that's actually it. no DRAM. no HBM. no cache hierarchy. 48 KB is your entire world per PE, where code and data must both fit.
the programming model is dataflow, and this is where it gets a bit more alien
data moves between PEs as 32-bit messages called wavelets, traveling along virtual channels called colors. when a wavelet arrives at a PE on a specific color, it activates a task (a chunk of code bound to that color at compile time). tasks run to completion, then hardware picks the next activated task. tasks cannot call each other. they can only be activated.
so instead of "launch N threads," you think of it like: "place code on PEs, define routes, let data flow."
the memory model is also very different from GPUs.

on an H100 you get 80 GB of HBM shared across all SMs. on the WSE-3, memory is 48 KB per PE, and there are 900,000 of them! this gives you
44 GB total on-chip SRAM at 21 PB/s aggregate bandwidth (vs 3 TB/s on H100). every access is single-cycle. no coalescing needed. no bank conflicts. no cache misses. but also no way to access another PE's memory. all inter-PE communication is explicit wavelet routing through the fabric.
take a distributed GEMV for example. you would write a layout file that physically routes wavelets across the mesh. two PEs sit side by side. the left
PE computes a partial result and routes it eastward. the right PE receives those wavelets from the west and accumulates. routing is defined at compile time. both operations are asynchronous and activate a task when they finish. you're physically routing data across silicon at 1 clock cycle per hop.
cerebras is an extremely technically interesting & alien beast. reports 95-210x speedups over
H100 on stencil computations. 3,000 tokens/ sec inference on gpt-oss-120B. $10B+ deal with OpenAl for 750 MW of inference infrastructure. very exciting times for alternative accelerators!

9706

Sungbin Jo @goranmoomin.dev · 28/07/2026

this would allow us to be able to do 2035-era things like: - copy text from a tui coding agent (e.g. pi, claude code) without having to trim whitespace from every god damn line of the selected text but how do we start proposing terminal escape sequences? does anyone know

120

Sungbin Jo @goranmoomin.dev · 28/07/2026

random thing, but we rly need some terminal protocol to allow text selection and copy/paste between blocks of text, so that we don't need tools.simonwillison.net/cleanup-clau.... prior work: the osc 8 hyperlink id param: gist.github.com/egmontkob/eb... and gitlab.freedesktop.org/terminal-wg/...

gist.github.com

Hyperlinks in Terminal Emulators

Hyperlinks in Terminal Emulators. GitHub Gist: instantly share code, notes, and snippets.

001

Reposted by Sungbin Jo

jfc @jfc.ninja · 15/07/2026

I realized earlier I'm treating my Fable usage like potions in video games "hmm, this problem seems hard. but what if something harder comes along and I've already used all my Fable?"

0492

Reposted by Sungbin Jo

Armin Ronacher @mitsuhiko.at · 16/07/2026

I wonder if this will make people reconsider their stance or become even more defensive.

1317318

Sungbin Jo @goranmoomin.dev · 13/07/2026

we need a way to compile a script with a directory of assets or other scripts into a single binary. unfortunate that this doesn't exist…

000

Sungbin Jo @goranmoomin.dev · 08/07/2026

oh by any chance, if anyone is around seoul or coex for icml 2026, more than happy to connect! dms are open :)

000

Sungbin Jo @goranmoomin.dev · 05/07/2026

ouch this hurts… i guess i can claim a status on this one :)

A coding agent session, with the model set as (anthropic) claude-fable-5, with the following text:

[user] look at @PLAN.md and try to figure out if there's anything that you feel is off or something that you would do differently. is there any feedback that you would give?
[assistant] I need to check the PLAN.md file to understand what I'm working with.
[assistant] read ~/Developer/Personal/melon/PLAN.md
[assistant] Error: This request triggered restrictions on violative cyber content and was blocked under Anthropic’s Content Usage Policy. To learn more, see https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback.

010

Sungbin Jo @goranmoomin.dev · 05/07/2026

It’s pretty funny that people who love the X protocol the most also hate web apps the most. Think about it, isn’t the web basically the most successful network-transparent display protocol? A server provides instructions to draw graphic primitives and the client renders them on the display.

100

Reposted by Sungbin Jo

Pirijan @pketh.org · 22/06/2026

One thing led to another, and I spent way too long making the nicest timezone picker UI, which most people will only need once if ever. It supports strumming the world map, searching by all major cities, and auto detection. Way beyond scope, but my inner Carmen Sandiego demanded it.

1171

Reposted by Sungbin Jo

olofkindgren.bsky.social @olofkindgren.bsky.social · 16/06/2026

Tired of reading yet another lukewarm take on access to AI models? Want to read some real news instead? Well, you're in luck, because I just made a write-up about the latest release of the award-winning SERV, the world's smallest RISC-V CPU. blog.award-winning.me/2026/06/serv...

blog.award-winning.me

SERV 1.4.0

SERV 1.4 is finally released. Actually, it has been out since October last year but I've been so busy that I never got around to make a prop...

062

Sungbin Jo @goranmoomin.dev · 23/06/2026

i found another bug in llvm while doing so, and got another PR merged, and now i want a third one, and i'm feeling the urge to handle a good-first-issue in the llvm repository, probably some instcombine optimization or sth like that…

010

Reposted by Sungbin Jo

aly @aly.codes · 23/06/2026

ive seen in-line python from LLMs before but why the fuck is codex using ruby lmao

6692

Reposted by Sungbin Jo

Sakana AI @sakanaai.bsky.social · 22/06/2026

Introducing Sakana Fugu: A full multi-agent orchestration system accessible via a single model API. Our ‘Fugu Ultra’ model matches the performance of Fable and Mythos, delivering frontier capability without the risk of export controls. Try it: sakana.ai/fugu 🐡

48921

Sungbin Jo @goranmoomin.dev · 22/06/2026

It's pretty funny to look at thinking traces from kimi k2.7 code: it thinks like Kevin's small talk from The Office: www.youtube.com/watch?v=_K-L.... One of the features of kimi-k2.7-code was optimized thinking (platform.kimi.ai/docs/guide/k...), and it seems the model got trained caveman speak :)

A screen capture of a coding agent session, showing Kimi K2.7 Code's thinking trace and tool calls with the following text:

```
Let's check if the local branch exists. git branch maybe. We can list.

$ git branch -a | grep format

chore/format-ci-check

Took 0.0s

Local branch exists. Good.

We should not push or delete without confirmation. The user said not to push. We can ask whether to delete the local branch. But they want how to test. We can explain.
```

000

Sungbin Jo @goranmoomin.dev · 21/06/2026

became curious: is there a way to delete a follower in bluesky? like how twitter allows deleting followers with block & unblock?

000

Sungbin Jo @goranmoomin.dev · 21/06/2026

I do hate Apple's software lockdown policies on iOS, but then I also see the Apple-enforced Privacy Requirements of browser engines, and then maybe Apple does have a point… Still, Apple could have just opened BrowserEngineKit on all regions, which would have saved a lot of grace… around this

020

Sungbin Jo @goranmoomin.dev · 20/06/2026

this actually in fact, got merged! :) i now have a contribution in llvm which makes me infinitely happy…

010

Reposted by Sungbin Jo

hailey @hailey.at · 20/06/2026

correct. no code - whether written by a human or an agent - should be getting reviewed by a human without being first reviewed by another agent. imo that includes reviewing yourself! respect your own time and dont review the agents work until it’s asked for review from another model.

724523

Reposted by Sungbin Jo

ponder @ponder.ooo · 20/06/2026

adding ".vibe" before the main file extension like "file.vibe.js" to indicate i haven't read the code

31135

Reposted by Sungbin Jo

antirez @antirez.bsky.social · 15/06/2026

Another important thing: Chinese models are not strong because they distill US models. Distillation of models via API is *impossible*. If somebody tells you the contrary, they don't understand machine learning:

119911

Sungbin Jo @goranmoomin.dev · 16/06/2026

> The output speed of frontier models … developers often spend more time waiting … than on prompting and reviewing code. I feel the same thing from models of Anthropic/OpenAI, and started using open-source models from other providers. Iterating on code with 300tok/s coding agents is awesome.

122

Reposted by Sungbin Jo

Xe @xeiaso.net · 15/06/2026

how do i even parody this shit

720115

Reposted by Sungbin Jo

amos @fasterthanli.me · 15/06/2026

!!! Anthropic is walking back "can't use subscriptions for programmatic use", whoa

We're pausing the Agent SDK credit change
Claude
Hi Amos,
In May, we sent you an email announcing that starting today, the Claude Agent SDK, claude
-p, and third-party apps built on the Agent SDK would stop drawing from subscription rate limits and move to a dedicated monthly credit. We're writing to let you know that we're not making this change today. We're working to update the plan to better support how users build with Claude subscriptions.
What this means for you
Nothing changes for now. Agent SDK, claude
-p, and third-party app usage continues to work with your subscription exactly as it did before today, and there's no credit to claim. Your

61487

Reposted by Sungbin Jo

cee @cee.wtf · 13/06/2026

idk it feels very futuristic to have a new AI and then the US government forces it to be rugpulled

013710

Reposted by Sungbin Jo

Samuel @samuel.fm · 11/06/2026

everyone hates on ai features but the Linear<->Slack integration is genuinely top tier, a massive quality of life improvement. just yap w/ collegues and get @linear to put it all into tickets after

5495

Reposted by Sungbin Jo

Gergely Orosz @gergely.pragmaticengineer.com · 11/06/2026

I am so fed up with products I pay for polluting the user interface with AI-related visual clutter Google Docs now has a Gemini button hovering with no way to turn off. I do NOT want it or need it. I pay for Google Workspace, but do not want Gemini, do not need Gemini, and yet cannot turn it off

1412010

Reposted by Sungbin Jo

NullVoxPopuli @nullvoxpopuli.com · 11/06/2026

AI supposedly helps with code out being the bottleneck ~ I assert this was never the bottleneck The bottleneck is making sure the code does what you want, covering all your use cases this involves: ~ review ~ tests ~ etc.. otherwise your users lose trust in you

051

Sungbin Jo @goranmoomin.dev · 11/06/2026

reported as a bug in rust in github.com/rust-lang/ru..., and will be reporting it to llvm as well. seems like the LINKEDIT string pool is aligned to pointer size in both ld64 (the apple linker) and lld-macho. it's only llvm-objcopy that doesn't do that.

github.com

Stripping debuginfo on macOS produces misaligned dylibs when `MACOSX_DEPLOYMENT_TARGET` >= 12.0 · Issue #157750 · rust-lang/rust

Bug When building a dylib / cdylib / proc-macro crate with -C strip=debuginfo and a deployment target that enables chained fixups (i.e. MACOSX_DEPLOYMENT_TARGET >= 12.0, e.g. 27.0), rustc produces ...

020

Sungbin Jo @goranmoomin.dev · 10/06/2026

did any dyld changes that happened in macos 27 golden gate? i was building some rust with MACOSX_DEPLOYMENT_TARGET=27.0, which triggers chained fixups, and seems like when the rustc is stripping debuginfo from proc macro dylibs, it mis-aligns the LINKEDIT string pool, and the dylib fails to load.

111

Reposted by Sungbin Jo

rain 🌦️ @sunshowers.io · 02/06/2026

iddqd, or the hardest kind of unsafe Rust. My first post for the @oxide.computer blog! Goes deep into the pitfalls of unsafe Rust and how to reason about it. A lot of labor went into the post -- please have a read! oxide.computer/blog/iddqd-u...

oxide.computer

iddqd, or the hardest kind of unsafe Rust | Oxide Computer Company

How our Rust collections library defends against adversarial trait implementations.

722339

Sungbin Jo @goranmoomin.dev · 02/06/2026

fr, i came back a few weeks ago and realized that it has become a much more pleasant experience and with a lot more content! (and with a good-enough algorithm with @foryou.club) it’s sticky and i don’t feel guilt in posting to this website so i might actually write posts…

070

Reposted by Sungbin Jo

patak @patak.cat · 02/06/2026

atproto is growing at such a healthy rate lately

Weekly users posting from non-Bluesky PDSes. As of 2026-05-11, there are 12881 users posting from these PDS, with a constant upward trend that started at the start of the year.

31118

Reposted by Sungbin Jo

Nathan Knowler @knowler.dev · 02/06/2026

I'm realizing that I can no longer rely on GitHub for exploring codebases. Way too slow. Faster to clone and explore in an editor.

8482

Reposted by Sungbin Jo

Martin Kleppmann @martin.kleppmann.com · 31/05/2026

You know how IDEs give you a red squiggly underline if you use a variable that hasn't been defined? I want the same thing for writing prose. Red underline if you refer to a non-obvious concept that hasn't previously been defined in the text.

1627413

Reposted by Sungbin Jo

apenwarr @apenwarr.ca · 31/05/2026

Another day, another bunch of entitled people yelling at free software maintainers. The older I get the more impressed I am about how nothing ever changes. apenwarr.ca/log/20211229

apenwarr.ca

The Gift of It's Your Problem Now

Recently a security hole in a certain open source Java library resulted in a worldwide emergency kerfuffle as, say, 40% of the possibly hund...

5949

Reposted by Sungbin Jo

antirez @antirez.bsky.social · 31/05/2026

Many of you forgot too fast the insane amount of shitty software we had to see and suffer in the pre-AI era.

713910

Sungbin Jo @goranmoomin.dev · 31/05/2026

omg this is very, very, very frustrating; i don't get to be the driver when i am programming? i hate the gpt experience.

000

Sungbin Jo @goranmoomin.dev · 31/05/2026

i'm trying out gpt-5.5 for the first time after using only the anthropic family of models, and it's very non-interactive? i always used the claude-based agents very interactively, i.e. with talking to them; it's very frustrating to daily drive the gpt models.

000