Haru

Web3 Developer
CURG 12th @curg_official

[DSRV @dsrvlabs]

🎖️ Special Mentions:

SuiBench: On-chain LLM benchmarking
🔗 https://t.co/D9G8SsEA9C
Sui-treaming: Web3 x NFT streaming service
🔗: https://t.co/iNMWVf0uPe
Aquads: On-chain ad auctions
🔗: https://t.co/8rZNYtyIlA
Sui-mming Pay: Invoice payments for freelancers
🔗: https://t.co/TZX4ZoeOvb

[Haru @hrkairdrop]

✨ Huge UI upgrade for Ontle! ✨

Check out the brand-new look below and let us know what you think! 👇

🔗 https://t.co/ePnR1anX7b
🌐 https://t.co/QwAIvjtoa4

@Mantle_Official #Mantle #TuringTestHackathon #Web3 #BuildOnMantle

[Haru @hrkairdrop]

Excited to introduce Ontle at The Turing Test Hackathon! 🚀

Ontle is on-chain data into a live, interactive visual map using real-time animations. Check out on link below!👇

🔗 https://t.co/hVMoLomaTQ

@Mantle_Official #Mantle #TuringTestHackathon

[Mantle @Mantle_Official]

Not a bad time to activate every group chat you’ve ever joined.

RWAlly your people, lock in their support, and the top 2 projects by engagement score in the Turing Test Hackathon community vote win $8,500 each.

More information below.

[Haru @hrkairdrop]

Now DeepDS is supporting 60FPS and live charts!
Improving on UI/UX🦾🦾

https://t.co/ag7mEtA2ZT

#Sui #Suioverflow2026 #DeepDS

[Haru @hrkairdrop]

@suidevelopers https://t.co/ag7mEtA2ZT

[Haru @hrkairdrop]

Sui Deepbook Predict on Nintendo 👀👀
#Sui #Suioverflow2026 #DeepDS

[Sui Korea 🇰🇷 @SuiNetworkKR]

Sui를 더 쉽게 시작할 수 있도록, Sui Docs가 한국어로 번역되었습니다!

영어 문서가 부담스러웠던 한국 개발자와 빌더분들도 이제 Sui의 구조와 Move 기반 개발 내용을 조금 더 편하게 확인할 수 있습니다!

현재 Sui 생태계에서는 Sui Overflow와 같은 빌더 중심 프로그램도 함께 진행되고 있습니다. Sui에서 직접 무언가를 만들어보고 싶었거나, 이번 기회에 Sui 개발을 시작해보고 싶은 분들은 한국어 문서를 참고해보시기 바랍니다!

[decipher @DecipherGlobal]

Monad Blitz Seoul 4th가 성황리에 마무리되었습니다. ⚡️

짧은 시간 안에 아이디어를 구체화하고, 실제 프로토타입으로 빠르게 구현해낸 빌더들의 에너지와 실행력이 돋보인 자리였습니다.
현장에서의 뜨거운 참여와 협업 덕분에 더욱 뜻깊은 시간이 되었습니다.

🏆 수상팀 소개

🥇 1등 — Monand
AI 에이전트가 고품질 프롬프트를 큐레이션·추천하고, 프롬프트 제작자는 사용량에 따라 보상을 받는 AI 프롬프트 플랫폼

🥈 2등 — MonadConnect
협업 과정의 상태 전이와 보상 정산을 하나의 온체인 상태 머신으로 연결한 Coordination Engine

🥉 3등 — MonArena
100자 텍스트로 캐릭터를 만들고, AI 판정으로 배틀하는 Monad 기반 PvP 플랫폼

함께해주신 모든 참가자분들, 연사분들, 그리고 행사를 함께 만들어주신 분들께 감사드립니다.
다음 Monad Blitz에서도 더 멋진 빌드와 만남이 이어지길 기대합니다.
@B__Harvest @monad

[Haru @hrkairdrop]

We just took 1st prize at Monad Blitz Seoul 4th with Monand. our powerful agentic image prompt marketplace!

What started as an amazing experience turned into something even bigger. Huge thanks to everyone who supported us and to my incredible teammate @beakerjin 🙌

[Haru @hrkairdrop]

Having an amazing time at Monad Blitz Seoul 4th.
The energy and vibes here are incredible 🔥
Cheer for our Monand team! 🚀
Huge shoutout to my teammate @beakerjin 🙌

Link below (demo only):
https://t.co/4eUAvNC0f0
@DecipherGlobal @B__Harvest @monad

[Haru @hrkairdrop]

@jade_o6 @WalrusProtocol @SuiNetwork This looks awesome. Really impressed by the fair-royalty approach. Keep it up! 💪

[zsystm @zsystmd]

Still cooking CometBFT Analyzer.

We've been focusing on the core use cases that really matter — and one of them is the "AppHash Dashboard".

It visualizes each module’s state hash in real time, so teams can instantly detect and debug AppHash mismatches — one of the most painful issues in Cosmos chains.

Perfect for production monitoring, testnet launches, or consensus forensics.

Code’s brewing at 👉 https://t.co/0Da6b68dBh
If you want early access or fast integration, DM me. ⚡️

#CometBFT #CosmosSDK #Observability #AppHash

[Haru @hrkairdrop]

Excited to share that our team won the Performance & Infrastructure track at Hackmos Croatia 2025 🏆 with CometBFT Analyzer!
Huge thanks to my amazing teammates and the organizers for an unforgettable hackathon in Croatia! 🇭🇷🚀
https://t.co/FsEy2w9X7b

#Hackmos2025 #CometBFT

[thebes @voooooogel]

why does this happen? the model believes there's a seahorse emoji, sure, but why does that make it output a *different* emoji? here's a clue from everyone's favorite underrated interpretability tool, logit lens!

in logit lens, we use the model's lm_head in a weird way. typically, the lm_head is used to turn the residual (the internal state built up over the model layers) into a set of token probabilities after the final layer. but in logit lens, we use the lm_head after *every* layer - showing us what tokens the model would output if that layer were the final layer.

for early layers, this results in hard-to-interpret states. but as we move through the layers, the model iteratively refines the residual first towards concepts useful for continuing the text, and then towards the final prediction.

looking at the image again, at the final layer, we have the model's actual output - ĠðŁ, IJ, ł - aka, an emoji byte prefix followed by the rest of the fish emoji.

(it looks like unicode nonsense because of a tokenization quirk - don't worry about it. if you're curious, ask claude about this line of code: `bytes([byte_decoder[c] for c in 'ĠðŁIJł']).decode('utf-8') == ' 🐠'`)

but look what happens in the middle layers - we don't just get emoji bytes! we get those *concepts*, specifically the concept of a seahorse. for example, on layer 52, we get "sea horse horse". later, in the top-k, we get a mixture of "sea", "horse", and that emoji prefix, "ĠðŁ".

so what is the model thinking about? seahorse + emoji! it's trying to construct a residual representation of a seahorse emoji.

why would it do that? well, let's look at how the lm_head actually works. the lm_head is a huge matrix of residual-sized vectors associated with token ids. when a residual is passed into it, it's going to compare that residual with each token vector, and in coordination with the sampler, select the token id with a vector most similar to the residual. (more technically: it's a linear layer without a bias, so v @ w.T does dot products with each unembedding vector, then log_softmax and argmax/temperature sample.)

so if the model wants to output the word "hello", it needs to construct a residual similar to the vector for the "hello" output token that the lm_head can turn into the hello token id. and if the model wants to output a seahorse emoji, it needs to construct a residual similar to the vector for the seahorse emoji output token(s) - which in theory could be any arbitrary value, but in practice is seahorse + emoji, word2vec style.

the only problem is the seahorse emoji doesn't exist! so when this seahorse + emoji residual hits the lm_head, it does its dot product over all the vectors, and the sampler picks the closest token - a fish emoji.

now, that discretization is valuable information! you can see in Armistice's example that when the token gets emplaced back into the context autoregressively, the model can tell it isn't a seahorse emoji. so it tries again, jiggles the residual around and gets a slightly different emoji, rinse and repeat until it realizes what's going on, gives up, or runs out of output tokens.

but until the model gets the wrong output token from the lm_head, it just doesn't know that there isn't a seahorse emoji in the lm_head. it assumes that seahorse + emoji will produce the token(s) it wants.

------------------

to speculate (even more), i wonder if this a part of the benefit of RL - it gives the models information about their lm_head that's otherwise difficult to get at because it's at the end of the layer stack. (remember that base models are not trained on their own outputs / rollouts - that only happens in RL.)

[Haru @hrkairdrop]

From the very beginning, i started this meaning about consistency. not perfection. I also wanted to leave an impression on those who visited my GitHub profile.

From today i will no longer make unnecessary commits.
I want to change daily consistency into meaningful results. 💪

[Haru @hrkairdrop]

Today marks one year since I started the daily commit challenge.

For a whole year I didn’t miss a single day and kept recording commits on GitHub.
ofc it’s impossible to push high-quality commits every day. Sometimes I just edited a README on mobile to keep the streak alive 😂

[Haru @hrkairdrop]

Recently i got lot of scam repo invites on GitHub.

When those repos go private or accounts get disabled, you’re left with ghost unread notifications you can’t open or clear.

I built a tiny cleaner to sweep them away 🧹
https://t.co/g8ZezX2fHa

#GitHub #devtools #OpenSource

[Haru @hrkairdrop]

Thrilled that my project Aquads won a Special Award at the Sui-mming Hackathon! 🏆🌊
Big shoutout to @SuiNetwork and @dsrvlabs for hosting such an amazing event.
This is just the beginning — excited to keep building with the community!

[Haru @hrkairdrop]

@web3BlueNode @nuWindovv 류미 (Ryumi, 류=龍 + 미=아름다움) 어떤가요

[Haru @hrkairdrop]

@saharaai @sskyss18 LFG 🕺