Co-founder of https://xitter.cf/t.co/o6cVhRuz6N, VC in Asia with 100+ portfolios. Seeded 1B$ value last 10 years. Investing in Deeptech startups for sustainability & ESG

Seoul, Singapore
Joined April 2009
이건 어떤 사용자가 필요로 하는거야? @grok
Meet Atomic Agent Desktop! ✦ Local AI-first agent for Mac, Windows & Linux ✦ 4x bigger context window on local models with TurboQuant ✦ Atomic Fusion orchestrates cloud and local models to reduce costs Download Free & Private Agent → atomicagent.io
2
1
137
이 기술로 뭘 만들수 있을까? grok
Neuphonic just released NeuDecide, an audio-first tiny decision model at only 43 MB, running on a single CPU thread!! > Speech directly to tool calls (no transcription needed) > Just 46 ms inference on an M3 MacBook > Runs fully offline, even on mobile & Raspberry Pi > 75.5% tool accuracy on SLURP
1
119
조만간 인간만 가입할 수 있는 틴더를 만들어야하지 않을까? 문제는 인간인지 아닌지를 누가 평가할 수 있을까? 안드로이드는 양의 꿈을 꾸는가? youtube.com/watch?v=lHw6yoyP…
163
Michael ByungSun Hwang retweeted
Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks. We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed. Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. 📄 Blog: contrastive-lm.notion.site 💻 Code: github.com/Contrastive-LM/CL… 🗣️ Discord: discord.gg/5dAQEDJBs 🤗 Data & Models: huggingface.co/Contrastive-L… More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵
190
641
171
5,543
1,062,878
엄청난 혁신 경쟁이 진행되고 있군요. 보안카메라에도 내장할 수 있을거 같군요.
Introducing Jev-Omni, the first multimodal system one model. (Other OSS versions miss atleast a modality) Supports all modalities: text, images, audio and video ! On Par with Jev on Typed-benchmarks. Scaled -> 30k examples on 8xH200 (data mix matters a lot) < 100ms on 1 H100 huggingface.co/akhilaaa3/Jev… More work is coming, so follow along !
1
254
Michael ByungSun Hwang retweeted
最近Jev爆火,拿到 Jev API 还不知道怎么玩的,可以按这个清单抄。 1. jev-ultrafast Browser Use 做的高速浏览器 Agent。Jev 每一步只判断「做什么、点哪个元素」,要打字才叫小模型。Google Flights 搜一次航班大约 7 秒。 github.com/browser-use/jev-u… 2. fast-jev-compaction 给 Claude Code 做上下文压缩。每次工具调用先让 Jev 判断还有没有用,没用的删掉,留下来的原文不重写。 github.com/tamaratran/fast-j… 3. json-render Vercel Labs 的生成式 UI 框架。实验里 Jev 不逐 token 写 JSON,只负责选组件、属性和布局。 github.com/vercel-labs/json-… 4. typesafe-mcp 最适合刚拿到 API 的人。把 Jev 接进 Claude Code、Claude Desktop、Codex 和 Pi,随时做 Choice / Score / Noul。 github.com/itsmostafa/typesa… 5. jev-mcp 现成的 Agent 判断工具箱:事实核验、内容筛查、语义排序、分类和信息提取。 github.com/jkudish/jev-mcp 6. SemDecide 把 Jev 做成命令行。Shell 里直接分类、打分、过滤,适合接爬虫、CI 和数据流水线。 github.com/sharziki/semdecid… 7. jev-codex-router 先让 Jev 判断这一轮编程任务有多难,再决定模型档位、推理深度和速度模式。 github.com/0xNatoshi/jev-cod… 8. Winnow 给 Claude Code 做上下文垃圾回收。Read / Bash / Grep 吐一大堆时,Jev 先判断哪些真和当前任务有关。 github.com/GhalebDweikat/win… 9. jev-review 代码审查前先过一遍 Jev,把高风险改动挑出来,再交给更贵的大模型或人。带本地看板。 github.com/devagrawal09/jev-… 10. Blink 把 Jev 当代码库导航器。每走到一层目录,就判断哪些文件和当前问题最相关,再继续往下找。 github.com/ellipsis-dev/blin… 11. agent-desktop 桌面自动化。读系统无障碍树,判断下一步该点哪个按钮、菜单或输入框。 github.com/lahfir/agent-desk… 12. typesafe-mario 让 Jev 玩《超级马里奥》。不看截图,直接读模拟器 RAM 里的结构化状态,再决定跑、跳、躲。 github.com/fhshaik/typesafe-… 13. jev-drone 拿 Jev 控无人机。底层飞控继续负责稳定和安全,Jev 只做爬升、刹车、穿越障碍这类上层判断。 github.com/RomanSlack/jev-dr… 14. OneVOneJev 浏览器里的 1v1 FPS。每个决策 tick 都要判断走位、视角、瞄准、开火和跳跃。 github.com/emrickgarrett/One… 15. jev-trader Monad 测试网上做高频做市。Jev 根据价差和成交方向判断下一步买还是卖,模型延迟大约 81ms。 github.com/jarrodwatts/jev-t… 16. Prism 不直接让 Jev 下单。它判断 toxic flow、市场压力、均值回归之类的状态,再交给原来的策略。 github.com/irfndi/prism-liqu… 17. neo4jev 把 Jev 塞进知识图谱。每走到一个节点,判断下一条最值得走的边,再一路找下去。 github.com/jexp/neo4jev 18. jev-curate 拿 Jev 筛训练数据。JSONL / Parquet 先做质量、相关性和风险判断,再决定哪些进后面的训练。 github.com/AkashPriyadarshii… 19. Canny 防 Coding Agent 嘴硬说自己做完了。看工具输出、代码 diff 和测试结果,再判断完成声明靠不靠谱。 github.com/qkal/Canny 20. killmyidea 输入一个创业点子,Jev 从多个维度打分,最后给你 KILL、FIX 或 SHIP。 github.com/monteduro/killmyi…
280
704
44
4,430
405,576
Michael ByungSun Hwang retweeted
Jev のクローン、2日で6本。本体が非公開なので全員が「中身の推測」で作っていて、6本ともアーキテクチャが違うのが面白い...。 ❶ Laya — ModernBERT系エンコーダ421M + PPO訓練 ❷ DiffusionGemmaJev — 拡散モデルで再現。ベンチは肉薄 ❸ Bespoke Nimble — Qwen3.5-9BのLoRA。自前評価90%(本家93%) ❹ SemIf(旧OpenJev) — Qwen3.5 + 3クラスNLI分類器。4Bと35B ❺ Jevlike — 埋め込み40KBの軽量アテンション ❻ Kev-0.5B — 0.5BのLoRA。MacBook Proで動く Latent Space の見立てでは、有力なのは ModernBERT 説と拡散モデル説か latent.space/p/ainews-here-a…
13
131
20
762
58,214
이젠 디지탈헬스케어에도 적용
During a live clinical consultation (with an ambient scribe), the transcript is periodically supplied to Jev, which traverses a medical ontology to classify symptoms, update the differential diagnosis, suggest treatments and identify red flags in real time.
246
Jev가 정말 쓸만한 LLM일지 궁금하네요.
Just found Jev. Jev looks at the page, sees the real buttons, and just picks: click this one, type here, scroll, wait, or done. - Click this - Scroll down - Type this - Buy / sell - Retry or stop - Send the job to another agent That’s it. This is useful for: • Browser agents that need to be fast • Auto QA • Trading loops • Agent routing • Games / robots • Anything where the AI has to choose an action over and over. The real trick is splitting thinking from “what do I do next”. Big model does the hard thinking. And the small fast model just picks the next click. Still looking into it, so forgive me for any false claims. Open source here: github.com/browser-use/jev-u…
1
1
292
AI를 위한 보험이라.
Today we’re launching Keythorn. AI isn’t just thinking anymore. It’s moving money, talking to customers, and taking action. Every new agent brings new risk. The future runs on AI. We’re building the insurance for it. Meet @TryKeythorn ↓
1
284
Michael ByungSun Hwang retweeted
Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead. You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier intelligence. You’ve also claimed the lead is widening because of recursive self-improvement. I don’t see what you see in the lab. If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible. But stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff. Stop pretending you need those same evaluators to police competitors who aren’t even at the frontier. Most of all, stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure if your products enable a truly damaging cyberattack. The market already punishes models that behave in unpredictable or unauthorized ways. After the Hugging Face episode, it is simply good business for OpenAI and Anthropic to trade some raw power for reliability and predictability. Call it alignment if you want. It is also just giving customers what they want. Pacing the frontier would also create breathing room for a more intelligent conversation about regulation than Bernie Sanders’ “shut it all down.” China is very unlikely to join a global agreement, as you know, and that has to be taken into account as well. So go ahead and pace the frontier. You are the ones setting it. The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it. If you do, you’ll buy goodwill for the next conversation. If you don’t, we’ll know this was just another bid for regulatory capture — or an election-season psyop.
3,409
12,296
2,707
71,988
9,626,826
하네스는 껍데기가 아니다. @grok 설명해봐
Harnesses often get dismissed as just scaffolding, just prompt engineering, and not real research. But that couldn't be farther from the truth. The same model weights that score 30% on ARC-AGI score 95% with a better harness. So we gathered a group of researchers and founders working at the frontier to do a deep dive into the state of harnesses. We cover how we got to this point, the case for making your harness as expressive as possible, and what YC learned building an agent for every employee in the company. 00:00 - @FrancoisChauba1: Why harnesses matter 04:27 - Building an auto-researcher by accident 07:13 - A five minute history of harnesses 13:56 - Self-improving harnesses 18:35 - @sethkarten: Prime Agent, a self-improving RLM harness 21:50 - Context as an L1, L2, L3 cache 24:51 - From Turing machine to von Neumann computer 28:33 - Messaging between agents 30:04 - ARC-AGI results 33:09 - Emulator Bench and GPU kernels 37:30 - @JonSaadFalcon: OpenJarvis, personal AI on personal devices 38:26 - How far behind are local models 39:21 - The five primitives of a personal AI stack 42:47 - Letting cloud models optimize your local stack 43:53 - 800x cheaper than the cloud 45:58 - @josh__france and @jbellregan: QM, YC's agent harness for work 47:29 - A history of YC's internal agents 49:24 - OpenClaw and a fleet of 50 agents 51:04 - Pulling the brain out of the sandbox 54:43 - Letting the agent choose its own sandbox and model 57:16 - The grind tool: budgets on goals 58:50 - Agents don't understand social context
1
1
314
이게 핵심이구만. 컨설턴트는 공짜로 하고 토큰 비용으로 받겠다?
Replying to @jinshang1997
amazing business model they got; they don't charge customers for the FDE, but for the services consumed (markup on tokens).
1
1
253
아직 이걸 배워야하는걸 보니 갈길이 멀구나.
Your Semantic Layer is not a Context Layer. And neither of them is an Ontology. If you are building data workflows for autonomous AI agents, it helps to know the difference. -Semantic Layer: The rigid logic compiling to SQL (metrics, dimensions). -Context Layer: The unstructured human knowledge (docs, wikis) agents desperately need to make decisions. -Ontology: The "digital twin" mapping your real-world business entities. Dive into Simon Späti's new primer to see how to organize your data architecture for the Agentic Era: motherduck.com/blog/context-…
224
Michael ByungSun Hwang retweeted
Replying to @marsnine
Anthropic의 Model Hardware Standard(MHS)는 AI 에이전트가 현미경·로봇팔 같은 실제 장비를 안전하게 조작할 수 있게 하는 공통 표준이야. 기존엔 장비마다 맞춤 연동 프로그램을 수주~수개월 걸려 만들었는데, MHS로 몇 시간 만에 연결되고 AI가 여러 장비를 동시에 제어·자율 실험까지 가능해. 기존 하드웨어 회사 입장에선 자사 제품에 MHS 드라이버만 지원하면 AI 자동화 시장에 바로 진입할 수 있는 기회야. 이미 Universal Robots·Doosan 등이 참여 중이고, 오픈소스화되면 사실상 업계 표준이 될 가능성이 높아.
1
84
Michael ByungSun Hwang retweeted
The next $1T company will be an AI-native services firm. But it must begin in a vertical with these traits: 1) Document-driven 2) Rules-bound 3) Verifiable Document-driven means the work product is a document. A contract. A claim. A filing. A medical record. AI reads, generates, and checks documents. The feedback loop is instant. Rules-bound means the rules are written down. Regulations, billing codes, compliance standards. Not "use your judgment." When rules are codified, AI doesn't need 20 years of instinct. It needs the rulebook. Verifiable means you can check the output against ground truth. Did the claim match the policy? Did the filing include the required disclosures? Verification kills hallucination risk at scale. The model labs got the trillion-dollar head start. The services firm that picks the right vertical gets the next one.
32
25
3
390
63,927