@0interestrates If you’re not creating agents that spawn meta vibe inferencing agents for every single possible timeline, calculating every outcome in advance and ensuring you become rich, you’re falling behind.
난 게임을 즐겨하지 않는데 이런건 진짜 유익함
만원으로 데이터 센터의 복잡한 구조와 컴퓨터 인프라를 이해하는 스팀게임 : Data Center
빈 방에서 시작해서
랙 구매 → 서버 장착 → 모든 케이블을 직접 손으로 하나하나 연결해야함
실제 데이터 센터처럼 고객 트래픽을 처리하는 시뮬레이션 게임
출시 48시간 만에 180개가 넘는 리뷰가 달렸고, 플레이어들은 “최근 본 시뮬레이션 게임 중 가장 몰입감 있다”, “컴퓨팅 인프라를 이해하는 데 최고”라는 평가를 하고 있습니다.
Today we're shipping Nemotron 3 Ultra.
A 550B MoE frontier-intelligence open model built for long-running agents.
It delivers 5x faster inference and lowers the cost of complex agentic tasks by up to 30% versus other open frontier models.
I've been using both Claude Code and Codex. On the surface they look the same. Both start around $20. Both meter by tokens. Both run 5-hour windows with weekly caps. So you'd assume they're the same product. They're not.
A few things I learned:
- Anthropic capped always-on usage and shut third-party clients out of Claude subscriptions. It wants the agent work happening on its own surfaces.
- OpenAI went the other way. It opened Codex up to third-party tools like OpenClaw and Hermes.
The token data shows the gap. On the same task, Claude Code burns 3 to 4 times more tokens than Codex.
The bigger difference is how you manage context. With Claude Code, you keep one thread healthy. You compact before it degrades. You clear when you pivot. With Codex, you plan around the auto-compact and front-load your context files so the agent survives it.
That changes how each one works. Claude thinks first. It lays out the design and the decisions before it writes. Codex just runs. You hand it a task and it gives you the result.
The benchmarks say the same thing. Neither wins outright. Opus 4.8 leads on real-world issue resolution. GPT-5.5 leads on terminal work. On some tests the winner flips depending on which harness you trust.
So asking which one is better misses the point; they reward different habits. Claude makes me think harder before I start. Codex makes me build cleaner so it can run without me.
#ClaudeCode #Codex #AgenticAI #SoftwareEngineering
Andrej Karpathy: the agent era already killed half of what you're learning
half of what you're grinding right now is already dead weight
senior engineers quietly stopped doing it themselves
the dead list: chasing every new framework, 40 research tabs open till 2am, prompt-and-pray workflows, "one genius model will save me", doing every task by hand
the pattern is obvious. effort that doesn't compound. busywork dressed up as skill. tabs that go viral in your head and close by spring
what actually compounds:
autoresearch - the loop that designs and runs experiments for you
model speciation - not one genius ai, a team of specialized ones
collaboration surfaces between you and your agents
orchestration over execution. you direct, you don't grind
the harness mindset. harness > model, always
taste and judgment - the part no agent replaces
the edge isn't being smarter. it's refusing to stay your own bottleneck while everyone else still grinds solo
book and study this
https://t.co/XDP4SIKwxD
Created a schema oriented MCP for ConnectWise Manage. Works incredibly, pretty sure I converged on the same design Kaseya will use. More information in the repo.
Excellent piece from the Anthropic engineering team on designing tools for agents: https://t.co/5lVYzDgnGF
A few ideas from the article that have shaped how I've been building:
- Designing tools for agents rather than developers
- Keeping the tool surface minimal so the agent can actually reason about its choices
- Returning informative context alongside results to avoid unnecessary round-trips
Solving problems for agents has been a genuinely different kind of engineering than anything I've done before. Worth a read if you're working on anything agentic
Things GitHub READMEs do that I keep forgetting:
- animated SVGs
- mermaid diagrams
- LaTeX math
- collapsible <details>
- direct video upload
Now I want to redo every repo.