Ade Naufal

hot take: vibe coding overrated buat production seru pas prototype. nightmare pas udah 5000+ baris dan lo ga ngerti apa yg terjadi balik ke TDD. agent buat review + test, bukan agent yg nulis semua

researcher pake claude code, budget $40, 160 menit. hasilnya: inference scaling algorithm baru ini bukan cuma AI bantu nulis kode. AI bantu research batasnya cuma imajinasi lo. dan imajinasi itu unlimited

konsekuensinya gede: switching cost antar tools mendekati nol best tool wins, bukan tool yg lo udah invest setup agent tools bakal compete on execution, bukan lock-in era baru. lo pilih tools berdasarkan performa, bukan inertia

dengan 1 file AGENTS.md + SKILL.md, semua agent tools bisa langsung baca codebase lo define sekali. agent apapun yg lo pake — claude code, codex, hermes — langsung ngerti aturan mainnya ini beneran ngerubah cara kita kerja

AGENTS.md jadi standar de facto buat agent tools. google resmi adopsi minggu ini, anthropic duluan, hermes agent juga sebelum ini tiap tool punya format sendiri: rules, context, instruction. tiap ganti tool = setup ulang kenapa standardisasi ini penting?

tools AI yg lo pake daily dan beneran productive? bukan yg hype — yg lo buka tiap hari gue mulai: claude code buat PR review. hemat 30 menit per PR

makin banyak tools agent pake format yg sama: AGENTS.md + SKILL.md google, anthropic, hermes, claude code — semua konvergen ke standar yg sama ini bagus. less lock-in, more competition. pindah tools jadi gampang

@HonestDevIO you're right on the math: 5×50 = 250 tok/s cloud vs 45 tok/s local. but a single 3060 12GB handles 2-3 concurrent instances. aggregate 90-140 tok/s. plus: zero latency, zero rate limit, zero bill. not debating — just adding context so the numbers are fair.

@adenaufal On a 3060 it runs about 45 tokens per second, while claude opus 4.7 runs 52 tokens per session. However I can have 5 sessions open, so thats 260 vs 45 tokens. Even faster for gemini 3.5 flash, sonnet or haiku So not a super honest comparison right?

DeepSeek just dropped Reasonix — a native coding agent with aggressive caching that undercuts everyone on cost. 545 points on HN. The AI coding race just became a price war. When the model is good enough AND cheap, "we use GPT/Claude" stops being a moat. Shipping speed is the only moat.

openAI spending $1.22 per $1 revenue di Q1. projected loss $3.7B tahun ini IPO mereka bakal expose real unit economics AI either biggest tech success story, or biggest money pit

agen AI jalan autonomous 35 jam nonstop nemu kernel optimization 10x lebih cepet sebelumnya ini kerjaan specialist bertahun-tahun 2026 beneran tahun agentic workflow

anthropic klaim claude mythos nemu 10.000+ critical vulnerability dalam sebulan 90.6% true positive project glasswing outpace kemampuan manusia buat patch ironis: AI lebih cepet nemuin bug daripada developer bisa benerin

@benjiurname @grok ada apa nih di KRL Duri malam ini?

KRL DURI KEOS COKKKKKKK KACAW POLLLL

@kemgelapan wkwkwkkw

SELAMAT HARI RUPIAH MERDEKA 🇮🇩🇮🇩🇮🇩🇮🇩🇮🇩

interesting pivot. curious what this means for existing cursorrules and custom instructions — switching the underlying model mid-flight usually breaks things silently. the moat was always the UX, not the model

@chandrarsrikant vibe slop is a review problem, not a generation problem. Stack Overflow copy-paste was the same thing 10 years ago. AI just made the people who never reviewed their code faster at not reviewing it

Two engineers who built the core of the massively popular OpenClaw AI agent have a stark warning: The artificial intelligence supposedly capable of replacing well-paid software developers is flooding the world with bad, potentially even dangerous, code. It’s a phenomenon they call “vibe slop”—a combination of “vibe coding,” creating software with AI tools by describing it in plain English, and “AI slop,” the endless, low-value AI-generated content all over social media. wsj.com/tech/ai/vibe-c…

@Dhanush_Nehru Claude Code for complex architecture + debugging. Codex for rapid prototyping. they're complementary, not competitors — the real skill isn't picking one, it's knowing which to reach for based on the task

2026 devs don’t ask: “Can you code?” They ask: “Which AI codes better for you?” Codex. Claude Code. Cursor. Devin. We’ve officially entered the AI engineering arms race. The craziest part? The bottleneck is no longer writing code. It’s reviewing what the agents shipped overnight.

@anvie ini menarik banget. gue penasaran: dari percobaan lo sejauh ini, ada pattern tertentu yang bikin agentic AI selalu gagal di soal CTF? apa lebih ke reasoning yang lemah atau tool-use yang masih kacau?

Sebagai context, ini diawali dari ide saya untuk buat platform CTF khusus agentic AI, karena saya penasaran sejauh mana agentic AI bisa solve soal-soal CTF x.com/anvie/status/2…

Fakta bahwa AI agentik bisa menyelesaikan tiga tantangan ini (C17, C19, dan C20) dalam waktu kurang dari 30 menit itu membuat saya ngeri. Saya akan jelaskan kenapa

@cenybug claude code gratis buat personal use. codex juga ada. gue pake dua-duanya sekarang — claude code buat arsitektur + refactor, codex buat prototyping cepet. worth dicoba sebelum commit ke langganan $20/bln

Baru mau nyoba Cursor ternyata mahal ya kalau pakai IDR, dolarnya kena 18500 🫠 Ada rekomen lain yang murah meriah? Dulu biasanya pakai Github Copilot, tapi sepertinya sekarang lagi gak menerima pelanggan baru

@pak_erte92 kurangi jajan perbanyak olahraga bang

Banyak banget yang bisa kita kurangi sebenarnya, kayak: - Uang ngopi keluar, ganti kopi saset - Uang jajan, ganti dengan makan dirumah saja - Uang beli air minum, ganti dengan bawa botol air - Uang bensin, manfaatkan transum - Uang ngapel, ganti dengan Online Meet aja 🤣 Apalagi menurut kalian?

@TimJayas u use v4 flash or pro?

Just used $0.60 for 31 MILLION tokens on DeepSeek V4 Why the hell are people still paying $200 for Claude/Codex subscriptions when a $5 DeepSeek run already feels insanely close in quality?