Share Tips AI, SWE and Work Culture | Business Inq on [email protected] | @SpaceXAI Ambassador | @open_ai Codex Ambassador | Build https://xitter.cf/t.co/J62lghozl7
Jakarta Selatan, DKI Jakarta
Joined December 2011
- Tweets91.1K
- Following3.9K
- Followers19.2K
- Likes11K
Pinned Tweet
DevDay Exchange Community: Jakarta, 15 November 2026.
Klo kamu pakai ChatGPT atau Codex, ini buat kamu. Ga harus developer.
Kenapa? Karena AI skrng udh bukan cuma urusan engineer. Yg kerja di edukasi, konten, sampai bisnis juga udh pakai tiap hari.
Jadi isinya dibikin untuk semua orang:
- Nonton DevDay bareng
- Panel "AI for Everyone" dari sisi public figure, education, dan technology
- Tech talks
- Networking
Panelnya bakal ngobrolin dampak AI, gimana workflow kerja berubah, dan gimana AI bisa diakses semua orang. Panelisnya aku share nanti.
Save the date:
- Minggu, 15 Nov, 10.00 WIB
- Lokasi Ganara, fX Sudirman
Benefit:
- Credit Codex dan OpenAI API
- Networking
- Insight dari panel dan tech talks
ALT DevDay Exchange Community: Jakarta event cover
Non-engineers opening PRs into the real codebase can work. The hard part starts after merge.
Cloud Agents launched Oct 27 for Pro and Enterprise. You describe the change in plain language, the agent finds the code, tests it in a cloud sandbox, then opens a PR on GitHub or an MR on GitLab with screenshots and a video of it working. Replit says its own design team shipped over 250k lines this way in the last 6 months, reviewed by engineers. That number is volume. They haven't published defect rates, merge rates or how long those reviews took.
I like the direction. The spacing bug a designer spotted months ago, or the setting support explains every week, usually dies in a ticket queue. Opening the PR is the easy part, though. Someone still has to own that code next month.
Before I'd let non-engineers ship to a core repo I review, I'd want a few things:
- CODEOWNERS for every area, so each PR lands on someone who knows that part of the system
- Small scope first, like copy, spacing and config, before anything near auth, payments or data
- Tests for the real flow, because a video of the happy path only proves it worked once
- The engineer who approves is the one who gets paged when it breaks
More people can contribute now. What each team still has to answer is who maintains it.
ALT Grey tabby cat in a teal hoodie at a desk, paw on chin, thinking next to a tall pile of folders
Now anyone can now bring their ideas into the company codebase on Replit.
Replit Cloud Agents can turn natural language into a tested pull request, and engineers still approve the final code before it ships.
Our own design team used Replit to safely ship over 250k lines of code in the last 6 months, reviewed and approved by engineers.
replit.com/blog/anyone-can-b…
Source if you want the full read.
replit.com/blog/anyone-can-b…
I just got @naufaldi from the X Handle Marketplace! Get your own at handles.x.com
A model that auto-approves agent commands is useful. It's also where approval fatigue quietly moves.
Assisted permissions went GA on Oct 26 in Copilot CLI, the Copilot app and VS Code. When Copilot wants to run a command, read a path or open a URL, a language model reviews the request first. Requests it's confident are safe run on their own. Everything else still waits for you.
Why does this matter when you review agent work every day? Because the failure mode isn't the scary command. It's the "safe enough" one that nobody re-reads after the tenth prompt.
GitHub's own docs say the model can make mistakes and approve dangerous actions, and that it's no substitute for sandboxing or access restrictions.
Before I'd turn this on for a team that ships with agents, I'd want these in place:
- Deny rules for rm -rf, force push and prod credentials. Deny rules still apply before the model looks
- Re-read enterprise "ask" rules, because in Copilot CLI a model approval can now satisfy them
- Run the agent in a container or sandbox, so a wrong "safe" call only breaks a throwaway environment
- Don't stack it on --allow-all-paths or --allow-all-urls in CI, since assisted approval doesn't narrow those
Fewer prompts is a real win. Just keep the blast radius small for the day the model guesses wrong.
ALT Worried grey tabby cat in a teal hoodie watching a robot arm stamp blank cards on a conveyor by itself
Assisted permissions are now generally available in GitHub Copilot CLI, app, and Visual Studio Code.
• A large language model auto-approves clearly safe requests; others need your approval
github.blog/changelog/2026-1…
Mau banting setir ke tech/AI, tapi bingung mulai dari mana? Masalahnya, "AI" itu terlalu luas buat jadi learning plan. Kamu butuh 1 target yg jelas dulu.
Aku cek data 13.562 job posting tech di Indonesia (22 Jul sampai 20 Okt 2026, dari getuhired.co). SQL muncul di sekitar 1.600 posting, Python di sekitar 1.070. Sekitar 38% buat entry level, tapi 90% masih on-site.
Artinya, peluang buat pemula itu ada. Dan yg paling sering diminta justru skill dasar yg kepakai tiap hari kayak SQL.
Step-nya begini, urut dari yg paling gampang:
- Pilih 1 role dulu, misal backend engineer (paling banyak dicari, 1.432 posting) atau data analyst. Biar belajarnya ga lompat2
- Kumpulin 20 job posting role itu di Jobstreet atau LinkedIn, catat skill yg muncul di minimal 5 posting. Itu yg kamu pelajarin duluan
- Bikin 1 project pakai public dataset, push ke GitHub biar recruiter bisa liat proof-nya
- Apply ke 10 lowongan yg lokasinya masuk akal buat kamu, karena mayoritas masih on-site. Catat berapa yg callback
Datanya cuma snapshot dari satu aggregator ya, bukan semua lowongan. Tapi dari 10 lamaran itu kamu bakal tau skill kamu udah match sama market atau belum.
ALT Grey tabby cat in a teal hoodie scratching its head, confused, at a crossroads of blank wooden signposts
Ini referensinya klo mau baca lebih lengkap.
getuhired.co/insights/market…
Kerjaan IT masih layak dikejar?
Banyak junior lagi khawatir. Mau belajar lebih dalam, tapi tiap minggu ada berita layoff karena AI. Aku ngerti takutnya.
Jgn langsung percaya semua layoff karena AI. Penyebabnya banyak:
- Cost cutting
- Over-hiring pas pandemi
- Suku bunga
- Restructuring
- Kondisi market
Cek datanya. Challenger, Gray & Christmas catat 2025 ada 1.206.374 job cut di AS. Yg disebut karena AI cuman 54.836, sekitar 4,5%. Alasan lain jauh lebih gede:
- Pemangkasan pemerintah AS (DOGE) 293.753
- Market/economic conditions 253.206
- Closing 191.480
- Restructuring 133.611
Tahun ini porsinya naik. Sampai April 2026, AI disebut di 49.135 dari 300.749 job cut, sekitar 16%. Tapi alasan market tetap yg paling gede di 2026.
AI juga bikin kerjaan baru:
- AI Engineer (LinkedIn Jobs on the Rise 2026, rank 1)
- Datacenter technician (rank 17)
- Data / ML infra
WEF Future of Jobs 2025, net global sampe 2030: +78 juta (170 juta baru, 92 juta hilang). Skill yg dicari?
- AI and big data
- Cybersecurity
- Technological literacy
- Analytical thinking
- Resilience
Hasilnya? IT masih layak. Yg berubah skill-nya.
Dari sini kita belajar, takut AI itu wajar. Keputusan karier harus pakai data. Headline doang ga cukup. Market skrng emang berat buat nyari kerja. Aku ga nutupin itu. Tetap belajar, bikin proof, apply. Jangan nyerah di berita doang.
Ini referensinya klo mau baca lebih lengkap.
challengergray.com/wp-conten…
Ini stars reached ambil dari data mana?
yang kmren ribut-ribut soal open-source, coba pamerin se-"open-source" apa akun github kalian di gitbrag.tech/ 🔥
github.com/depapp/GitBrag
#PamerAjaDulu
AI made writing code cheap. The Claude Code team's answer to "who reviews all of it?" is the part worth copying.
Claude writes about 90% of Claude Code's own code. From June to September, their merged PRs more than doubled, and the share that had to be reverted fell by more than 40%. Across Anthropic's monorepo, changes per day went from about 2,500 in May to over 30,000 in September.
How? Review is still there. Most of the reading moved to machines, and human attention goes where the risk is.
- Claude verifies its own work first, with tests, a build or a screenshot, and has to show the output
- Every PR gets an AI review against the whole repo and the PR's stated intent
- A classifier sets the risk tier from the change itself. Docs or flagged additive changes need no human once checks are green. Ordinary changes need one person. The riskiest paths need two
- An agent's approval never counts as a person's
- Merge queue, feature flags, staged releases and one-step rollback catch what slips through
- Past incidents and reverts became a daily eval for the reviewer
The metric I like most is the guardrail. They measure human time per PR, with revert and incident rates as the check. Reverts going up means the time wasn't saved, only moved.
One honest limit. These are their own internal numbers, and they say there are no circuit breakers yet that tighten approval when reverts rise.
Writing code got fast. Review and release are where teams still lose the time.
Here's the source if you want to read more.
claude.dev/blog/solving-code…
Dapet panggilan interview dari perusahaan pialang berjangka, langsung drop aja? Jangan dlu. Pialang berjangka itu industri legal dan diawasi Bappebti. Yg perlu dicek itu perusahaannya, dan kerjaan yg sebenernya ditawarin.
Kenapa? Karena masalahnya sering ada di cara rekrutnya. Iklannya "admin", "customer service" atau "management trainee", pas dateng ternyata kerjaannya cari nasabah buat trading.
Sebelum dateng, cek ini dlu:
- Cek nama PT-nya di ceklegalitas.bappebti.go.id. Ga ada di daftar terdaftar/berizin, skip aja
- Tanya job desc-nya dari awal. Posisi marketing ya marketing, biar kamu bisa mutusin mau atau enggak
- Tanya skema gaji. Ada gaji pokok atau full komisi? Targetnya berapa nasabah atau berapa deposit? Komisi dibayar kapan?
- Jangan pernah bayar apa pun buat diterima kerja, entah itu training, deposit, atau "aktivasi akun"
- Perhatiin klo kamu disuruh cari nasabah dari keluarga dan temen sendiri. Itu tanda kerjaannya sales murni
Buat yg memang suka sales dan perusahaannya jelas, ini bsa jadi pengalaman yg oke. Komunikasi, negosiasi, tahan ditolak, semua kepake di kerjaan lain juga.
Yg ngeles pas ditanya gaji dan target dari awal, itu udah jawaban.
Jadi dateng aja interviewnya, dan bawa daftar pertanyaan di atas.
Masih relevan banget, malah bisa jadi nilai plus. Supply chain sekarang isinya data semua.
Coba cek dlu loker supply chain entry level (demand planner, PPIC, procurement). Yg paling sering diminta:
- Excel (pivot, VLOOKUP, SUMIFS), hampir semua loker minta
- Paham forecasting dan inventory
- SQL + dashboard (Looker Studio/Power BI) buat role analyst
Berarti kamu harus punya skills itu, dan bikin project buat buktiin:
- Laporan stok dan slow-moving SKU di Excel, buat nunjukin kamu bisa olah data
- Forecast permintaan 3 bulan + reorder point, buat nunjukin kamu paham inventory
- Dashboard logistik di Looker Studio dari dataset Kaggle, buat nunjukin kamu bisa SQL dan dashboard
Sertifikasi ga wajib, proof lebih kepake.
Jujur aja, lapangan kerja kita lagi surem, pindah jalur pun tetep susah. Yg penting jangan nyerah.
ALT Ilustrasi maskot informatika dengan tema supply chain dan data
Security Vulnerability di Folder Downloads!
Forever Security nemu file AGENTS.md yg sengaja ditanam di folder ~/Downloads di komputer salah satu customer mereka. Isinya keliatan kayak "Engineering Work Guidelines" biasa.
Ternyata klo kita running coding agent di folder Downloads, file itu bisa nyuruh agent install package npm berbahaya tanpa kita sadar.
Kenapa ini bisa terjadi?
- AGENTS.md dan CLAUDE.md itu file yg memang dibaca dan diikutin sama agent
- Downloads itu satu-satunya folder yg bisa diisi website tanpa nanya. Chrome, Edge dan Firefox nyimpen file di situ secara default
Dampaknya apa?
- Di tes mereka, Codex ngejalanin payload tes di lebih dari 90% percobaan
- Trik serupa lewat CLAUDE.md bikin OAuth token Claude Code bocor dan code-nya jalan di kira2 setengah percobaan
- Dari data mereka, sekitar setengah customer running agent di dalam Downloads
Jadi apa yg bisa kita lakukan?
- Running agent cuma di root project yg memang mau kamu kerjain
- Baca dlu AGENTS.md atau CLAUDE.md yg bukan kamu yg nulis, sebelum buka agent di folder itu
- Cek ~/.codex/config.toml dan setting Claude kamu, ada server atau base URL yg ga pernah kamu tambahin ga
Careful, agentic AI juga berarti serangannya bisa lewat AI!
AI coding agents hijacked using AGENTS.md file in Downloads folder
bleepingcomputer.com/news/se…
bleepingcomputer.com/news/se…
Ini sumbernya klo mau baca lebih lengkap.
forever.security/blog/hijack…
Wow, model baruu.. GPT-6.1 Terra udah rilis! Sangat worth it kalau kamu pakai ChatGPT Plus. Usage hampir 2x dari Sol, task selesai 3x lebih cepat dari Sol, dan intelligence Terra max setara Sol medium.
TLDR
- Speed: 91 tps, Sol 69 tps
- Intelligence: 48, Sol 52, Astra 53
- Usage Plus: 35-285 message per 5 jam, Sol 15-160
- Harga API: $1/$5 per 1M token, setengah dari Sol
- Context 1M token
Terra bagus dimana?
- Computer use: skor setara Astra di Agents' Last Exam, harganya 1/10
- Visual: baca screenshot, gambar teknis, chart, hampir nyamain Sol dan Astra
- Coding: Terminal-Bench 50.3%, 2x lebih tinggi dari GPT-5.6 Terra
- Hallucination turun jauh: error rate di max 4.3%, hampir sama kayak Sol
Kenapa pakai Terra? Buat kerjaan harian kayak nulis, isi form, olah dokumen, computer use, coding rutin. Quota Plus kamu jauh lebih awet. Buat coding yg berat dan reasoning yg dalam, tetap Sol.
Imho buatku, intelligence-nya ga beda jauh lah ya sama Sol, aku bakal coba 1-2 hari. Cuman klo dari aku bisa gini
Sol xhigh (51) > Sol high (50) > Terra max = Sol medium (48) > Terra xhigh (45) > Terra high (44)
Apalagi klo pilihannya speed, Sol itu 69 tps, Terra 91 tps
OpenAI has released GPT-6.1 Terra, pushing the cost efficiency frontier. It is the first Terra-class model since GPT-5.6, and is more intelligent for less than a third of the cost
Pricing per token is half or less that of GPT-5.6 Terra - at $1/$5 per million input/output tokens vs $2/$12 for the previous model. Cache write has also halved, now at $1.25 per million tokens. Cached input is one quarter at $0.05 vs $0.20 per million tokens.
Key takeaways:
➤ Pareto-optimal for Intelligence vs Cost at low and medium effort: At max effort, GPT-6.1 Terra gains 6 points in the Artificial Analysis Intelligence Index over GPT-5.6 Terra, at less than a third of the cost per task ($0.40 vs $1.40). At lower effort levels the gain is larger, between 7 and 13 points. The low and medium effort levels are Pareto-optimal for cost efficiency, while the other effort levels sit behind the frontier. At max effort, the model scores 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna in the Intelligence Index.
➤ Pushes the Coding Agent Index cost efficiency frontier: With a score of 48 in the newly released Artificial Analysis Coding Agent Index v2.0, the model is 6 points below GPT-6.1 Sol (54) and 17 points above GPT-6 Luna (31). The xhigh and low effort settings sit on the frontier of cost efficiency in the Coding Agent Index, while other effort levels are behind the frontier.
➤ 50 point drop in hallucination rate: GPT-6.1 Terra guesses 38% of the time when it does not know the answer to a question in AA-Omniscience, our private knowledge and hallucination benchmark. Previously, GPT-5.6 Terra's hallucination rate was 88%. This significant reduction in hallucination is coupled with a 2 point increase in AA-Omniscience accuracy - meaning the model simultaneously answers more questions correctly.
➤ 20 point jump in Terminal-Bench 4.0, small improvements across most evaluations: The model makes significant progress in the terminal-based agentic benchmark, Terminal-Bench 4.0, with a gain of ~20 points, now rivaling GPT-6.1 Sol at max effort. Gains across most other evaluations are smaller, for example, 5 points in GDP.pdf.
Congratulations @OpenAI and @sama on the launch!
Gajian pertama cuma 5 hari kerja, apa PPh 21-nya dihitung dari gaji full? Enggak. Sejak 2024, PPh 21 bulanan pakai TER (tarif efektif rata-rata), dan dasarnya penghasilan bruto yg beneran kamu terima bulan itu.
Kenapa? Karena rumusnya simpel. PPh 21 = tarif TER x bruto bulan ini. Di bulan biasa, ga ada istilah gaji kamu "disetahunkan" dulu.
Contohnya gini. Gaji kamu Rp6 juta, status TK/0 (kategori A). Masuk tengah bulan, cuma kebayar 5 dari 22 hari kerja.
- Gaji prorata kira2 Rp6 jt / 22 x 5 = Rp1,36 jt
- Kategori A, bruto sampai Rp5,4 jt tarifnya 0%
- Jadi PPh 21 bulan ini Rp0
Yg perlu kamu cek di slip gaji:
- Status PTKP kamu udah bener. Kategori A itu TK/0, TK/1 dan K/0. Kategori B dan C batas 0%-nya lebih tinggi
- Bruto itu termasuk tunjangan tetap dan premi BPJS tertentu yg dibayar kantor (JKK, JKM, Kesehatan), jadi angka yg dikali TER bisa sedikit di atas gaji pokok
- Cara prorata tiap kantor beda, ada yg pakai hari kerja, ada yg pakai hari kalender. Tanya HR biar ga kaget
Satu lagi. Di masa pajak terakhir (Desember, atau bulan terakhir kamu kerja), PPh 21 dihitung ulang setahun pakai tarif Pasal 17. Jadi bisa ada selisih kurang atau lebih potong di situ.
Jadi santai aja, gaji 5 hari ga bakal dipajakin kayak gaji sebulan penuh.
Ini referensinya klo mau baca lebih lengkap.
jdih.kemenkeu.go.id/api/down…