Coretax as SaaS is the moat
Ya, SaaS resmi mati, dan pasti akan mati!!!!!
Menurutku hanya beberapa yang bisa bertahan:
- Berhasil sebanyak mungkin data locks sebelum negara ai menyerang.
- Punya MOAT yang 100x.
- Perusahaan AI belum berhasil menggantikannya.
- Punya jaringan member yang super kuat. Misalnya, gojek grab akan bisa bertahan karena mempunya jutaan driver dan puluhan juta pelanggan. Perusahaan AI sulit menggantikannya.
Ga punya salah satu atau salah dua dari itu, nunggu waktu kukut.
Ditto
Muse assistant is truly amazing. Somehow it’s the first moment I feel the AGI.
It shows that you don’t need the best LLM, just a really well thought out product.
I love ChatGPT/Claude/Gemini and those are the ones I would use to design a house, create complex software, etc… but Muse is directly useful and just makes the day to day easier and more fun.
Boring admin tasks are suggested, it’s not overwhelming to switch insurance when it notices you’re being screwed, give you a little script here to help call for the cancellation.
I assume it’s also heavily tuned for consumer tasks, chatting on many interleaved topics at the same time, updated daily with the relevant knowledge (how airlines will consider your infant seat, what are the good credit cards, etc), instead of researching it all the time.
This is just my assumption and if it’s not how it works, I’m sure that’s what they’ll do, it just makes sense for efficiently serving people.
And it actually feels like a real personal assistant tuned for me with my interests, etc. Not nudging me all the time for interaction, just proposing few actions a day.
It can also easily be tweaked, all by LLM, even just for checking a box from its TODO list (which makes sense because it may impact some other internal memory).
All the reminders, goals, etc are visible and well organized.
It’s not the smartest, but it’s more than smart enough for all the tasks required. And when you don’t agree or want to make more research, you can just orient it, or do your side search or your other leading LLM and feed back that response to it.
On that note we need a way for agents / different providers to communicate with each other, sharing memory, reaching out to a session, a bit like Claude/Codex are able to do within their own sessions. I see myself using all of them (ChatGPT/Claude/Gemini/Muse/Grok) for different tasks and the main friction now is sharing the knowledge between each other, where I end up trying to design my own system but I probably should not and a leading company will end up establishing a normalized way (a bit like Anthropic did with MCP’s and skills).
In any way congrats Meta for designing such a great product. From this step, it’s a straightforward shot for continuously improving it!
For the others you should take inspiration from it, though I understand if their own future agents (ChatGPT Dot + whatever the others do) evolve differently as they target different use cases (business, coders, etc) and I don’t think you can build an agent that’s just the best interface for everything.
Instead each person will connect the agent of their choice for different tasks and they will need to be efficiently interlinked (Muse for personal stuff / consuming, Claude/Codex for coding, ChatGPT for finances/admin, Gemini for deep research/knowledge, local models for private/confidential stuff, etc).
Presiden aja disadap.
Meanwhile, my homies:
“Kita nggak disadap, bro. Kita sukarela spill seluruh hidup ke ChatGPT.” 😭
AI sekarang bisa ngedit vlog dari footage mentah, tinggal dikasih foldernya
kami tetep bisa 40 menit galau milih thumbnail. the bottleneck was never the tool
jadwal shift karyawan tiap minggu bikin pusing? kasih AI 3 hal. jam buka, siapa ga bisa hari apa, minimal berapa orang per shift
terus minta dia tandain yg BENTROK, jangan dibenerin diem diem. yg mutusin tetep lo
limit AI lo abis sebelum makan siang? ini 4 kebiasaan yg bikin jatahnya awet
satu tugas satu chat baru, kerjaan ringan pake model ringan, file dikirim sekali aja, minta jawaban singkat dulu
tanda lo kebanyakan pake AI di kantor
- chat ke pasangan pake bullet point
- bilang "coba regenerate" pas rapat
- nanya ke temen "boleh kasih konteks lebih?"
- bales grup keluarga "noted, ini versi ringkasnya"
Jangan jangan lagi disadap
rekap pengeluaran sebulan dari foto struk
1. kirim 10 foto struk sekaligus
2. minta tabel tanggal, toko, total, kategori
3. yg buram tulis CEK, jangan ditebak
4. minta total per kategori
5. cocokin ke mutasi rekening
6. tanya 3 pos yg bisa dipangkas
AI yg ngetik, lo yg ngecek
istilah baru dari dunia engineer, cognitive surrender
alias pencet approve tanpa baca. di kantor biasa disebut "udah acc aja, yg penting jalan"
akun AI sharing 25rb sebulan emang murah. yg ikut ke-share bukan cuma login
riwayat chat, file yg lo upload, sama memori soal kerjaan lo. data customer jangan ditaruh di akun barengan
minta AI "tulis ulang kayak manusia" itu kayak bilang ke tukang cukur "yg bagus ya"
kami ganti jadi 3 hal. tempel 1 contoh tulisan lo sendiri, sebut kata yg dilarang, kasih batas panjang. hasilnya kedengeran kayak lo, bukan kayak brosur
Artificial Analysis sekarang ngukur kecepatan AI pake waktu beres per tugas, bukan token per detik
GPT-6 Astra mode max sekitar 18 menit, mode default ~6 menit. Opus 5.5 26 menit vs ~7
buat bales email sama rekap harian, mode default udah cukup. max simpen buat yg beneran ribet














