Abid Famasya

@evanpurnama Wah menarik, udah lama ngga hands-on di bidang ini. Terakhir waktu ngulik transformers sebelum chatgpt lahir. Will drop an email mas.

Yuk LLM researchers yang ada di Indonesia, mari kita bersama-sama mengembangkan AI ecosystem di Indonesia. Masih banyak banget area yg bisa kita research dan kembangkan selain building the foundation model itself dan impactnya juga penting for the industry and society.

@rbayuokt Anyway, identation di zed ini udah bener belum ya? Last usage gabisa auto detect dan fallback ke space mau dipaksa gimanapun juga. Cukup nyebelin sebagai tab indent user wkwk

udah mau seminggu make Zed dan udah terbiasa, kayanya gak bakal balik lagi ke vscode. cuma sayang aja di Zed ini gaada paste JSON to TS langsung buat interface, apa ada ya ? nyari" ga nemu, kalau ada yg tau rep coba 🙏🏻

We've been building an internal Claude Code plugin system at Intercom with 13 plugins, 100+ skills, and hooks that turn Claude into a full-stack engineering platform. Lots done, more to do. Here's a thread of some highlights.

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

Indonesian Flood Explorer Google baru-baru ini share dataset banjir dari artikel berita dari tahun 2000 sampai dengan tahun 2026. Datanya cukup besar sekitar 600MB tetapi data tersebut bisa di filter untuk region Indonesia saja dan size sekitar 25MB. Data ini berasal dari Groundsource, inisiatif metodologi dari Google yang memanfaatkan Gemini, yang mengubah jutaan laporan publik menjadi arsip data berkualitas tinggi untuk membantu memprediksi krisis. Kalau saya lihat data dari tahun ke tahun banjir di Indonesia memang tambah parah. App di bawah ini saya buat dgn vibe coding pake GPT-5.4. tujuan hanya ingin render data scr offline & belum ada tujuan khusus (kalau ada ide komen aja) Tech Stack: - Tauri + Rust - DuckDB - Svelte Link app menyusul. Bookmark kalau berkenan.

ini cuman saya apa emang token cache di codex setinggi ini ya?

Halo! Lulusan Linguistik di sini! Saya menemukan video menarik ini dari akun Instagram [@]batyalael. Sejujurnya, pertanyaan “nyeleneh” kayak gini, hampir pasti ada jawabannya. Ini observasi yang tajam banget, dan justru jadi pintu masuk ke beberapa poin dalam teori pemerolehan bahasa.

tadi di masjid tempat gue jumatan, khatibnya ngomong sendiri

We believe the coding agent is dead. Soon, Amp will look very different. ampcode.com/news/the-codin…

@SkylerMiao7 Any plans to bring out a CLI version? Seems legit to try

you guys are probably tired of me yapping about code and agents. but MiniMax-M2.5 isn't just SOTA at coding, it's a legit workspace companion. Research, organize, deliver. All from one prompt. coming soon.

Salah satu akar dari kisruh BPJS PBI The Ethics of R-squared: A True Story of How Regression Decides Who Live... youtu.be/2FP3DrSGCkE?si… via @YouTube

Here's what GitHub's last CEO @ashtom has been up to: building an agent-first dev platform. I'm an investor. They just shipped Checkpoints. It's open source, easy to use. It adds agent context (eg trajectories, prompts, token usage etc) to PRs. Get it at entire.io

@levifikri Wow, pake model apa ini mas? Claude?

Bab2nya ini. Dan ini cuma 2-3 menitan jadi. Imagine, bisa dikirimin konten seperti ini tiap hari 🤩 Saya juga minta si bot decide sendiri saya harus belajar apa nextnya. Tapi kalau lagi pingin belajar specific thing, simply tinggal minta aja

Sekalian deh, minta bikinin website. Full dilakukan via tablet tanpa coding apapun 😁. Paling tadi bantu setup GoDaddy, simply karena masih worry kasih GoDaddy key ke OpenClaw learn.levifikri.com

@levifikri Given current architecture (transformers/next token prediction), seems unlikely. Makin lama makin kebeli sama argumen @ylecun kalo current model belum cukup pintar dan efisien sampai tipping point advancing human

Yang pertama bisa bikin self improving AI, dia akan menguasai dunia.. *provided that punya plenty of resources untuk jalanin itu

@fess10nopember Coba lapor ke satgas PPKS instagram.com/satgasppks.its

Permisi min aku juga salah satu mahasiswa ITS. Pacarku juga pernah ngalamin kayak gitu cuma ngak parah. dia sering di catcalling pas dia pulang kuliah lewat gate utama yang arah gebang. Malah yang catcalling itu dari pegawai yang jaga gate mungkin ngak pacarku aja mungkin beberapa mahasiswi ITS mengalami itu juga min kalau 1 kali 2 kali okeh sih min tapi kalau setiap pulang ada yang catcalling ya risih. Sampai pernah aku tegur si pegawai yang catcalling itu malah dianya senyum2 kayak ngak ngerasa bersalah. Semoga ITS lebih berbenah lagi mengenai ini agar lingkungan ITS lebih nyaman untuk mahasiswa ITS terutama mahasiswa cewek.🙏

Single biggest improvement to your CLAUDE.md / AGENTS.md: "When I report a bug, don't start by trying to fix it. Instead, start by writing a test that reproduces the bug. Then, have subagents try to fix the bug and prove it with a passing test."

I used to work as a programmer at Credit Suisse. I was working with some guys on an API for financial instrument static reference data. Basically data for bonds, equities, mutual funds and other financial instruments. Data that wasn't supposed to change very often. Stuff like: - A bond's name - A bond's expiry date - The country a mutual fund is domiciled in Anyway there was this team that was responsible for inputting mutual fund data into some system we were replacing. We asked them to use our web UI to input the data, but they said: "No. We like using Excel, we want to input the data with Excel" The data to input was kind of complicated and required lots of validation. We didn't know how to make such a complicated Excel spreadsheet. Luckily there was a team at Credit Suisse called DeskDev. They specialised in making Excel spreadsheets. It was a team of about 10 people and they all smoked and had really bad skin. They had to code in VBA. A lot of traders used the spreadsheets they made so they spent a lot of time on the phone getting yelled at by traders. Basically the job sucked, so people who joined this team got paid extra money. Extra money in exchange for being miserable. We asked DeskDev to make this spreadsheet for us but they said "No, we have too many spreadsheets to support as it is." We talked a bit more and eventually they agreed to make it, but the team I worked on had to support it. They made the sheet, and it worked pretty well. Then a few months later I got an email from a user saying the spreadsheet wasn’t working. I went over to where DeskDev sat and they found they weren’t at their desks. Damn. I then tried out the spreadsheet and it wasn’t working. Shit. In 2009 I had a job writing VBA. So I opened up the code to figure out what was going wrong. Oh man what the hell is this… I tracked through it and found for some reason it was throwing an error when trying to parse JSON into an object. I reproduced the error with a small function like this: // String text = “{ “key”: “value” }” // Object parsed = parse_json(text); The code crashed on the second line. I then looked over to the DeskDev area and I saw one guy was at his desk. A guy called Tommy. I walked over and said “Hey, I have a sheet that’s crashing when trying to parse JSON. Can you help me take a look at it” “Oh, I don’t know how to do that” “It’s two lines of code, I think you’ll get it.” He comes over to my desk and I show him the two lines: “Hmm that’s strange, looks okay but really I don’t know how this stuff works so I can’t really help you” He went back to his desk and I thought “How does he not know how to parse JSON? DeskDev must do that all the time. Most of our APIs give out JSON responses...” I went to lunch and when I came back I saw the boss of DeskDev was at his desk. I asked him to look at the crash. “It’s crashing because you need to call initaliase_json_parser() before using that library.” We add that line and the sheet works. “Thanks man. Hey I asked Tommy about this and he didn’t know how to fix it. How come?” “Oh haha, Tommy doesn’t know anything.” “Oh okay, so why don’t you get rid of him” “Hey come on, he’s a nice guy. Also he has kids. Think about them.” Tommy was a nice guy. He was so nice that people were willing to let him have a job even though he didn’t seem to do or know anything.

@aidenybai how's repogrep is better than deepwiki? i am now hooked on deepwiki so much

want to know how the new X algo works? replace github.​com/twitter/the-algorithm with repogrep​.com to instantly get your questions answered

just encountered a weird bug in opencode, @thdxr. is this opencode or ghostty?

@ainunnajib Memang terjadi secara global. Sepertinya memang X udah mapping demografi user base nya, terus naikin exposure contrarian belief buat naikin engagement Referensi dikasih chatgpt

@famasya udah A/BCDEFGHIJKLMNOPQRSTUVWXYZ test kalau saya jelas cuma diviralkan ketika sesekali nyeletuk yang bisa disalahfahami oleh zeitgeist amuk massa netizen X indonesia

bangsa yang secara kolektif tidak menghargai kalangan pendidiknya, berarti secara kolektif memang layak untuk tidak terdidik //TW rage bait: netizen yang mengeluhkan rendahnya gaji guru & dosen di sini, banyak juga yang mengolok-olok budaya mengamplopi ahli ilmu dibilang feodal