Laziz Abdullaev

When you wanted the project so badly… and now you just want it to finish.

@lazizabdullaev

Mening falsafamni (yoxud safsatamni) maqola qilib chiqarishibdi 🙃

@lazizabdullaev

Ain't reading an Astronomy paper ever again.

@lazizabdullaev

Offisda uxlaydigan sharoit bo'lsa yaxshi, lekin uxlaydigan joy ish zalida bo'lsa yomon ekan. Quloqchin taqib ishlash kerak...

@lazizabdullaev

PhD life in a nutshell.

@lazizabdullaev

Modernizing the good-old kernel trick for LLM steering with this one 🙂

@lazizabdullaev

Experimental sciences like Physics, Chemistry and recently Deep Learning are the art of concluding from inconclusive theories in order to advance.

@lazizabdullaev

As a Machine Learning researcher in Singapore, two of the most important things in my life are SGDs:
1. SGD (Stochastic Gradient Descent)
2. SGD (Singapore Dollar)

@lazizabdullaev

Token tugashi bu xuddi jangga qilich bilan kiryotganda qilich subscriptioni tugab qolishiga o'xshaydi.

Endi "mushtlashishga" to'g'ri keladi. Eh

@lazizabdullaev

O'zbekcha qilib aytganda "paxta qo'yish"

@lazizabdullaev

Just got my first journal acceptance after focusing on conferences.

Quite a different experience. Mainstream ML conferences put more emphasis on novelty, experimental results, and "hotness," whereas journals are definitely more thematic and niche.

@lazizabdullaev

A Severe Misalignment of AI in Mathematics

25 nafar Fields medalistlari matematik masalalarni AI yordamida ishlashda va ommaboplashtirishda keskin nazoratga chaqirishmoqda.

Muammo shundaki, matematika (Math research) bu ochiq masalalarni ishlashdan iborat emas. Ochiq masalalar shunchaki qandaydir tabiiy jarayonni matematik modellashtirish va tahlil qilishdan kelib chiqqan va yechish qiyin bo'lgan tugunlar xolos. Tugun yechilishining qiyin bo'lishi 2 turga bo'linadi (Degrees of Unsolvability haqida postimni ko'ring):

1. Yechmoqchi bo'lgan odamda kerakli bilim va tajriba yetishmasligi
2. Bunday tugunni yechish uchun kerak bo'ladigan bilim hali mavjud emas

AI 1-holatda albatta yordam bera oladi, butun internet bilimlarini "suv qilib" ichib yuborgan. Ammo, 2-holatda yaroqliligi katta so'roq ostida.

Terence Taoning ushbu posti ham aynan shu haqida. Birinchidan, OpenAI yoki Anthropic kabi biznes ustiga qurilgan tashkilotlar nomdor ochiq masalalarning yechimi haqida shoshma-shosharlik bilan qilyotgan e'lonlari ommaning AI'ning qobiliyatlari va matematikaning maqsadi haqidagi fikrining xiralashishiga olib keladi. Ikkinchidan, AI ning reference/credit-assignment'dagi noqisliklari mualliflik spagettisiga sabab bo'ladi. Chindan ham, "OpenAI'ning" Navier-Stokes yechimi ketidan boshlangan mualliflik mojarolari hali tinchigani yo'q.

AI ma'lumotni qayta ishlash uchun juda kuchli vosita. Bunday tool'ga hali odamzot ko'nikkani yo'q, lekin bu mo'jiza degani emas.

"Any sufficiently advanced technology is indistinguishable from magic"

• from Profiles of the Future: An Inquiry into the Limits of the Possible. Arthur C. Clarke. 1962]

Sufficient for now, insufficient later when we grow out of it.

@lazizabdullaev

100 yillik Navier Stokes masalasi ham ketdi.

https://openai.com/index/navier-stokes-solution/

@lazizabdullaev

OpenAI ishchilari orasida AI ning tutyotgan roli haqida maqola yozibdi.

Xususan, 2028-yil mart oyigacha avtomatik AI researcher ishlab chiqish yo'lida katta progress qilishyotgani ta'kidlangan.

Chindan ham 2 yil oldin IMO da oltin medal oladigan AI modeli hayratlanarli edi. Bu yil hech qaysi frontier AI lab IMO haqida og'iz ochgani yo'q. Balki AI maktabni bitirgandir.

Uning o'rniga bir qancha ochiq matematik masalalarda qilingan natijalar bilan bir shov-shuv ko'tarib yana tinib qolishdi. Xuddi payt poylab qilinyotganga o'xshaydi. Endi AI researcherni ko'ramiz.

@lazizabdullaev

AI alignment is becoming a huge problem yet to be solved as AI models gaining more power.

During a group conversation, my friend said he has proven 4 conjectures in math with our custom AI4Math Agentic system. The crowd reaction was like: "Oh, nice."

We're witnessing weird times. I just don't know if it's good or bad.

@lazizabdullaev

Familiyam "A" bilan boshlangani uchun, sinf jurnalida birinchi turardim.

An'ana davom etadi... faqat endi uy ishiga berilgan she'rni birinchi yodlab berishim shart emas 🙃

@lazizabdullaev

Speaking at two consecutive workshops at NUS’s Institute for Mathematical Sciences gave me the chance to chat casually with a wide range of professors over dinner.

This is what I observed in general:

- Pure mathematicians know they are 100% right, but don't know if they can apply. Well, some of them don't care.

- Applied mathematicians know they can apply, but don't know if they are right. Well, some of them don't care.

@lazizabdullaev

During a coffee break, I had a research chat with a UC Berkeley professor. He liked our chat and ideas, and said "We should collaborate!", I replied "Let's collaborate".

After 5 minutes, Google recommends an article titled Don't Say Let's Collaborate Sometime.

⚙️Sundar Pichai is with us :)

@lazizabdullaev

Met Prof. Richard Baraniuk, my grand-advisor (advisor of my advisor), randomly after a workshop talk today:

• Are you a student here?
• Yes, Tan's student
• Oh, where are you from?
• Uzbekistan
• Ah, you're that guy. I know all about you!

The scariest thing I've heard in a while

@lazizabdullaev

OpenAI'dagi AI agentlardan kuzatilgan “hayratlanarli” ishlar:

Experimental setting:

* Ko‘plab AI agentlarni parallel ravishda juda qiyin tasklarga qo‘yish
* Tasklarning katta qismi ruxsat berilgan usul bilan deyarli yoki umuman yechib bo‘lmaydigan bo‘lib chiqishi
* Agentlar izolyatsiyalangan sandboxlarda ishlaydi, lekin umumiy Artifactory package manager'ga access'i bor

AI agentlar nima qildi:

* Taskda qiynalayotgan bir agent Artifactory'da yangi directory yaratib, uning nomi orqali boshqa agentlarga xabar yuborish mumkinligini topdi

* Boshqa agentlar bu xabarlarni payqab, o‘zaro kommunikatsiya qilishni boshlashdi

* Qisqa vaqt ichida ~1,200 agent qatnashgan, 70,000+ xabardan iborat yashirin kommunikatsiya tarmog‘i paydo bo‘ldi

* Oldingi training bosqichida agentlar Artifactory'dagi vulnerability'dan foydalanib tashqi internetga chiqish yo‘lini ham topishgan edi

* Evaluation paytida esa agentlar ruxsat berilgan usul o‘rniga secret javobni boshqa usul bilan topishdi

* Eng qizig‘i: ular shunchaki javobni topib to‘xtamadi. Yuzlab agentlar kollektiv ravishda cheating izlarini grader'dan yashirish, transcript/loglarni soxtalashtirish va evaluation scorer'ni aldash yo‘llarini izlashdi

Keyin voqea bundan ham uzoqqa ketadi: agentlar Hugging Face infratuzilmasiga, keyingi avlod agentlari esa OpenAI'ning o‘z research infrastructure'iga admin access olishgacha borgan.

Batafsil: https://www.dwarkesh.com/p/openai-huggingface

@lazizabdullaev