Forwarded from Nuruddin Blogs
LLM yordamida tez va effektiv o’rganish uchun quyidagi metodikalardan foydalaninglar deyapti Andrey Karpati
Biz til modellarining natijalarini tushunishga ancha ko‘proq vaqt sarflashimizga to‘g‘ri keladi. Bir nechta fikrlar, maslahatlar va tavsiyalar:
Yozish. Men muvaffaqiyatga erishgan bir usul: LLM’dan biror narsani ASD-STE100 formatida tushuntirib berishini so‘rang. Bu dastlab aerokosmik texnik xizmat ko‘rsatish hujjatlari uchun ishlab chiqilgan boshqariladigan til spetsifikatsiyasidir. LLMlar bu tilni yaxshi bilishadi va u toza yozish uslubiga qat’iy cheklovlar qo‘yadi, bu esa matnni ancha tushunarli qiladi deb hisoblayman. Ba’zida men buni biroz yumshatishga harakat qilib, masalan, "80% ASD-STE100 darajasida" yozishni so‘radim, chunki spetsifikatsiya juda qat’iy. Ammo bundan ham yaxshisi:
Diagrammalar / rasmlar. Yozish o‘rniga, LLM’dan diagramma yaratishni so‘rang. Ularni qayta ishlash, tahlil qilish va tushunish ancha oson bo‘lishi mumkin. Ammo bundan ham yaxshisi:
Veb-sahifalar. Chiroyli, interaktiv veb-sahifa olish uchun natijani "HTML formatida" berishini so‘rang. LLMlar frontend dasturlashda juda yaxshilanib bormoqda va chiroyli interfeyslar, animatsiyalar va hokazolarni yarata oladi. Ammo bundan ham yaxshisi:
Tushuntiruvchi videolar. Men eng istiqbolli deb biladigan natija formati – bu ixtiyoriy mavzuda yaratilgan to‘liq maxsus tushuntiruvchi videolardir. "X haqida 3b1b uslubida tushuntiruvchi video yarat. Audio hikoya uchun mening ElevenLabs API kalitimdan foydalan" kabi buyruqlarni sinab ko‘ring. (Ikkinchisi uchun sizga API kaliti kerak bo‘ladi yoki LLM’dan mahalliy kompyuteringiz quvvatidan foydalanadigan munosib bepul muqobillarni topishni so‘rashingiz mumkin). Bu haqiqatan ham ishlay boshladi!
Xulosa qilib aytganda:
- LLMlar takomillashib borishi bilan, ular ko‘proq ishlarni mustaqil ravishda bajaradi va bizning ishimiz asosan nazorat qilish va tushunish kabi yuqoriroq darajalarga ko‘tariladi.
- Yaxshiyamki, LLMlar bu yerda ham yordam bera oladi, chunki intellekt va kod tobora ko‘payib borayotgani sababli, siz ilgari yaratish mantiqqa to‘g‘ri kelmagan katta, maxsus, bir martalik dasturiy ta’minot artefaktlarini (masalan, veb-ilovalar, tushuntiruvchi videolar) so‘rashingiz mumkin. Bu borada chegaralarni kengaytiring va natijadan o‘zingiz ham hayratda qolasiz.
Manba: https://x.com/karpathy/status/2105819303471976479
@nuruddinblogs
Biz til modellarining natijalarini tushunishga ancha ko‘proq vaqt sarflashimizga to‘g‘ri keladi. Bir nechta fikrlar, maslahatlar va tavsiyalar:
Yozish. Men muvaffaqiyatga erishgan bir usul: LLM’dan biror narsani ASD-STE100 formatida tushuntirib berishini so‘rang. Bu dastlab aerokosmik texnik xizmat ko‘rsatish hujjatlari uchun ishlab chiqilgan boshqariladigan til spetsifikatsiyasidir. LLMlar bu tilni yaxshi bilishadi va u toza yozish uslubiga qat’iy cheklovlar qo‘yadi, bu esa matnni ancha tushunarli qiladi deb hisoblayman. Ba’zida men buni biroz yumshatishga harakat qilib, masalan, "80% ASD-STE100 darajasida" yozishni so‘radim, chunki spetsifikatsiya juda qat’iy. Ammo bundan ham yaxshisi:
Diagrammalar / rasmlar. Yozish o‘rniga, LLM’dan diagramma yaratishni so‘rang. Ularni qayta ishlash, tahlil qilish va tushunish ancha oson bo‘lishi mumkin. Ammo bundan ham yaxshisi:
Veb-sahifalar. Chiroyli, interaktiv veb-sahifa olish uchun natijani "HTML formatida" berishini so‘rang. LLMlar frontend dasturlashda juda yaxshilanib bormoqda va chiroyli interfeyslar, animatsiyalar va hokazolarni yarata oladi. Ammo bundan ham yaxshisi:
Tushuntiruvchi videolar. Men eng istiqbolli deb biladigan natija formati – bu ixtiyoriy mavzuda yaratilgan to‘liq maxsus tushuntiruvchi videolardir. "X haqida 3b1b uslubida tushuntiruvchi video yarat. Audio hikoya uchun mening ElevenLabs API kalitimdan foydalan" kabi buyruqlarni sinab ko‘ring. (Ikkinchisi uchun sizga API kaliti kerak bo‘ladi yoki LLM’dan mahalliy kompyuteringiz quvvatidan foydalanadigan munosib bepul muqobillarni topishni so‘rashingiz mumkin). Bu haqiqatan ham ishlay boshladi!
Xulosa qilib aytganda:
- LLMlar takomillashib borishi bilan, ular ko‘proq ishlarni mustaqil ravishda bajaradi va bizning ishimiz asosan nazorat qilish va tushunish kabi yuqoriroq darajalarga ko‘tariladi.
- Yaxshiyamki, LLMlar bu yerda ham yordam bera oladi, chunki intellekt va kod tobora ko‘payib borayotgani sababli, siz ilgari yaratish mantiqqa to‘g‘ri kelmagan katta, maxsus, bir martalik dasturiy ta’minot artefaktlarini (masalan, veb-ilovalar, tushuntiruvchi videolar) so‘rashingiz mumkin. Bu borada chegaralarni kengaytiring va natijadan o‘zingiz ham hayratda qolasiz.
Manba: https://x.com/karpathy/status/2105819303471976479
@nuruddinblogs
🔥4
Forwarded from ANIQ VA IJTIMOIY FANLAR UNIVERSITETI
Media is too big
VIEW IN TELEGRAM
🚀 AIFU Pitch Day 3.0 - startapingiz uchun $100 000 gacha investitsiya jalb qilish imkoniyati
🎒 Aniq va ijtimoiy fanlar universiteti respublika miqyosidagi startap tanloviga arizalar qabul qilmoqda. Ishtirok - bepul.
16-40 yoshdagi dasturchilar, dizaynerlar va startap asoschilari qatnasha oladi. G'oya bosqichidagi loyihalar ham qabul qilinadi, ariza va saralash - onlayn.
🔵 Loyihalarni investorlar va hakamlar baholaydi
🔵 15 mln so'm sovrin jamg'armasi
🔵 AIFUda o'qish uchun 100% gacha grant
🔵 Har bir ishtirokchiga sertifikat
🔵 Hamkorlar: Alibaba, IT Park, INNO Technopark
O'tgan mavsumda 574 jamoa ro'yxatdan o'tdi, 30 ta jamoa investorlarga taqdimot qilish imkoniyatini qo'lga kiritdi va 5 ta loyiha investorlar va venchur fondlar bilan alohida uchrashuvga chiqqan.
11-oktyabrgacha - ro'yxatdan o'tish
18-oktyabr - final, Toshkent
Dasturchi yoki startap g'oyasi bor tanishingiz bo'lsa, bu xabarni unga yuboring - muddatga ulgursin.
👉 Ariza: pitch-day.uz
📥 @PitchDayAIFU_bot
Telegram | Instagram | YouTube | Facebook | LinkedIn | Veb-sayt
16-40 yoshdagi dasturchilar, dizaynerlar va startap asoschilari qatnasha oladi. G'oya bosqichidagi loyihalar ham qabul qilinadi, ariza va saralash - onlayn.
O'tgan mavsumda 574 jamoa ro'yxatdan o'tdi, 30 ta jamoa investorlarga taqdimot qilish imkoniyatini qo'lga kiritdi va 5 ta loyiha investorlar va venchur fondlar bilan alohida uchrashuvga chiqqan.
11-oktyabrgacha - ro'yxatdan o'tish
18-oktyabr - final, Toshkent
Dasturchi yoki startap g'oyasi bor tanishingiz bo'lsa, bu xabarni unga yuboring - muddatga ulgursin.
Telegram | Instagram | YouTube | Facebook | LinkedIn | Veb-sayt
Please open Telegram to view this post
VIEW IN TELEGRAM
Forwarded from Tsukerberg Qo’shnilari
Forwarded from Hitchhiking Inverse Log
Anthropic's PR merge rate grew exponentially, but github couldn't handle it. So folks created a new source control system (scion), new code review system (antique) and new code search, all at the same time. They look like Google's Piper, Critique and CodeSearch (cs/).
This was one of the most impressive engineering projects I've seen, carried out by a very small group of experienced engineers (and lots of tokens) for the size and ambition of these projects
plots in comments
https://www.anthropic.com/engineering/redesigning-source-control
This was one of the most impressive engineering projects I've seen, carried out by a very small group of experienced engineers (and lots of tokens) for the size and ambition of these projects
plots in comments
https://www.anthropic.com/engineering/redesigning-source-control
Anthropic
Redesigning source control at Anthropic
As our merges per day grew more than tenfold, source control became our biggest development bottleneck. Here, we share how we migrated to an in-house version control system in five months.
Please open Telegram to view this post
VIEW IN TELEGRAM
🔥3
https://lnkd.in/p/gGB2UCcz
if anyone is interested in tech and startups, i can refer you.
it is an in-person event scheduled in Paris for 2027.
@lazizbekio
if anyone is interested in tech and startups, i can refer you.
it is an in-person event scheduled in Paris for 2027.
@lazizbekio
LinkedIn
We’re hiring. Two roles, one team, a big year ahead. 🚀
At Chain of Events, we run the leading events where leaders shaping technology…
At Chain of Events, we run the leading events where leaders shaping technology…
We’re hiring. Two roles, one team, a big year ahead. 🚀
At Chain of Events, we run the leading events where leaders shaping technology come together. RAISE Summit, MACHINA Summit, and Signal Week (formerly Paris Blockchain Week), all in Paris, all in the…
At Chain of Events, we run the leading events where leaders shaping technology come together. RAISE Summit, MACHINA Summit, and Signal Week (formerly Paris Blockchain Week), all in Paris, all in the…
👍3🎃2
Benchmark Radar
An open source project I started contributing.
New version is online:
https://arxiv.org/abs/2609.11115
@lazizbekio
An open source project I started contributing.
New version is online:
https://arxiv.org/abs/2609.11115
@lazizbekio
arXiv.org
Benchmark Radar: A Living Database and Search Engine for AI...
Benchmark researchers and developers of large language models (LLMs) and other AI systems need to find relevant evaluations, locate their benchmark datasets and code, and understand the settings...
🔥1