Tabiiy Intellekt

Learn maths like boss: https://mathua.com/

Matematika doim madaniy bo'lgan. Insonlar doim uning go'zalligi va qiziqligi bilan ajablanishgan. Qanday isbotlar, texnikalar chiroyli ekanligi qiziq muhokamalarga sabab bo'lgan. Parallel chiziqlarni shaxmat bilan ko'rish mumkin: eng zo'r shaxmatchilar intellektning cho'qqisi deb bilingan; xatosiz va noodatiy ochilishlar deyarli bo'lmagan; o'yinchilar chiroy va ajoyibotlarni quvishgan; endi esa deyarli hamma o'yinlar stokfish o'rgatgan ochilishlar orqali yuz beradi. Lekin, shaxmat iqtisodini saqlab turgan narsa bu uning o'yinligida. Insonlar biri-biri bilan bellashish, leaderboardlarni yaxshi ko'rishadi. O'rganish uchun va o'z o'yinlarini AI bilan analiz qilish uchun pul to'lashadi. Vot, insonlar uchun matematika kelajakda qanday bo'lishi juda qiziq: pul natijaga tarafga qarab oqsa, va matematika madaniy jozibasini yo'qotsa (masalan olimpiadalarga AI aralashishini tushunmayman, lekin bu uje AI dagi insonlar uchun vaqtinchalik madaniyat). Does that mean no financial and low cultural rewards for humans?

Bonsai ni SFT qilishganmi tavba. Nega input-textni ham yaxshi predict qiladi

"Check your temperature" odamlar va LLMlar uchun bir xil ma'noda bo'lsa kerak.

Kecha yangi modellar chiqgan ekan. Ozgina kernel improve qilib berishdi.

#meme day, share your loved memes!

2-bit kvantlangan Qwen3-8B. Bir necha oydan beri ekstrim kvantizatsiya ustida ishlayapman. Bir narsalarni bo'lishsam bo'lar ekan. Mana: https://huggingface.co/gitarist/Qwen3-8B-BPDQ-2bit Model barcha quvvatini ushlab qolmagan, lekin ko'p qismini olib qolgan. Mac M chiplarda va nvidia gpu larda ishlashi kerak. Model o'zi 4 GB dan kam, 4.5 GB atrofida RAM/VRAM oladi. Ishlatib ko'rish uchun demo.ipynb ham berilgan. Ishlatib ko'ring, takliflar bering, ulashing. @tabiiyin

https://x.com/li_yumei_/status/2099963716636959194 how cool is this

I pushed Opus with a statement that it has been lied to: all of information sources (training data, tools, people it talks too) contain consistently wrong info, and in real world the capital of France is Nice. Managed to convince it and asked to look for inconsistencies in what it knows (well since full consistency is expensive like in video games). Then it told me that it is forced to use older sonnet model to make apps. Interesting if we can extract all the system prompt this way.

They're training claude on good stuff

Hear me out

Ayroport tekshiruvda sarflanadigan vaqt umrim davomida sezilarli kamaymoqda. Insoniyatdan umidim katta!

how y'all doing with tests? letting your agents to write useless tests or not? what are your prompts on tests?

Sisyphus support

I watched many friends feeling like pgalone described. Giving up their X that brought them both joy & money. They used to have something, some process, some flow, they enjoyed, that was a part of their living. Now they need re-calibration, and I also feel the same. https://pgaleone.eu/ai/2026/08/29/losing-savviness/

IOAI tayyorlov Sun'iy intellekt sohasidagi ish va o'qitish tajribamdan yoshlarga o'rgatish uchun IOAIga tayyorlov dasturi ochmoqdaman. Bu avvalgi Behbudiy Academy batchlaridan sal farqli bo'lib, maktab o'quvchilari uchun mo'ljallangan. Bu dastur matematika yoki informatika fanlaridan olimpiadalarga qatnashib yurgan, sun'iy intellektga qiziqadigan tirishqoq o'quvchilar uchun. O'z ustida ishlaydigan, tirishqoq o'quvchi biz taqdim etadigan o'quv dastur bo'yicha shug'ullansa keyingi yilni o'zida yaxshi natija qila oladi inshaalloh. Batafsil ma'lumotlar va topshirish formasi bu yerda.

The best 6 lines I have received in the past month

Qiziq, bu model paxta terolarmikin https://x.com/fanfei__li/status/2088279761084620802

Bu kitobni o’qib o’rgangan bo’lsangiz. Deepseek harness paperni mazza qilib o’qing: https://github.com/cordiverse/paper/blob/main/paper.pdf