Ukuzithuthukisa okuyisiphindaphindo, ngaphansi kwendawo efihliwe

Lokho okuhlanganiswa yile vidiyo

• 1965 → 2008: U-"kuqhuma kobuhlakani" kuka-I. J. Good, i-AI yembewu ka-Yudkowsky — umshini obhala kabusha izisindo zawo
• 2025: I-AlphaEvolve — ukuphindaphinda okungu-48 kwe-matrix engu-4x4, u-0.7% we-compute ye-Google ibuyisiwe, ama-Gemini kernels ashesha ngama-23%
• 2026: I-Dream-RSI — inqubomgomo iyathuthuka, umbhali wamakhodi, umahluleli nombhali kabusha bonke bafriziwe; umyalo uthi "Ungayixazululi inkinga yesayensi"
• U-X: "IGoogle isanda kuthuthukisa ukuzithuthukisa okuyisiphindaphindo" (ukungathandi okungu-819) vs HN: "ukubiza lokhu nge-RSI kubukeka kudukisa"
• Izinombolo ezisetshenzisiwe: §4.1 I-Lasso 550 → izingcingo zabenzeli ezingu-317, 3587 → 2931 ms; Ithebula 1 ukufaka isiyingi 2.635983 = AlphaEvolveV2; §4.3 I-KernelBench 2.43× / 1.79× izizukulwane ezincane, kuze kube yi-2.09× ngokushesha; Isijobelelo B.2 umyalo (konke okuvela ku-PDF engenhla)

Umbhalo obhaliwe ohunyushwe

Kuhunyushwe kusuka ekulandiseni kokuqala kwesiNgisi. Umsindo namazwibela atholakalayo alawulwa yi-YouTube.

0:00 Ukuzithuthukisa okuyisiphindaphindo umbono wokuthi i-AI izenza ihlakaniphe, bese isebenzisa lokho kuhlakanipha ukukwenza futhi, futhi kuleli sonto iGoogle ishicilele iphepha elinamawo lawo magama esihlokweni, lapho izisindo zemodeli zingalokothi zinyakaze. Izinombolo ezintathu. I-CEO yakwa-Anthropic ithi le loop ibilokhu isebenza kusukela ehlobo, okuyisizathu sakhe sokubambezela yonke imboni. I-tweet ememezela ukuthi iGoogle isanda kuyithuthukisa yaqoqa ukuthandwa okungamakhulu ayisishiyagalombili ngosuku.

0:24 Futhi umphumela wesihloko sepepha uqobo lwawo uyingcingo yemodeli engu-317 esikhundleni sika-550, okuyisaphulelo, hhayi ukusuka. Ngemizuzu emithathu: lapho igama livela khona, yini empeleni ephindaphindeka ngaphakathi I-Dream-RSI, nokuthi ungalifunda kanjani iphepha elilandelayo eline-RSI ekhavini. Lena yi-The Daily Diff, ngaphansi kwendawo efihliwe. 1965. U-I. J. Good, isazi sezibalo sase-Bletchley Park esasebenza eduze no-Turing, ubhalela ukuthi umshini ohlakaniphe kakhulu ungakha nemishini engcono kakhulu,

0:52 akubize ngokuthi ukuqhuma kobuhlakani kanye nokuqanjwa kokugcina umuntu okufanele akwenze, inqobo nje uma umshini uthobekile ngokwanele ukusitshela ukuthi singawulawula kanjani, okuyilo nalelo nalabo bantu abayeka ukufunda ngemuva kokhefana. Iminyaka engamashumi amane leli gama lisho lokho kanye: uhlelo oluhlela olwalo umthombo noma izisindo bese luphuma luhlakaniphe kakhulu ngaso sonke isikhathi. Indatshana ka-Eliezer Yudkowsky yango-2008 yayenza igama elithwala umthwalo we-AI ukuphepha, futhi akekho owayenomshini wokukuhlola kuwo, okuyisimo esifanele se- incazelo. Bese ngoMeyi 2025 i-DeepMind yethula i-AlphaEvolve,

1:23 i-ejenti ye-Gemini ehlongoza ikhodi, iyahlolwa, igcina abawinile bese iphinde iyashintsha. Yaqala ukuphindaphinda ama-matrix ayinkimbinkimbi angama-four-by-four ngezinyathelo ezingu-48, iwehlula irekhodi u-Strassen alisetha ngo-1969, futhi ibuyisa cishe u-zero point seven percent we-compute ye-Google emhlabeni jikelele, okwizinga le-Google kuyi- isikhungo sedatha esitholakala ngaphansi kukasofa. I-Gemini manje ibisisiza ukuqeqesha i-Gemini, futhi leli gama lahamba buthule lisuka ekuphepheni emabhlogi laya ezitatimendeni zokukhishwa kwabezindaba. I-Dream-RSI iqinisa isilawuli phezu kwalowo loop.

1:52 Inqubomgomo, uhlelo lwe-Python lwangempela, lunquma ukuthi yimaphi amagatsha esihlahla sokusesha okufanele awanwe, mangaki ngokufana, nokuthi kufanele unikezele nini. I-Gemini ibhala ikhodi yokuba umuntu, futhi umhloli omisiwe uyayihlomula. Konke okwenziwayo kushiya isihlahla sokuthi kwazanywa ini nokuthi kwahlomula ini. Leso sihlahla siba umhloli wokudlala kabusha: i-Gemini yesibili, efriziwe ngokulinganayo, ibhala kabusha inqubomgomo, futhi inqubomgomo entsha ihlolwa ngokumelene nemiphumela erekhodiwe esikhundleni sama-GPU wangempela.

2:16 Iphepha libiza lokhu ngokuthi ukuphupha, futhi ukusebenza okukodwa kwangempela kubhadala izinkulungwane zokuphupha ngaphandle kwezindleko zokusebenza. Owine uya ku-inthanethi futhi, inqwaba yemihlaba erekhodiwe iyakhula, phinda. Futhi umyalo ku-Isijobelelo B.2 utshela i-AI ezithuthukisayo, ngokubhala: hlela leli fayela elilodwa kuphela, futhi ungayixazululi inkinga yesayensi. Kulinganiswe. Ku-task yokuxazulula i-Lasso, ukuhlola okumisiwe kwasebenzisa izingcingo ze-Gemini ezingu-550 ukuze kufinyelele imizuzwana emithathu point six, i-Dream-RSI yasebenzisa engu-317 ukufinyelela amabili point nine, futhi bobabili bahlula i-scikit-learn.

2:46 Ku-KernelBench yafinyelela isivinini se-kernel esifanayo ngezikhathi ezingaba amabili point four izizukulwane ezimbalwa. Futhi ekuphakeni kwesiyingi yahlomula amabili point six three five nine eight three, okuyikho impela, kuze kufike ezinombolweni eziyisithupha, lokho i-AlphaEvolve version two yakuhlomula ngonyaka odlule. Ngakho u-X wafunda isihloko: IGoogle isanda kuthuthukisa ukuzithuthukisa okuyisiphindaphindo. I-Hacker News yafunda i-PDF: ukubiza lokhu nge-RSI kubukeka kudukisa. Futhi kulelo sonto i-CEO yomqhudelani yathi i-loop ibilokhu isebenza kusukela ehlobo futhi wonke umuntu kufanele anciphise isivinini.

3:13 Imisho emithathu, amagama amabili afanayo, imishini emithathu ehlukene. Ngakho, ngoMsombuluko, ukuthi ungalifunda kanjani iphepha elilandelayo le-RSI. Okokuqala: buza ukuthi yisiphi isihloko esishintshayo, izisindo, ikhodi, noma izilungiselelo zokusesha; I-Dream-RSI ishintsha okwesithathu. Okwesibili: buza ukuthi ubani ofriziwe; lapha umbhali wamakhodi, umahluleli nombhali kabusha, bonke abathathu. Okwesithathu: buza ukuthi inombolo yesihloko iyiyunithi yani. I-compute egciniwe iyilungiselelo; ibhentshimakhi imodeli eyayingakwazi ukuyifinyelela ngaphambili yi- ukusuka, futhi akekho okushicilele lokho okwamanje.

3:40 Isahlulelo, ngaphansi kwendawo efihliwe: IDINGA UKUBUYEKEZWA. I-loop ingokoqobo, izindleko zilinganisiwe, futhi amagama amabili ekhavini achaza umshini ongekho ephepheni. Uma ungathanda ukukufunda lokhu kunokuba ungizwe ngikusho, i-diff ifika ebhokisini lakho leposi njalo ekuseni, mahhala ku-the daily diff dot dev, isixhumanisi esingezansi. Ngitshele ukuthi yini okufanele ngiyivule ngokulandelayo kumazwana. Futhi yilokho i-diff yanamuhla. Ngingu-Niko ovela kwa-Axrisi.

4:00 Hlanganisa ngokuzibophezela.

Imithombo

1. Paper: Dream-RSI (Zheng et al., Google / Google DeepMind / UMD / UVA, 14 Sep 2026)arxiv.org
2. Code (293 stars, no license) — https://github.com/zhengkid/Dream-RSI · project pagedream-rsi.com
3. Hacker News thread (169 points)news.ycombinator.com
4. Hugging Face papers (278 upvotes)huggingface.co
5. I. J. Good, 1965, "Speculations Concerning the First Ultraintelligent Machine"en.wikipedia.org
6. Eliezer Yudkowsky, "Recursive Self-Improvement", LessWrong, 1 Dec 2008www.lesswrong.com
7. Google DeepMind, AlphaEvolve (14 May 2025): 0.7% compute, 48 multiplications, 23% kernel speed-updeepmind.google
8. Dario Amodei, "We Must Pace the Frontier" (12 Sep 2026)darioamodei.com
9. Tweetx.com