{"id":167,"date":"2023-11-16T11:00:33","date_gmt":"2023-11-16T04:00:33","guid":{"rendered":"https:\/\/xuhuongai.com\/?p=167"},"modified":"2023-11-18T12:17:17","modified_gmt":"2023-11-18T05:17:17","slug":"chi-can-hai-nha-may-dien-hat-nhan-de-cung-cap-nang-luong-cho-nhu-cau-tri-tue-nhan-tao-cua-nhan-loai-trong-nam-toi","status":"publish","type":"post","link":"https:\/\/xuhuongai.com\/?p=167","title":{"rendered":"K\u1ef9 s\u01b0 Meta: Ch\u1ec9 c\u1ea7n hai nh\u00e0 m\u00e1y \u0111i\u1ec7n h\u1ea1t nh\u00e2n \u0111\u1ec3 cung c\u1ea5p n\u0103ng l\u01b0\u1ee3ng cho nhu c\u1ea7u AI c\u1ee7a nh\u00e2n lo\u1ea1i trong n\u0103m t\u1edbi"},"content":{"rendered":"\n
\n

B\u1ea3n tin \u0111\u01b0\u1ee3c d\u1ecbch v\u00e0 t\u00f3m t\u1eaft b\u1edfi n\u1ec1n t\u1ea3ng t\u1ea1o tr\u1ee3 l\u00fd AI – KamiMind<\/a>.<\/p>\nNgu\u1ed3n: Matt Marshall, “Meta engineer: Only two nuclear power plants needed to fuel AI inference next year<\/a>“, VentureBeat, 13\/11\/2023.<\/cite><\/blockquote>\n\n\n

\n
\"\"
\u1ea2nh: Tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o v\u1ec1 h\u1eadu qu\u1ea3 h\u1ea1t nh\u00e2n c\u1ee7a \u0110\u1ea1i h\u1ecdc Tokyo<\/figcaption><\/figure>\n<\/div>\n\n\n

Gi\u00e1m \u0111\u1ed1c k\u1ef9 thu\u1eadt c\u1ee7a Meta v\u1ec1 AI t\u1ea1o sinh, Sergey Edunov, tin r\u1eb1ng ch\u1ec9 c\u1ea7n hai nh\u00e0 m\u00e1y \u0111i\u1ec7n h\u1ea1t nh\u00e2n m\u1edbi s\u1ebd \u0111\u1ee7 \u0111\u1ec3 \u0111\u00e1p \u1ee9ng nhu c\u1ea7u ng\u00e0y c\u00e0ng t\u0103ng v\u1ec1 \u1ee9ng d\u1ee5ng tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o trong n\u0103m t\u1edbi. Edunov \u0111\u00e3 \u0111\u01b0a ra tuy\u00ean b\u1ed1 n\u00e0y trong m\u1ed9t bu\u1ed5i th\u1ea3o lu\u1eadn t\u1ea1i Di\u1ec5n \u0111\u00e0n C\u00f4ng nh\u00e2n K\u1ef9 thu\u1eadt s\u1ed1 t\u1ea1i Thung l\u0169ng Silicon. \u00d4ng gi\u1ea3i th\u00edch r\u1eb1ng nh\u1eefng nh\u00e0 m\u00e1y \u0111i\u1ec7n n\u00e0y s\u1ebd c\u00f3 kh\u1ea3 n\u0103ng cung c\u1ea5p n\u0103ng l\u01b0\u1ee3ng cho nhu c\u1ea7u tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o c\u1ee7a nh\u00e2n lo\u1ea1i trong m\u1ed9t n\u0103m. Edunov \u01b0\u1edbc t\u00ednh r\u1eb1ng n\u1ebfu t\u1ea5t c\u1ea3 c\u00e1c GPU H100 do Nvidia ph\u00e1t h\u00e0nh v\u00e0o n\u0103m t\u1edbi \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng \u0111\u1ec3 t\u1ea1o ra c\u00e1c token cho c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef, th\u00ec ch\u1ec9 c\u1ea7n hai nh\u00e0 m\u00e1y h\u1ea1t nh\u00e2n \u0111\u1ec3 cung c\u1ea5p n\u0103ng l\u01b0\u1ee3ng cho ch\u00fang. \u00d4ng c\u0169ng th\u1ea3o lu\u1eadn v\u1ec1 nh\u1eefng th\u00e1ch th\u1ee9c trong vi\u1ec7c \u0111\u00e0o t\u1ea1o c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef, nh\u1ea5n m\u1ea1nh s\u1ef1 c\u1ea7n thi\u1ebft c\u1ee7a m\u1ed9t l\u01b0\u1ee3ng d\u1eef li\u1ec7u \u0111\u1ee7. \u00d4ng suy \u0111o\u00e1n r\u1eb1ng GPT4 c\u00f3 th\u1ec3 \u0111\u00e3 \u0111\u01b0\u1ee3c \u0111\u00e0o t\u1ea1o tr\u00ean to\u00e0n b\u1ed9 internet. Tuy nhi\u00ean, \u00f4ng l\u01b0u \u00fd r\u1eb1ng c\u00f3 th\u1ec3 kh\u00f4ng c\u00f3 \u0111\u1ee7 d\u1eef li\u1ec7u c\u00f4ng c\u1ed9ng \u0111\u1ec3 hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh t\u01b0\u01a1ng lai, v\u00e0 c\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u \u0111ang nghi\u00ean c\u1ee9u c\u00e1c k\u1ef9 thu\u1eadt hi\u1ec7u qu\u1ea3 v\u00e0 t\u00ecm ngu\u1ed3n d\u1eef li\u1ec7u thay th\u1ebf \u0111\u1ec3 gi\u1ea3i quy\u1ebft th\u00e1ch th\u1ee9c n\u00e0y. Nh\u00ecn chung, c\u00e1c di\u1ec5n gi\u1ea3 trong bu\u1ed5i th\u1ea3o lu\u1eadn \u0111\u1ed3ng \u00fd r\u1eb1ng c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef \u0111\u00e3 ch\u1ee9ng minh \u0111\u01b0\u1ee3c gi\u00e1 tr\u1ecb \u0111\u00e1ng k\u1ec3 v\u00e0 c\u00e1c doanh nghi\u1ec7p c\u00f3 th\u1ec3 s\u1ebd tri\u1ec3n khai ch\u00fang r\u1ed9ng r\u00e3i trong v\u00f2ng hai n\u0103m t\u1edbi. H\u1ecd c\u0169ng d\u1ef1 \u0111o\u00e1n r\u1eb1ng trong ba \u0111\u1ebfn b\u1ed1n n\u0103m t\u1edbi, li\u1ec7u tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o t\u1ed5ng qu\u00e1t (AGI) c\u00f3 kh\u1ea3 thi v\u1edbi c\u00f4ng ngh\u1ec7 hi\u1ec7n t\u1ea1i hay kh\u00f4ng.<\/p>\n\n\n\n

B\u1ea3n t\u00f3m t\u1eaft ti\u1ebfng Anh<\/summary>\n

Meta’s director of engineering for Generative AI, Sergey Edunov, believes that just two new nuclear power plants would be sufficient to meet the increasing demand for AI applications in the next year. Edunov made this statement during a panel session at the Digital Workers Forum in Silicon Valley. He explained that these power plants would be able to power humanity’s AI needs for a year. He specifically focused on the power requirements for AI inference, which is the process of deploying AI in applications to respond to questions or make recommendations. Edunov estimated that if all the H100 GPUs released by Nvidia next year were used to generate tokens for language models, it would require only two nuclear reactors to power them. He also discussed the challenges of training LLMs, emphasizing the need for a sufficient amount of data. He speculated that GPT4, for example, may have been trained on the entire internet. However, he noted that there may not be enough public data available for training future models, and researchers are exploring efficiency techniques and alternative data sources to address this challenge. Overall, the panelists agreed that LLMs have already demonstrated significant value and that enterprises will likely start deploying them widely within the next two years. They also predicted that within three to four years, it will become clear whether artificial general intelligence (AGI) is possible with current technology.<\/p>\n<\/details>\n\n\n\n

B\u1ea3n d\u1ecbch Anh – Vi\u1ec7t<\/summary>\n

Gi\u00e1m \u0111\u1ed1c k\u1ef9 thu\u1eadt c\u1ee7a Meta v\u1ec1 AI t\u1ea1o sinh, Sergey Edunov, c\u00f3 m\u1ed9t c\u00e2u tr\u1ea3 l\u1eddi \u0111\u00e1ng ng\u1ea1c nhi\u00ean v\u1ec1 vi\u1ec7c c\u1ea7n bao nhi\u00eau c\u00f4ng su\u1ea5t h\u01a1n \u0111\u1ec3 x\u1eed l\u00fd nhu c\u1ea7u ng\u00e0y c\u00e0ng t\u0103ng v\u1ec1 \u1ee9ng d\u1ee5ng AI trong n\u0103m t\u1edbi: ch\u1ec9 c\u1ea7n hai nh\u00e0 m\u00e1y \u0111i\u1ec7n h\u1ea1t nh\u00e2n m\u1edbi. Edunov \u0111ang d\u1eabn \u0111\u1ea7u c\u00e1c n\u1ed7 l\u1ef1c \u0111\u00e0o t\u1ea1o c\u1ee7a Meta cho m\u00f4 h\u00ecnh c\u01a1 s\u1edf ngu\u1ed3n m\u1edf Llama 2, \u0111\u01b0\u1ee3c coi l\u00e0 m\u1ed9t trong nh\u1eefng m\u00f4 h\u00ecnh h\u00e0ng \u0111\u1ea7u. Trong m\u1ed9t bu\u1ed5i phi\u00ean th\u1ea3o m\u00e0 t\u00f4i \u0111\u00e3 \u0111i\u1ec1u ph\u1ed1i t\u1ea1i Di\u1ec5n \u0111\u00e0n C\u00f4ng nh\u00e2n K\u1ef9 thu\u1eadt s\u1ed1 tu\u1ea7n tr\u01b0\u1edbc \u1edf Thung l\u0169ng Silicon, \u00f4ng n\u00f3i r\u1eb1ng hai nh\u00e0 m\u00e1y \u0111i\u1ec7n s\u1ebd c\u00f3 v\u1ebb \u0111\u1ee7 \u0111\u1ec3 cung c\u1ea5p \u0111\u1ee7 n\u0103ng l\u01b0\u1ee3ng cho nhu c\u1ea7u AI c\u1ee7a nh\u00e2n lo\u1ea1i trong m\u1ed9t n\u0103m, v\u00e0 \u0111i\u1ec1u n\u00e0y c\u00f3 v\u1ebb ch\u1ea5p nh\u1eadn \u0111\u01b0\u1ee3c. \u0110\u1ec1 c\u1eadp \u0111\u1ebfn c\u00e2u h\u1ecfi v\u1ec1 vi\u1ec7c th\u1ebf gi\u1edbi c\u00f3 \u0111\u1ee7 kh\u1ea3 n\u0103ng \u0111\u1ec3 x\u1eed l\u00fd nhu c\u1ea7u ngu\u1ed3n \u0111i\u1ec7n gia t\u0103ng c\u1ee7a AI, \u0111\u1eb7c bi\u1ec7t l\u00e0 do s\u1ef1 gia t\u0103ng c\u1ee7a c\u00e1c \u1ee9ng d\u1ee5ng AI t\u1ea1o sinh \u0111\u00f2i h\u1ecfi nhi\u1ec1u n\u0103ng l\u01b0\u1ee3ng, \u00f4ng n\u00f3i r\u1eb1ng: “Ch\u00fang t\u00f4i ch\u1eafc ch\u1eafn c\u00f3 th\u1ec3 gi\u1ea3i quy\u1ebft v\u1ea5n \u0111\u1ec1 n\u00e0y.”<\/p>\n\n\n\n

Edunov cho bi\u1ebft r\u00f5 r\u1eb1ng \u00f4ng ch\u1ec9 th\u1ef1c hi\u1ec7n ph\u00e9p t\u00ednh \u0111\u01a1n gi\u1ea3n d\u1ef1a tr\u00ean s\u01a1 \u0111\u1ed3 gi\u1ea5y khi chu\u1ea9n b\u1ecb c\u00e2u tr\u1ea3 l\u1eddi c\u1ee7a m\u00ecnh. Tuy nhi\u00ean, \u00f4ng n\u00f3i r\u1eb1ng n\u00f3 cung c\u1ea5p m\u1ed9t \u01b0\u1edbc t\u00ednh kho\u1ea3ng c\u1ee7a c\u00f4ng su\u1ea5t c\u1ea7n thi\u1ebft \u0111\u1ec3 th\u1ef1c hi\u1ec7n nh\u1eefng g\u00ec \u0111\u01b0\u1ee3c g\u1ecdi l\u00e0 “inferencing” AI. Inferencing l\u00e0 qu\u00e1 tr\u00ecnh m\u00e0 AI \u0111\u01b0\u1ee3c tri\u1ec3n khai trong m\u1ed9t \u1ee9ng d\u1ee5ng \u0111\u1ec3 ph\u1ea3n h\u1ed3i m\u1ed9t c\u00e2u h\u1ecfi ho\u1eb7c \u0111\u01b0a ra m\u1ed9t \u0111\u1ec1 xu\u1ea5t.<\/p>\n\n\n\n

Inferencing kh\u00e1c bi\u1ec7t v\u1edbi vi\u1ec7c “training” m\u00f4 h\u00ecnh AI, trong \u0111\u00f3 m\u1ed9t m\u00f4 h\u00ecnh \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u00ean l\u01b0\u1ee3ng d\u1eef li\u1ec7u l\u1edbn \u0111\u1ec3 s\u1eb5n s\u00e0ng th\u1ef1c hi\u1ec7n inferencing.<\/p>\n\n\n\n

Vi\u1ec7c hu\u1ea5n luy\u1ec7n c\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLMs) \u0111\u00e3 nh\u1eadn \u0111\u01b0\u1ee3c s\u1ef1 quan t\u00e2m g\u1ea7n \u0111\u00e2y, v\u00ec n\u00f3 \u0111\u00f2i h\u1ecfi x\u1eed l\u00fd l\u1edbn, tuy ch\u1ec9 ban \u0111\u1ea7u. Khi m\u1ed9t m\u00f4 h\u00ecnh \u0111\u00e3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n, n\u00f3 c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng l\u1eb7p \u0111i l\u1eb7p l\u1ea1i cho c\u00e1c nhu c\u1ea7u inferencing, \u0111\u00f3 l\u00e0 n\u01a1i m\u00e0 \u1ee9ng d\u1ee5ng th\u1ef1c s\u1ef1 c\u1ee7a AI di\u1ec5n ra.<\/p>\n\n\n\n

Nhu c\u1ea7u v\u1ec1 c\u00f4ng su\u1ea5t cho inferencing \u0111\u01b0\u1ee3c ki\u1ec3m so\u00e1t<\/p>\n\n\n\n

Edunov \u0111\u01b0a ra hai c\u00e2u tr\u1ea3 l\u1eddi ri\u00eang bi\u1ec7t \u0111\u1ec3 gi\u1ea3i quy\u1ebft inferencing v\u00e0 training. C\u00e2u tr\u1ea3 l\u1eddi \u0111\u1ea7u ti\u00ean c\u1ee7a \u00f4ng \u0111\u1ec1 c\u1eadp \u0111\u1ebfn inferencing, n\u01a1i ph\u1ea7n l\u1edbn qu\u00e1 tr\u00ecnh x\u1eed l\u00fd s\u1ebd di\u1ec5n ra khi c\u00e1c t\u1ed5 ch\u1ee9c tri\u1ec3n khai c\u00e1c \u1ee9ng d\u1ee5ng AI. \u00d4ng gi\u1ea3i th\u00edch c\u00e1ch \u00f4ng th\u1ef1c hi\u1ec7n t\u00ednh to\u00e1n \u0111\u01a1n gi\u1ea3n cho ph\u00eda inferencing: \u00d4ng n\u00f3i r\u1eb1ng Nvidia, nh\u00e0 cung c\u1ea5p ch\u1ee7 \u0111\u1ea1o c\u1ee7a b\u1ed9 x\u1eed l\u00fd cho AI, c\u00f3 v\u1ebb \u0111\u00e3 s\u1eb5n s\u00e0ng ra m\u1eaft t\u1eeb m\u1ed9t tri\u1ec7u \u0111\u1ebfn hai tri\u1ec7u GPU H100 c\u1ee7a m\u00ecnh v\u00e0o n\u0103m t\u1edbi. N\u1ebfu t\u1ea5t c\u1ea3 s\u1ed1 GPU \u0111\u00f3 \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng \u0111\u1ec3 t\u1ea1o ra “token” cho c\u00e1c LLM c\u00f3 k\u00edch th\u01b0\u1edbc h\u1ee3p l\u00fd, \u00f4ng n\u00f3i r\u1eb1ng n\u00f3 t\u01b0\u01a1ng \u0111\u01b0\u01a1ng v\u1edbi kho\u1ea3ng 100.000 token cho m\u1ed7i ng\u01b0\u1eddi tr\u00ean h\u00e0nh tinh m\u1ed7i ng\u00e0y, \u00f4ng th\u1eeba nh\u1eadn r\u1eb1ng \u0111\u00f3 l\u00e0 m\u1ed9t s\u1ed1 l\u01b0\u1ee3ng kh\u00e1 l\u1edbn.<\/p>\n\n\n\n

Token l\u00e0 c\u00e1c \u0111\u01a1n v\u1ecb c\u01a1 b\u1ea3n c\u1ee7a v\u0103n b\u1ea3n m\u00e0 LLMs s\u1eed d\u1ee5ng \u0111\u1ec3 x\u1eed l\u00fd v\u00e0 t\u1ea1o ra ng\u00f4n ng\u1eef. Ch\u00fang c\u00f3 th\u1ec3 l\u00e0 t\u1eeb, c\u00e1c ph\u1ea7n c\u1ee7a t\u1eeb ho\u1eb7c th\u1eadm ch\u00ed l\u00e0 c\u00e1c k\u00fd t\u1ef1 \u0111\u01a1n l\u1ebb, t\u00f9y thu\u1ed9c v\u00e0o c\u00e1ch m\u00e0 LLM \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf. V\u00ed d\u1ee5, t\u1eeb “xin ch\u00e0o” c\u00f3 th\u1ec3 l\u00e0 m\u1ed9t token duy nh\u1ea5t, ho\u1eb7c n\u00f3 c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c chia th\u00e0nh hai token: “xin” v\u00e0 “ch\u00e0o”. C\u00e0ng nhi\u1ec1u token m\u00e0 LLM c\u00f3 th\u1ec3 x\u1eed l\u00fd, ng\u00f4n ng\u1eef m\u00e0 n\u00f3 c\u00f3 th\u1ec3 t\u1ea1o ra c\u00e0ng ph\u1ee9c t\u1ea1p v\u00e0 \u0111a d\u1ea1ng h\u01a1n.<\/p>\n\n\n\n

V\u1eady ch\u00fang ta c\u1ea7n bao nhi\u00eau \u0111i\u1ec7n \u0111\u1ec3 t\u1ea1o ra nhi\u1ec1u token nh\u01b0 v\u1eady? Nh\u01b0 v\u1eady, m\u1ed7i GPU H100 c\u1ea7n kho\u1ea3ng 700 watt, v\u00e0 v\u1edbi vi\u1ec7c b\u1ea1n c\u1ea7n m\u1ed9t s\u1ed1 \u0111i\u1ec7n \u0111\u1ec3 h\u1ed7 tr\u1ee3 trung t\u00e2m d\u1eef li\u1ec7u v\u00e0 l\u00e0m m\u00e1t, Edunov n\u00f3i r\u1eb1ng \u00f4ng l\u00e0m tr\u00f2n l\u00ean 1KW cho m\u1ed7i GPU. T\u1ed5ng c\u1ed9ng l\u1ea1i, ch\u1ec9 c\u1ea7n hai nh\u00e0 m\u00e1y \u0111i\u1ec7n h\u1ea1t nh\u00e2n \u0111\u1ec3 cung c\u1ea5p \u0111\u1ee7 n\u0103ng l\u01b0\u1ee3ng cho t\u1ea5t c\u1ea3 c\u00e1c H100 \u0111\u00f3. “V\u1edbi quy m\u00f4 c\u1ee7a nh\u00e2n lo\u1ea1i, kh\u00f4ng ph\u1ea3i l\u00e0 qu\u00e1 nhi\u1ec1u,” Edunov n\u00f3i. “T\u00f4i ngh\u0129 nh\u01b0 m\u1ed9t x\u00e3 h\u1ed9i, nh\u00e2n lo\u1ea1i c\u00f3 th\u1ec3 chi tr\u1ea3 cho vi\u1ec7c s\u1eed d\u1ee5ng t\u1ed1i \u0111a 100.000 token m\u1ed7i ng\u00e0y cho m\u1ed7i ng\u01b0\u1eddi tr\u00ean h\u00e0nh tinh n\u00e0y. V\u00ec v\u1eady, v\u1ec1 ph\u00eda inferencing, t\u00f4i ngh\u0129 nh\u01b0 hi\u1ec7n t\u1ea1i ch\u00fang ta c\u00f3 th\u1ec3 \u1ed5n.”<\/p>\n\n\n\n

Sau bu\u1ed5i h\u1ed9i th\u1ea3o, Edunov \u0111\u00e3 l\u00e0m r\u00f5 v\u1edbi VentureBeat r\u1eb1ng \u00fd ki\u1ebfn c\u1ee7a \u00f4ng li\u00ean quan \u0111\u1ebfn n\u0103ng l\u01b0\u1ee3ng c\u1ea7n thi\u1ebft cho s\u1ef1 t\u00ednh to\u00e1n AI b\u1ed5 sung t\u1eeb s\u1ef1 gia t\u0103ng m\u1edbi c\u1ee7a Nvidia H100, \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1eb7c bi\u1ec7t \u0111\u1ec3 x\u1eed l\u00fd c\u00e1c \u1ee9ng d\u1ee5ng AI v\u00e0 do \u0111\u00f3 l\u00e0 m\u1ed9t trong nh\u1eefng c\u00f4ng ngh\u1ec7 \u0111\u00e1ng ch\u00fa \u00fd nh\u1ea5t. Ngo\u00e0i c\u00e1c m\u1eabu GPU H100, c\u00f2n c\u00f3 c\u00e1c m\u00f4 h\u00ecnh GPU Nvidia c\u0169 h\u01a1n, c\u0169ng nh\u01b0 CPU AMD v\u00e0 Intel, c\u0169ng nh\u01b0 c\u00e1c gia t\u1ed1c vi\u00ean AI chuy\u00ean d\u1ee5ng \u0111\u1ec3 th\u1ef1c hi\u1ec7n inferencing cho AI.<\/p>\n\n\n\n

\u0110\u1ed1i v\u1edbi vi\u1ec7c hu\u1ea5n luy\u1ec7n generative AI, v\u1ea5n \u0111\u1ec1 l\u1edbn nh\u1ea5t l\u00e0 vi\u1ec7c c\u00f3 \u0111\u1ee7 d\u1eef li\u1ec7u \u0111\u1ec3 hu\u1ea5n luy\u1ec7n ch\u00fang. Edunov cho bi\u1ebft r\u1eb1ng c\u00f3 nhi\u1ec1u suy \u0111o\u00e1n r\u1ed9ng r\u00e3i r\u1eb1ng GPT4 \u0111\u00e3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u00ean to\u00e0n b\u1ed9 internet. \u00d4ng \u0111\u00e3 \u0111\u01b0a ra m\u1ed9t s\u1ed1 gi\u1ea3 \u0111\u1ecbnh \u0111\u01a1n gi\u1ea3n kh\u00e1c. \u00d4ng n\u00f3i r\u1eb1ng to\u00e0n b\u1ed9 internet c\u00f4ng khai c\u00f3 kho\u1ea3ng 100 ngh\u00ecn t\u1ef7 token, n\u1ebfu b\u1ea1n ch\u1ec9 t\u1ea3i xu\u1ed1ng n\u00f3, \u00f4ng n\u00f3i r\u1eb1ng b\u1ea1n c\u00f3 th\u1ec3 gi\u1ea3m d\u1eef li\u1ec7u \u0111\u00f3 xu\u1ed1ng c\u00f2n 20 t\u1ef7 \u0111\u1ebfn 10 t\u1ef7 token sau khi l\u00e0m s\u1ea1ch v\u00e0 lo\u1ea1i b\u1ecf c\u00e1c d\u1eef li\u1ec7u tr\u00f9ng l\u1eb7p. V\u00e0 n\u1ebfu b\u1ea1n t\u1eadp trung v\u00e0o c\u00e1c token ch\u1ea5t l\u01b0\u1ee3ng cao, s\u1ed1 l\u01b0\u1ee3ng token s\u1ebd c\u00f2n \u00edt h\u01a1n. “S\u1ed1 l\u01b0\u1ee3ng ki\u1ebfn th\u1ee9c tinh luy\u1ec7n m\u00e0 nh\u00e2n lo\u1ea1i \u0111\u00e3 t\u1ea1o ra qua c\u00e1c th\u1ebf k\u1ef7 kh\u00f4ng l\u1edbn l\u1eafm,” \u00f4ng n\u00f3i, \u0111\u1eb7c bi\u1ec7t l\u00e0 n\u1ebfu b\u1ea1n c\u1ea7n ti\u1ebfp t\u1ee5c th\u00eam d\u1eef li\u1ec7u v\u00e0o c\u00e1c m\u00f4 h\u00ecnh \u0111\u1ec3 m\u1edf r\u1ed9ng ch\u00fang \u0111\u1ea1t hi\u1ec7u su\u1ea5t t\u1ed1t h\u01a1n.<\/p>\n\n\n\n

\u00d4ng \u01b0\u1edbc t\u00ednh r\u1eb1ng c\u00e1c m\u00f4 h\u00ecnh ti\u1ebfp theo v\u1edbi hi\u1ec7u su\u1ea5t cao h\u01a1n s\u1ebd y\u00eau c\u1ea7u nhi\u1ec1u h\u01a1n g\u1ea5p 10 l\u1ea7n d\u1eef li\u1ec7u. V\u00ec v\u1eady, n\u1ebfu GPT4 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u00ean kho\u1ea3ng 20 ngh\u00ecn t\u1ef7 token, th\u00ec m\u00f4 h\u00ecnh ti\u1ebfp theo s\u1ebd y\u00eau c\u1ea7u kho\u1ea3ng 200 ngh\u00ecn t\u1ef7 token. \u00d4ng cho bi\u1ebft c\u00f3 th\u1ec3 kh\u00f4ng c\u00f3 \u0111\u1ee7 d\u1eef li\u1ec7u c\u00f4ng c\u1ed9ng \u0111\u1ec3 l\u00e0m \u0111i\u1ec1u \u0111\u00f3. \u0110\u00f3 l\u00e0 l\u00fd do t\u1ea1i sao c\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u \u0111ang l\u00e0m vi\u1ec7c v\u1ec1 c\u00e1c k\u1ef9 thu\u1eadt hi\u1ec7u qu\u1ea3 \u0111\u1ec3 l\u00e0m cho m\u00f4 h\u00ecnh tr\u1edf n\u00ean hi\u1ec7u qu\u1ea3 v\u00e0 th\u00f4ng minh h\u01a1n tr\u00ean l\u01b0\u1ee3ng d\u1eef li\u1ec7u nh\u1ecf h\u01a1n. C\u00e1c m\u00f4 h\u00ecnh LLM c\u0169ng c\u00f3 th\u1ec3 ph\u1ea3i s\u1eed d\u1ee5ng c\u00e1c ngu\u1ed3n d\u1eef li\u1ec7u thay th\u1ebf, v\u00ed d\u1ee5 nh\u01b0 d\u1eef li\u1ec7u \u0111a ph\u01b0\u01a1ng th\u1ee9c nh\u01b0 video. “\u0110\u00f3 l\u00e0 m\u1ed9t l\u01b0\u1ee3ng d\u1eef li\u1ec7u r\u1ea5t l\u1edbn c\u00f3 th\u1ec3 t\u1ea1o \u0111i\u1ec1u ki\u1ec7n cho s\u1ef1 m\u1edf r\u1ed9ng trong t\u01b0\u01a1ng lai,” \u00f4ng n\u00f3i.<\/p>\n\n\n\n

Edunov \u0111\u00e3 n\u00f3i trong m\u1ed9t bu\u1ed5i th\u1ea3o lu\u1eadn mang t\u1ef1a \u0111\u1ec1: “T\u1ea1o ra Token: \u0110i\u1ec7n n\u0103ng c\u1ee7a th\u1eddi \u0111\u1ea1i GenAI,” v\u00e0 \u00f4ng \u0111\u00e3 tham gia c\u00f9ng v\u1edbi Nik Spirin, gi\u00e1m \u0111\u1ed1c GenAI c\u1ee7a Nvidia, v\u00e0 Kevin Tsai, Tr\u01b0\u1edfng ki\u1ebfn tr\u00fac gi\u1ea3i ph\u00e1p, GenAI, c\u1ee7a Google.<\/p>\n\n\n\n

Spirin \u0111\u1ed3ng \u00fd v\u1edbi Edunov r\u1eb1ng c\u00f3 c\u00e1c ngu\u1ed3n d\u1eef li\u1ec7u kh\u00e1c n\u1eb1m ngo\u00e0i internet c\u00f4ng c\u1ed9ng, bao g\u1ed3m sau t\u01b0\u1eddng l\u1eeda v\u00e0 di\u1ec5n \u0111\u00e0n, m\u1eb7c d\u00f9 ch\u00fang kh\u00f4ng d\u1ec5 d\u00e0ng truy c\u1eadp. Tuy nhi\u00ean, c\u00e1c t\u1ed5 ch\u1ee9c c\u00f3 quy\u1ec1n truy c\u1eadp v\u00e0o d\u1eef li\u1ec7u \u0111\u00f3 c\u00f3 th\u1ec3 s\u1eed d\u1ee5ng \u0111\u1ec3 t\u00f9y ch\u1ec9nh d\u1ec5 d\u00e0ng c\u00e1c m\u00f4 h\u00ecnh c\u01a1 b\u1ea3n.<\/p>\n\n\n\n

X\u00e3 h\u1ed9i quan t\u00e2m \u0111\u1ebfn vi\u1ec7c \u1ee7ng h\u1ed9 c\u00e1c m\u00f4 h\u00ecnh c\u01a1 b\u1ea3n m\u00e3 ngu\u1ed3n m\u1edf t\u1ed1t nh\u1ea5t, \u0111\u1ec3 tr\u00e1nh ph\u1ea3i h\u1ed7 tr\u1ee3 qu\u00e1 nhi\u1ec1u n\u1ed7 l\u1ef1c \u0111\u1ed9c l\u1eadp, Spirin n\u00f3i. \u0110i\u1ec1u n\u00e0y s\u1ebd ti\u1ebft ki\u1ec7m c\u00f4ng su\u1ea5t t\u00ednh to\u00e1n, v\u00ec ch\u00fang c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n tr\u01b0\u1edbc m\u1ed9t l\u1ea7n v\u00e0 h\u1ea7u h\u1ebft c\u00f4ng s\u1ee9c c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c d\u00e0nh cho vi\u1ec7c t\u1ea1o ra c\u00e1c \u1ee9ng d\u1ee5ng th\u00f4ng minh ph\u00eda d\u01b0\u1edbi. \u00d4ng n\u00f3i r\u1eb1ng \u0111\u00e2y l\u00e0 m\u1ed9t c\u00e1ch \u0111\u1ec3 tr\u00e1nh g\u1eb7p b\u1ea5t k\u1ef3 gi\u1edbi h\u1ea1n d\u1eef li\u1ec7u n\u00e0o trong th\u1eddi gian t\u1edbi.<\/p>\n\n\n\n

Tsai c\u1ee7a Google b\u1ed5 sung r\u1eb1ng m\u1ed9t s\u1ed1 c\u00f4ng ngh\u1ec7 kh\u00e1c c\u0169ng c\u00f3 th\u1ec3 gi\u00fap gi\u1ea3m \u00e1p l\u1ef1c \u0111\u00e0o t\u1ea1o. Vi\u1ec7c t\u0103ng c\u01b0\u1eddng sinh ra (RAG) c\u00f3 th\u1ec3 gi\u00fap c\u00e1c t\u1ed5 ch\u1ee9c \u0111i\u1ec1u ch\u1ec9nh c\u00e1c m\u00f4 h\u00ecnh c\u01a1 b\u1ea3n v\u1edbi c\u00e1c kho d\u1eef li\u1ec7u c\u1ee7a h\u1ecd. M\u1eb7c d\u00f9 RAG c\u00f3 nh\u1eefng gi\u1edbi h\u1ea1n c\u1ee7a n\u00f3, c\u00e1c c\u00f4ng ngh\u1ec7 kh\u00e1c m\u00e0 Google \u0111\u00e3 th\u1eed nghi\u1ec7m, ch\u1eb3ng h\u1ea1n nh\u01b0 vector ng\u1eef ngh\u0129a th\u01b0a th\u1edbt, c\u00f3 th\u1ec3 gi\u00fap. “C\u1ed9ng \u0111\u1ed3ng c\u00f3 th\u1ec3 \u0111\u1ed3ng h\u00e0nh v\u1edbi nhau v\u1edbi nh\u1eefng m\u00f4 h\u00ecnh h\u1eefu \u00edch c\u00f3 th\u1ec3 t\u00e1i s\u1eed d\u1ee5ng \u1edf nhi\u1ec1u n\u01a1i. V\u00e0 \u0111\u00f3 l\u00e0 c\u00e1ch ti\u1ebfp t\u1ee5c, \u0111\u00fang kh\u00f4ng, cho tr\u00e1i \u0111\u1ea5t,” \u00f4ng n\u00f3i.<\/p>\n\n\n\n

D\u1ef1 \u0111o\u00e1n: Trong ba ho\u1eb7c b\u1ed1n n\u0103m t\u1edbi, ch\u00fang ta s\u1ebd bi\u1ebft li\u1ec7u AGI c\u00f3 kh\u1ea3 thi v\u1edbi c\u00f4ng ngh\u1ec7 hi\u1ec7n t\u1ea1i hay kh\u00f4ng, v\u00e0 c\u00e1c m\u00f4 h\u00ecnh LLM s\u1ebd mang l\u1ea1i gi\u00e1 tr\u1ecb “to l\u1edbn” cho doanh nghi\u1ec7p.<\/p>\n\n\n\n

Cu\u1ed1i bu\u1ed5i th\u1ea3o lu\u1eadn, t\u00f4i \u0111\u00e3 h\u1ecfi c\u00e1c di\u1ec5n gi\u1ea3 v\u1ec1 d\u1ef1 \u0111o\u00e1n c\u1ee7a h\u1ecd v\u1ec1 hai \u0111\u1ebfn ba n\u0103m t\u1edbi v\u1ec1 c\u00e1ch LLMs s\u1ebd ph\u00e1t tri\u1ec3n trong kh\u1ea3 n\u0103ng, v\u00e0 n\u01a1i ch\u00fang s\u1ebd \u0111\u1ea1t \u0111\u1ebfn gi\u1edbi h\u1ea1n. N\u00f3i chung, h\u1ecd \u0111\u1ed3ng \u00fd r\u1eb1ng trong khi ch\u01b0a r\u00f5 r\u00e0ng LLMs c\u00f3 th\u1ec3 c\u1ea3i thi\u1ec7n \u0111\u1ebfn \u0111\u00e2u, \u0111\u00e3 c\u00f3 \u0111\u01b0\u1ee3c gi\u00e1 tr\u1ecb \u0111\u00e1ng k\u1ec3 v\u00e0 c\u00e1c doanh nghi\u1ec7p c\u00f3 th\u1ec3 tri\u1ec3n khai LLMs theo s\u1ed1 l\u01b0\u1ee3ng l\u1edbn trong kho\u1ea3ng hai n\u0103m t\u1edbi.<\/p>\n\n\n\n

Edunov c\u1ee7a Meta n\u00f3i r\u1eb1ng c\u1ea3i ti\u1ebfn cho LLMs c\u00f3 th\u1ec3 ti\u1ebfp t\u1ee5c theo h\u00e0m s\u1ed1 m\u0169 ho\u1eb7c b\u1eaft \u0111\u1ea7u gi\u1ea3m \u0111i, \u00f4ng d\u1ef1 \u0111o\u00e1n r\u1eb1ng ch\u00fang ta s\u1ebd c\u00f3 c\u00e2u tr\u1ea3 l\u1eddi trong ba ho\u1eb7c b\u1ed1n n\u0103m t\u1edbi xem tr\u00ed tu\u1ec7 t\u1ed5ng qu\u00e1t nh\u00e2n t\u1ea1o (AGI) c\u00f3 kh\u1ea3 thi v\u1edbi c\u00f4ng ngh\u1ec7 hi\u1ec7n t\u1ea1i hay kh\u00f4ng. Spirin c\u1ee7a Nvidia n\u00f3i r\u1eb1ng d\u1ef1a tr\u00ean c\u00e1c l\u00e0n s\u00f3ng c\u00f4ng ngh\u1ec7 tr\u01b0\u1edbc, bao g\u1ed3m c\u00f4ng ngh\u1ec7 AI ban \u0111\u1ea7u, c\u00e1c c\u00f4ng ty doanh nghi\u1ec7p s\u1ebd ch\u1eadm ch\u00e2n trong vi\u1ec7c \u00e1p d\u1ee5ng ban \u0111\u1ea7u. Nh\u01b0ng trong v\u00f2ng hai n\u0103m, \u00f4ng mong \u0111\u1ee3i c\u00e1c c\u00f4ng ty s\u1ebd nh\u1eadn \u0111\u01b0\u1ee3c gi\u00e1 tr\u1ecb “to l\u1edbn” t\u1eeb \u0111\u00f3. “\u00cdt nh\u1ea5t l\u00e0 tr\u01b0\u1eddng h\u1ee3p v\u1edbi l\u00e0n s\u00f3ng c\u00f4ng ngh\u1ec7 AI tr\u01b0\u1edbc,” \u00f4ng n\u00f3i.<\/p>\n\n\n\n

Tsai c\u1ee7a Google ch\u1ec9 ra r\u1eb1ng gi\u1edbi h\u1ea1n chu\u1ed7i cung \u1ee9ng – do s\u1ef1 ph\u1ee5 thu\u1ed9c c\u1ee7a Nvidia v\u00e0o b\u1ed9 nh\u1edb b\u0103ng th\u00f4ng cao cho GPU c\u1ee7a m\u00ecnh – \u0111ang l\u00e0m ch\u1eadm qu\u00e1 tr\u00ecnh c\u1ea3i ti\u1ebfn m\u00f4 h\u00ecnh v\u00e0 r\u1eb1ng n\u00fat th\u1eaft n\u00e0y ph\u1ea3i \u0111\u01b0\u1ee3c gi\u1ea3i quy\u1ebft. Nh\u01b0ng \u00f4ng n\u00f3i r\u1eb1ng \u00f4ng v\u1eabn c\u1ea3m th\u1ea5y kh\u00edch l\u1ec7 b\u1edfi c\u00e1c \u0111\u1ed5i m\u1edbi, nh\u01b0 Blib-2, m\u1ed9t d\u1ef1 \u00e1n nghi\u00ean c\u1ee9u t\u1eeb Salesforce, \u0111\u1ec3 t\u00ecm c\u00e1ch x\u00e2y d\u1ef1ng c\u00e1c m\u00f4 h\u00ecnh nh\u1ecf h\u01a1n, hi\u1ec7u qu\u1ea3 h\u01a1n. Nh\u1eefng m\u00f4 h\u00ecnh n\u00e0y c\u00f3 th\u1ec3 gi\u00fap LLMs v\u01b0\u1ee3t qua c\u00e1c r\u00e0ng bu\u1ed9c chu\u1ed7i cung \u1ee9ng b\u1eb1ng c\u00e1ch gi\u1ea3m y\u00eau c\u1ea7u x\u1eed l\u00fd c\u1ee7a ch\u00fang, \u00f4ng n\u00f3i.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":"

Gi\u00e1m \u0111\u1ed1c k\u1ef9 thu\u1eadt c\u1ee7a Meta v\u1ec1 AI t\u1ea1o sinh, Sergey Edunov, tin r\u1eb1ng ch\u1ec9 c\u1ea7n hai nh\u00e0 m\u00e1y \u0111i\u1ec7n h\u1ea1t nh\u00e2n m\u1edbi s\u1ebd \u0111\u1ee7 \u0111\u1ec3 \u0111\u00e1p \u1ee9ng nhu c\u1ea7u ng\u00e0y c\u00e0ng t\u0103ng v\u1ec1 \u1ee9ng d\u1ee5ng tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o trong n\u0103m t\u1edbi.<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[21,49,10],"class_list":["post-167","post","type-post","status-publish","format-standard","hentry","category-ai-news","tag-chatgpt","tag-mo-hinh-ngon-ngu-lon-llm","tag-tri-tue-nhan-tao"],"_links":{"self":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/167","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=167"}],"version-history":[{"count":25,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/167\/revisions"}],"predecessor-version":[{"id":439,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/167\/revisions\/439"}],"wp:attachment":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=167"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=167"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=167"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}