{"id":552,"date":"2023-11-24T11:18:03","date_gmt":"2023-11-24T04:18:03","guid":{"rendered":"https:\/\/xuhuongai.com\/?p=552"},"modified":"2023-11-24T13:29:08","modified_gmt":"2023-11-24T06:29:08","slug":"khi-noi-ve-mo-hinh-ai-lon-hon-khong-phai-luc-nao-cung-tot-hon","status":"publish","type":"post","link":"https:\/\/xuhuongai.com\/?p=552","title":{"rendered":"Khi n\u00f3i v\u1ec1 M\u00f4 h\u00ecnh AI, l\u1edbn h\u01a1n kh\u00f4ng ph\u1ea3i l\u00fac n\u00e0o c\u0169ng t\u1ed1t h\u01a1n"},"content":{"rendered":"\n
\n

B\u1ea3n tin \u0111\u01b0\u1ee3c t\u00f3m t\u1eaft Anh-Vi\u1ec7t b\u1edfi n\u1ec1n t\u1ea3ng t\u1ea1o tr\u1ee3 l\u00fd AI \u2013 KamiMind<\/a>.<\/p>\nNgu\u1ed3n: Lauren Leffer, “When It Comes to AI Models, Bigger Isn\u2019t Always Better<\/a>“, Scientific American, 21\/11\/2023.<\/cite><\/blockquote>\n\n\n\n

C\u00e1c m\u00f4 h\u00ecnh tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o (AI) \u0111ang ng\u00e0y c\u00e0ng tr\u1edf n\u00ean l\u1edbn h\u01a1n, \u0111i k\u00e8m v\u1edbi \u0111\u00f3 l\u00e0 c\u00e1c b\u1ed9 d\u1eef li\u1ec7u \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng \u0111\u1ec3 hu\u1ea5n luy\u1ec7n ch\u00fang. Tuy nhi\u00ean, vi\u1ec7c thu nh\u1ecf c\u00f3 th\u1ec3 gi\u1ea3i quy\u1ebft m\u1ed9t s\u1ed1 v\u1ea5n \u0111\u1ec1 l\u1edbn c\u1ee7a AI. C\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (LLMs) t\u1ea1o ra c\u00e1c chatbot n\u1ed5i ti\u1ebfng nh\u01b0 ChatGPT c\u1ee7a OpenAI v\u00e0 Bard c\u1ee7a Google, \u0111\u01b0\u1ee3c t\u1ea1o th\u00e0nh t\u1eeb h\u01a1n 100 t\u1ef7 tham s\u1ed1. Tuy nhi\u00ean, khi c\u00e1c m\u00f4 h\u00ecnh tr\u1edf n\u00ean l\u1edbn h\u01a1n, ch\u00fang c\u0169ng tr\u1edf n\u00ean kh\u00f3 ki\u1ec3m so\u00e1t, ti\u00eau t\u1ed1n n\u0103ng l\u01b0\u1ee3ng v\u00e0 kh\u00f3 x\u00e2y d\u1ef1ng h\u01a1n. C\u00e1c m\u00f4 h\u00ecnh v\u00e0 b\u1ed9 d\u1eef li\u1ec7u nh\u1ecf h\u01a1n c\u00f3 th\u1ec3 gi\u00fap gi\u1ea3i quy\u1ebft v\u1ea5n \u0111\u1ec1 n\u00e0y.<\/p>\n\n\n\n

V\u00ed d\u1ee5, v\u00e0o th\u00e1ng 9 v\u1eeba qua, m\u1ed9t nh\u00f3m c\u00e1c nh\u00e0 nghi\u00ean c\u1ee9u c\u1ee7a Microsoft \u0111\u00e3 c\u00f4ng b\u1ed1 m\u1ed9t b\u00e1o c\u00e1o k\u1ef9 thu\u1eadt v\u1ec1 m\u1ed9t m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef m\u1edbi c\u00f3 t\u00ean l\u00e0 phi-1.5. Phi-1.5 bao g\u1ed3m 1.3 t\u1ef7 tham s\u1ed1, ch\u1ec9 b\u1eb1ng m\u1ed9t ph\u1ea7n tr\u0103m k\u00edch th\u01b0\u1edbc c\u1ee7a GPT-3.5, m\u00f4 h\u00ecnh n\u1eb1m d\u01b0\u1edbi phi\u00ean b\u1ea3n mi\u1ec5n ph\u00ed c\u1ee7a ChatGPT. Nh\u01b0ng d\u00f9 k\u00edch th\u01b0\u1edbc t\u01b0\u01a1ng \u0111\u1ed1i nh\u1ecf, phi-1.5 v\u1eabn “th\u1ec3 hi\u1ec7n nhi\u1ec1u \u0111\u1eb7c \u0111i\u1ec3m c\u1ee7a c\u00e1c LLMs l\u1edbn h\u01a1n nhi\u1ec1u,” theo nh\u01b0 t\u00e1c gi\u1ea3 vi\u1ebft trong b\u00e1o c\u00e1o c\u1ee7a h\u1ecd.<\/p>\n\n\n\n

Hu\u1ea5n luy\u1ec7n v\u00e0 ch\u1ea1y m\u1ed9t m\u00f4 h\u00ecnh AI v\u1edbi h\u01a1n 100 t\u1ef7 tham s\u1ed1 t\u1ed1n r\u1ea5t nhi\u1ec1u n\u0103ng l\u01b0\u1ee3ng. M\u1ed9t ng\u00e0y ti\u00eau chu\u1ea9n c\u1ee7a vi\u1ec7c s\u1eed d\u1ee5ng ChatGPT tr\u00ean to\u00e0n c\u1ea7u c\u00f3 th\u1ec3 ti\u00eau t\u1ed1n \u0111i\u1ec7n n\u0103ng t\u01b0\u01a1ng \u0111\u01b0\u01a1ng v\u1edbi kho\u1ea3ng 33,000 h\u1ed9 gia \u0111\u00ecnh \u1edf M\u1ef9, theo \u01b0\u1edbc t\u00ednh c\u1ee7a k\u1ef9 s\u01b0 m\u00e1y t\u00ednh Sajjad Moazeni t\u1eeb \u0110\u1ea1i h\u1ecdc Washington. AI nh\u1ecf h\u01a1n c\u1ea7n \u00edt c\u00f4ng su\u1ea5t t\u00ednh to\u00e1n v\u00e0 n\u0103ng l\u01b0\u1ee3ng \u0111\u1ec3 ch\u1ea1y, theo Matthew Stewart, m\u1ed9t k\u1ef9 s\u01b0 m\u00e1y t\u00ednh t\u1ea1i \u0110\u1ea1i h\u1ecdc Harvard. \u0110i\u1ec1u n\u00e0y gi\u00fap c\u1ea3i thi\u1ec7n t\u00ednh b\u1ec1n v\u1eefng.<\/p>\n\n\n\n

H\u01a1n n\u1eefa, AI \u00edt t\u1ed1n t\u00e0i nguy\u00ean h\u01a1n c\u0169ng d\u1ec5 ti\u1ebfp c\u1eadn h\u01a1n. Hi\u1ec7n t\u1ea1i, ch\u1ec9 c\u00f3 m\u1ed9t s\u1ed1 \u00edt c\u00f4ng ty t\u01b0 nh\u00e2n c\u00f3 \u0111\u1ee7 kinh ph\u00ed v\u00e0 kh\u00f4ng gian m\u00e1y ch\u1ee7 \u0111\u1ec3 x\u00e2y d\u1ef1ng, l\u01b0u tr\u1eef, hu\u1ea5n luy\u1ec7n v\u00e0 ch\u1ec9nh s\u1eeda c\u00e1c LLMs l\u1edbn nh\u1ea5t. C\u00e1c m\u00f4 h\u00ecnh nh\u1ecf h\u01a1n c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c ph\u00e1t tri\u1ec3n v\u00e0 nghi\u00ean c\u1ee9u b\u1edfi nhi\u1ec1u ng\u01b0\u1eddi h\u01a1n. Vi\u1ec7c ngh\u0129 nh\u1ecf “c\u00f3 th\u1ec3 d\u00e2n ch\u1ee7 h\u00f3a AI,” theo Eva Portelance, m\u1ed9t nh\u00e0 nghi\u00ean c\u1ee9u ng\u00f4n ng\u1eef h\u1ecdc t\u00ednh to\u00e1n v\u00e0 nh\u1eadn th\u1ee9c t\u1ea1i Vi\u1ec7n Tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o Mila-Quebec.<\/p>\n\n\n\n

Cu\u1ed1i c\u00f9ng, c\u00f3 m\u1ed9t v\u1ea5n \u0111\u1ec1 c\u01a1 b\u1ea3n h\u01a1n v\u1ec1 kh\u1ea3 n\u0103ng gi\u1ea3i th\u00edch: m\u1ee9c \u0111\u1ed9 m\u00e0 m\u1ed9t m\u00f4 h\u00ecnh h\u1ecdc m\u00e1y c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c hi\u1ec3u b\u1edfi nh\u00e0 ph\u00e1t tri\u1ec3n c\u1ee7a n\u00f3. \u0110\u1ed1i v\u1edbi c\u00e1c m\u00f4 h\u00ecnh AI l\u1edbn h\u01a1n, vi\u1ec7c ph\u00e2n t\u00edch vai tr\u00f2 c\u1ee7a t\u1eebng tham s\u1ed1 g\u1ea7n nh\u01b0 kh\u00f4ng th\u1ec3. Trong c\u00e1c m\u00f4 h\u00ecnh nh\u1ecf h\u01a1n, vi\u1ec7c x\u00e1c \u0111\u1ecbnh nguy\u00ean nh\u00e2n v\u00e0 k\u1ebft qu\u1ea3 v\u00e0 \u0111i\u1ec1u ch\u1ec9nh theo \u0111\u00f3 d\u1ec5 d\u00e0ng h\u01a1n, m\u1eb7c d\u00f9 th\u01b0\u1eddng v\u1eabn kh\u00f3 kh\u0103n. “T\u00f4i th\u00edch c\u1ed1 g\u1eafng hi\u1ec3u m\u1ed9t tri\u1ec7u tham s\u1ed1 h\u01a1n l\u00e0 m\u1ed9t t\u1ef7 tham s\u1ed1,” Brenden Lake, m\u1ed9t nh\u00e0 khoa h\u1ecdc v\u1ec1 t\u00ednh to\u00e1n nh\u1eadn th\u1ee9c nghi\u00ean c\u1ee9u l\u0129nh v\u1ef1c tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o t\u1ea1i \u0110\u1ea1i h\u1ecdc New York, n\u00f3i.<\/p>\n\n\n\n

V\u00ec v\u1eady, \u0111\u00f4i khi nh\u1ecf h\u01a1n c\u00f3 th\u1ec3 th\u00f4ng minh h\u01a1n.<\/p>\n\n\n\n

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

Artificial Intelligence (AI) models are experiencing a paradigm shift, moving away from sheer size towards more compact, yet equally potent models. The large language models (LLMs) such as OpenAI’s ChatGPT and Google’s Bard, which consist of over 100 billion parameters, are becoming increasingly cumbersome, energy-intensive, and challenging to manage. Thus, the tech industry is exploring the potential of smaller AI models, as exemplified by Microsoft’s new language model, phi-1.5, which comprises 1.3 billion parameters, yet demonstrates capabilities akin to much larger models.<\/p>\n\n\n\n

The benefits of smaller AI models are evident in their lower energy requirements and increased accessibility. Current mega models are not only power-hungry but also restricted to a few big tech companies with the necessary resources. Conversely, smaller models democratize AI, enabling more people to engage in their development and use. Moreover, they can fit into smaller devices, providing enhanced functionality without the need for cloud-based operations, thereby enhancing data security.<\/p>\n\n\n\n

Furthermore, smaller AI models have the advantage of interpretability \u2013 the ability for developers to understand the role of each parameter, which is virtually impossible in larger models. This interpretability can lead to insights into human learning that can be replicated in AI, creating more “cognitively plausible” models.<\/p>\n\n\n\n

However, it’s important to note that while smaller models are making strides, larger AIs like Bard, GPT-3.5, and GPT-4 are still more capable. Yet, the emergence and success of compact models like phi-1.5 and phi-2 suggest that small AI models can indeed be mighty and can potentially solve some of the problems posed by larger AI models. As AI continues to evolve, the focus seems to be shifting towards a more economical, sustainable, and inclusive approach.<\/p>\n<\/details>\n","protected":false},"excerpt":{"rendered":"

C\u00e1c m\u00f4 h\u00ecnh tr\u00ed tu\u1ec7 nh\u00e2n t\u1ea1o (AI) \u0111ang ng\u00e0y c\u00e0ng tr\u1edf n\u00ean l\u1edbn h\u01a1n, \u0111i k\u00e8m v\u1edbi \u0111\u00f3 l\u00e0 c\u00e1c b\u1ed9 d\u1eef li\u1ec7u \u0111\u01b0\u1ee3c s\u1eed d\u1ee5ng \u0111\u1ec3 hu\u1ea5n luy\u1ec7n ch\u00fang. Tuy nhi\u00ean, vi\u1ec7c thu nh\u1ecf c\u00f3 th\u1ec3 gi\u1ea3i quy\u1ebft m\u1ed9t s\u1ed1 v\u1ea5n \u0111\u1ec1 l\u1edbn c\u1ee7a AI. […] V\u00ec v\u1eady, \u0111\u00f4i khi nh\u1ecf h\u01a1n c\u00f3 th\u1ec3 th\u00f4ng minh h\u01a1n.<\/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":[21,105,49,106],"class_list":["post-552","post","type-post","status-publish","format-standard","hentry","category-ai-news","tag-chatgpt","tag-google-bard","tag-mo-hinh-ngon-ngu-lon-llm","tag-phi-1-5"],"_links":{"self":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/552","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=552"}],"version-history":[{"count":2,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/552\/revisions"}],"predecessor-version":[{"id":557,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=\/wp\/v2\/posts\/552\/revisions\/557"}],"wp:attachment":[{"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=552"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=552"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xuhuongai.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=552"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}