OpenAI 共同創辦人 Ilya Sutskever 在 NeurIPS 2024 上表示，AI 預訓練的時代即將結束

OpenAI 共同創辦人 Ilya Sutskever 最近在加拿大溫哥華舉行的神經資訊處理系統 (NeurIPS) 2024 會議上發表演講，認為人工智慧預訓練的時代正在結束，並預測 AI 超級智慧的崛起。

OpenAI co-founder Ilya Sutskever gave a lecture at the Neural Information Processing Systems (NeurIPS) 2024 conference in Vancouver, arguing that the age of artificial intelligence pre-training is coming to an end and predicting the rise of an AI superintelligence.

OpenAI 共同創辦人 Ilya Sutskever 在溫哥華神經資訊處理系統 (NeurIPS) 2024 會議上發表演講，認為人工智慧預訓練的時代即將結束，並預測 AI 超級智慧的崛起。

Sutskever’s lecture highlighted several key points:

Sutskever的演講強調了幾個要點：

**Computing Power Outpacing Data Availability for AI Pre-Training**

**人工智慧預訓練的運算能力超過資料可用性**

According to Sutskever, the rate at which computing power is increasing — thanks to better hardware, software, and machine-learning algorithms — outpaces the total amount of data available for AI model training. He likened data to fossil fuels, which will eventually run out.

Sutskever 表示，由於更好的硬體、軟體和機器學習演算法，運算能力的成長速度超過了可用於人工智慧模型訓練的資料總量。他將數據比喻為最終會耗盡的化石燃料。

“Computing power is increasing exponentially, but the amount of data is not,” said Sutskever. “This means that at some point, we will reach the limits of what can be achieved with pre-training.”

「運算能力正在指數級增長，但數據量卻沒有指數級增長，」Sutskever 說。 “這意味著在某個時候，我們將達到預訓練所能達到的極限。”

**Agentic AI, Synthetic Data, Inference Time Computing as Next Steps**

**下一步是代理人工智慧、合成資料、推理時間計算**

The OpenAI co-founder predicted that agentic AI, synthetic data, and inference time computing are the next evolutions of artificial intelligence that will eventually give rise to an AI superintelligence.

OpenAI 聯合創始人預測，代理人工智慧、合成數據和推理時間計算是人工智慧的下一個演進，最終將催生人工智慧超級智慧。

“Agentic AI is able to make decisions without human input and will be able to operate in the real world,” explained Sutskever. “Synthetic data is generated by AI models and can be used to train other AI models on a massive scale.”

「代理人工智慧能夠在沒有人類輸入的情況下做出決策，並且能夠在現實世界中運行，」Sutskever 解釋道。 “合成數據由人工智慧模型生成，可用於大規模訓練其他人工智慧模型。”

**AI Agents in the Crypto World: A Closer Look**

**加密世界中的人工智慧代理：仔細觀察**

AI agents are advancing rapidly in the crypto space, going beyond current chatbot models by being able to make decisions without human input.

人工智慧代理在加密貨幣領域正在迅速發展，超越了當前的聊天機器人模型，能夠在沒有手動輸入的情況下做出決策。

This capability has made AI agents a central topic in the crypto narrative with the rise of AI memecoins and large-language models (LLMs) like Truth Terminal.

隨著人工智慧迷因幣和大語言模型（LLM）（如 Truth Terminal）的興起，這種能力使人工智慧代理人成為加密貨幣敘事的中心主題。

Truth Terminal, an LLM, went viral by promoting a memecoin called Goatseus Maximus (GOAT), which soared to a market capitalization of $1 billion — grabbing the attention of retail investors and venture capitalists.

法學碩士 Truth Terminal 透過推廣一種名為 Goatseus Maximus (GOAT) 的迷因幣而走紅，該幣的市值飆升至 10 億美元，吸引了散戶投資者和風險投資家的關注。

“AI agents are able to interact with the world in a way that is not possible for humans,” said Sutskever. “This opens up new possibilities for how we can use AI to solve problems.”

「人工智慧代理商能夠以人類不可能的方式與世界互動，」蘇茨克弗說。 “這為我們如何使用人工智慧解決問題開闢了新的可能性。”

**Google's Gemini 2.0 to Power AI Agents**

**Google 的 Gemini 2.0 為人工智慧代理提供動力**

Google’s DeepMind artificial intelligence laboratory introduced Gemini 2.0, an artificial intelligence model that will power AI agents.

谷歌的 DeepMind 人工智慧實驗室推出了 Gemini 2.0，這是一種將為人工智慧代理提供動力的人工智慧模型。

According to Google, agents built with the Gemini 2.0 framework will be able to assist in complex tasks such as coordinating between multiple websites and logical reasoning.

據谷歌稱，使用 Gemini 2.0 框架構建的代理將能夠協助完成複雜的任務，例如多個網站之間的協調和邏輯推理。

“Gemini 2.0 is a major advance in AI technology and will enable the creation of a new generation of AI agents,” said DeepMind in a statement.

DeepMind 在聲明中表示：“Gemini 2.0 是人工智慧技術的重大進步，將有助於創建新一代人工智慧代理。”

Advancements in AI agents that can independently act and reason will pave the way for AI to overcome the issue of data hallucinations.

能夠獨立行動和推理的人工智慧代理的進步將為人工智慧克服數據幻覺問題鋪平道路。

AI hallucinations, a phenomenon that occurs due to incorrect data sets, are becoming more prevalent as AI pre-training relies heavily on using older LLMs to train newer LLMs, which degrades performance over time.

AI 幻覺是一種由於不正確的資料集而發生的現象，它變得越來越普遍，因為AI 預訓練嚴重依賴使用舊的LLM 來訓練新的LLM，這會隨著時間的推移而降低性能。