Andrej Karpathy 相信法學碩士將使 PyTorch 等深度學習框架變得多餘

法學碩士正在成為企業和研究人員面臨的許多問題的預設解決方案。即使涉及語言和文本之外的領域，人們也一直在嘗試使用 LLM 來猜測和預測下一個標記，這引發了圍繞 PyTorch 等其他工具的需求的有趣對話，因為 LLM 將來可以完成全部工作。

The rapid advancement of large language models (LLMs) has sparked a discussion about their potential to replace other tools, such as deep learning frameworks like PyTorch, for a wide range of problems.

大型語言模型 (LLM) 的快速發展引發了關於它們是否有可能取代其他工具（例如 PyTorch 等深度學習框架）來解決各種問題的討論。

LLMs are typically designed to predict the next token in a sequence, whether that sequence consists of words, images, or other types of information. This next token prediction framework can be applied to a diverse set of problems, extending beyond text.

LLM 通常旨在預測序列中的下一個標記，無論該序列由單字、圖像或其他類型的信息組成。下一個令牌預測框架可以應用於多種問題，並擴展到文字之外。

Andrej Karpathy suggests that deep learning frameworks, like PyTorch and its counterparts, might be overly general for the majority of problems in the future.

Andrej Karpathy 認為，像 PyTorch 及其對應產品這樣的深度學習框架對於未來的大多數問題可能過於通用。

LLMs have evolved beyond their initial specialization in language. The term “language” is used historically because these models were first trained to predict the next word in a sentence. However, LLMs can work on any kind of data that’s broken down into small pieces, called tokens.

法學碩士已經超越了最初的語言專業。 「語言」一詞在歷史上一直被使用，因為這些模型首先被訓練來預測句子中的下一個單字。然而，法學碩士可以處理任何類型的數據，這些數據被分解成小塊，稱為代幣。

Imagine LLMs as a super-smart guessing game. If you’re building a car, a house, or an animal using Legos, you’re just putting blocks together. LLMs don’t care if the tokens (blocks) represent words, images, or even molecules—they just focus on predicting what the next block should be based on what’s already there.

將法學碩士想像成超級聰明的猜謎遊戲。如果您使用樂高積木建造汽車、房屋或動物，您只需將積木拼湊在一起即可。法學碩士不關心標記（區塊）是否代表單字、圖像甚至分子，他們只是專注於根據現有的內容預測下一個區塊應該是什麼。

For example, protein prediction models like AlphaFold and ESMFold are built on top of generative language models. Calling such intricate models LLMs might be limiting.

例如，AlphaFold 和 ESMFold 等蛋白質預測模型是建立在生成語言模型之上的。將如此複雜的模型稱為法學碩士可能有限制。

Karpathy further highlights that the term “language” might be leading people to believe that LLMs are limited to text applications, which is not the case.

Karpathy 進一步強調，「語言」一詞可能會讓人們認為法學碩士僅限於文本應用程序，但事實並非如此。

“I don’t think this is true but I think it’s half true,” Karpathy said in response to a thread about LLMs being able to handle almost all types of problems.

「我不認為這是真的，但我認為這只是對了一半，」卡帕蒂在回應有關法學碩士能夠處理幾乎所有類型問題的帖子時說道。

“Probably the name should change.”

“也許這個名字應該改變一下。”

“Definitely needs a new name. ‘Multimodal LLM’ is extra silly, as the first word contradicts the third word,” replied Elon Musk in the same thread.

「肯定需要一個新名字。 「多模式法學碩士」特別愚蠢，因為第一個詞與第三個字相矛盾，」伊隆馬斯克在同一則貼文中回答。

Meanwhile, Yann LeCun is more concerned about why this doesn’t make sense for all the types of problems.

同時，Yann LeCun 更關心的是為什麼這對所有類型的問題都沒有意義。

“It only works with discretized outputs (discrete symbols) and only makes sense with symbol sequences with a natural order (not images). Text, DNA, proteins, musical scores, etc. are discrete or easily discretized,” said LeCun.

「它僅適用於離散輸出（離散符號），並且僅對具有自然順序的符號序列（而非圖像）有意義。文本、DNA、蛋白質、樂譜等都是離散的或容易離散的，」LeCun 說。

For something like images, which are continuous and don’t naturally have a strict sequence of discrete symbols (each pixel doesn’t follow a clear ‘order’ like text), LLMs don’t work as naturally. To use an LLM for images, you would first need to somehow convert the image into discrete chunks (like dividing the image into small patches), but this doesn’t follow the same natural order that exists in text or DNA.

對於像圖像這樣的東西，它們是連續的，並且自然不具有嚴格的離散符號序列（每個像素不遵循像文本那樣的明確“順序”），法學碩士不能自然地工作。要將 LLM 用於圖像，您首先需要以某種方式將圖像轉換為離散的區塊（例如將圖像劃分為小塊），但這並不遵循文字或 DNA 中存在的相同自然順序。

Agreeing with Karpathy, and a little with LeCun, Gary Marcus said that statistical modelling of token streams works well if reasoning or planning isn’t required.

Gary Marcus 同意 Karpathy 的觀點，也同意 LeCun 的觀點，他表示，如果不需要推理或規劃，令牌流的統計建模效果很好。

Last month, Eliezer Yudkowsky also said that predicting the next token can solve almost all the well-posed problems.

上個月，Eliezer Yudkowsky 也表示，預測下一個代幣可以解決幾乎所有適定問題。

“literally any well-posed problem is isomorphic to ‘predict the next token of the answer’,” he said.

“實際上，任何適定問題都同構於‘預測答案的下一個標記’，”他說。

“Throwing an LLM at it”

“攻讀法學碩士”

The idea that many problems can be reduced to a token-stream prediction model is intriguing, especially since domains like images, audio, and even molecules can be broken down into sequences of tokens. This suggests that a unified approach like LLMs could handle diverse tasks, reducing the need for highly specialised architectures, such as PyTorch.

許多問題可以簡化為令牌流預測模型的想法很有趣，特別是因為圖像、音訊甚至分子等領域可以分解為令牌序列。這表明像法學碩士這樣的統一方法可以處理不同的任務，從而減少對 PyTorch 等高度專業化架構的需求。

However, frameworks like PyTorch provide more than just flexibility in creating neural network models. They allow for a variety of deep learning operations that aren’t necessarily relevant for LLMs but are critical for other areas like reinforcement learning, generative models, and non-sequential tasks.

然而，像 PyTorch 這樣的框架不僅僅提供創建神經網路模型的靈活性。它們允許進行各種深度學習操作，這些操作不一定與法學碩士相關，但對於強化學習、生成模型和非序列任務等其他領域至關重要。

While it’s true that LLMs could dominate many applications, not every problem is best framed as “next token prediction.” We may see a simplification or specialization of deep learning frameworks to accommodate the increasing dominance of LLM-based models. Still, the complete redundancy of frameworks like PyTorch might be too extreme of a prediction.

雖然法學碩士確實可以主導許多應用程序，但並不是每個問題都最好被定義為「下一個代幣預測」。我們可能會看到深度學習框架的簡化或專業化，以適應基於法學碩士的模型日益占主導地位。儘管如此，像 PyTorch 這樣的框架的完全冗餘可能是一個過於極端的預測。

OpenAI’s newest model o1 gives a sense of why LLMs would be able to solve a lot of problems outside of the realm that is currently considered achievable. With the reasoning tokens in place, the model can go beyond just ‘predicting’ the next token and giving reasons for why it did so.

OpenAI 的最新模型 o1 讓人了解為什麼法學碩士能夠解決目前認為可以實現的領域之外的許多問題。有了推理標記，模型就不僅僅是「預測」下一個標記並給出這樣做的理由。

PyTorch and similar frameworks may not become redundant but could evolve to become more focused on token-based models, while still offering tools for more diverse problems outside that paradigm.

PyTorch 和類似的框架可能不會變得多餘，但可能會發展為更加關注基於代幣的模型，同時仍然為該範式之外的更多樣化的問題提供工具。

Though, currently only in language or text format, the capabilities might extend beyond it soon. Calling LLMs as LLMs might be underrepresenting their capabilities. Moreover, “it just predicts the next token” is a thought-terminating cliche.

雖然目前僅採用語言或文字格式，但功能可能很快就會超越它。將法學碩士稱為法學碩士可能低估了他們的能力。此外，「它只是預測下一個標記」是一個終結思想的陳腔濫調。