在追踪人工智能潜力的增量进展时，人类有一种奇怪的倾向，那就是用我们可能从小就没玩过的棋盘游戏来思考。尽管不乏例子，甚至是最近的例子，都突显了人工智能完全拥有纸板游戏领域的能力，但这些测试只能说明该技术在解决现实世界问题方面的有效性。
一个可能更好的“挑战”是让人工智能与人类并肩参加编程比赛。alphabet旗下的DeepMind就利用其AlphaCode模型做到了这一点。结果呢?AlphaCode表现得很好，但也不是特别出色。根据发表在《科学》杂志上与Gizmodo分享的一篇论文，该模型的整体性能相当于一个经过几个月到一年训练的“新手程序员”。今年早些时候，DeepMind公布了部分研究结果。
在测试中，AlphaCode能够达到“接近人类水平的性能”，并通过预测代码片段并创建数百万个潜在解决方案来解决以前从未见过的自然语言问题。在生成过多的解决方案后，AlphaCode将它们过滤到最多10个解决方案，研究人员称所有这些解决方案都是在“没有任何关于计算机代码结构的内置知识”的情况下生成的。
在Codeforces竞争性编码平台上最近的编码比赛中，当每个问题只能生成10个解决方案时，AlphaCode在模拟评估中平均排名前54.3%。然而，其中66%的问题在第一次提交时就得到了解决。
这听起来可能并不那么令人印象深刻，特别是与在复杂的棋盘游戏中与人类相比，似乎更强大的模型表现相比，尽管研究人员指出，在编码比赛中取得成功是非常困难的。为了取得成功，AlphaCode必须首先理解自然语言中复杂的编码问题，然后“推理”无法预见的问题，而不是简单地记忆代码片段。AlphaCode能够解决以前从未见过的问题，研究人员声称，他们没有发现证据表明他们的模型只是简单地从训练数据中复制核心logix。研究人员表示，综合这些因素，使AlphaCode的性能“向前迈出了一大步”。
卡耐基梅隆大学博世人工智能中心教授J. Zico Kolter在最近的一篇评论文章中写道:“最终，AlphaCode在以前从未见过的编码挑战中表现得非常好，无论它‘真正’理解任务的程度如何。”
AlphaCode并不是唯一一个考虑编码的人工智能模型。最值得注意的是，OpenAI已经对其GPT-3自然语言模型进行了调整，以创建一个自动补全函数，该函数可能会对代码行产生偏差。GitHub也有自己流行的人工智能编程工具Copilot。然而，这两个程序在解决复杂的竞争问题上都没有表现出与人类竞争的能力。
虽然我们还处于人工智能辅助代码生成的相对早期阶段，但DeepMind的研究人员相信，AlphaCode最近的成功将为人类程序员带来有用的应用。研究人员说，除了提高总体生产力，AlphaCode还可以“让新一代开发人员更容易编程”。研究人员表示，在最高层面上，AlphaCode有一天可能会导致编程文化的转变，即人类主要存在于制定问题，然后由人工智能来解决问题。
与此同时，人工智能领域的一些批评者对支撑许多高级人工智能模型的核心训练模型的有效性提出了质疑。就在上个月，一位名叫马修·巴特里克(Matthew Butterick)的程序员首次对微软旗下的GitHub提起诉讼，称其人工智能助手工具Copilot在学习和测试阶段公然无视或删除软件工程师提交的许可。巴特里克认为，随意使用其他程序员的代码相当于“规模空前的软件盗版”。这场诉讼的结果可能会在决定人工智能开发人员(尤其是那些用过去的人类代码训练他们的模型的人)改进和改进他们的模型的难易程度方面发挥重要作用。
原文：
When it comes to tracking the incremental advances of AI potential, humans have an odd tendency to think in terms of board games we probably haven’t played since childhood. Though there’s no shortage of examples, even recent ones, highlighting AI’s ability to utterly own the cardboard gaming space, those tests only go so far in illustrating the tech’s effectiveness at solving real world problems.
A potentially far better “challenge,” would be to put an AI side by side with humans in a programming competition. Alphabet-owned DeepMind did just that with its AlphaCode model. The results? Well, AlphaCode performed well but not exceptional. The model’s overall performance, according to a paper published in Science shared with Gizmodo, corresponds to a “novice programmer” with a few months to a year of training. Part of those findings were made public by DeepMind earlier this year.
In the test, AlphaCode was able to achieve “approximately human-level performance” and solve previously unseen, natural language problems in a competition by predicting segments of code and creating millions of potential solutions. After generating the plethora of solutions, AlphaCode then filtered them down to a maximum of 10 solutions, all of which the researchers say were generated, “without any built-in knowledge about the structure of computer code.”
AlphaCode received an average ranking in the top 54.3% in simulated evaluations in recent coding competitions on the Codeforces competitive coding platform when limited to generation 10 solutions per problem. 66% of those problems, however, were solved using its first submission.
That might not sound all that impressive, particularly when compared to seemingly stronger model performances against humans in complex board games, though the researchers note that succeeding at coding competitions are uniquely difficult. To succeed, AlphaCode had to first understand complex coding problems in natural languages and then “reason” about unforeseen problems rather than simply memorizing code snippets. AlphaCode was able to solve problems it hadn’t seen before, and the researchers claim they found no evidence that their model simply copied core logix from the training data. Combined, the researchers say those factors make AlphaCode’s performance a “big step forward.”
“Ultimately, AlphaCode performs remarkably well on previously unseen coding challenges, regardless of the degree to which it ‘truly’ understands the task,” Carnegie Mellon University, Bosch Center for AI Professor J. Zico Kolter wrote in a recent Perspective article commenting on the study.
AlphaCode isn’t the only AI model being developed with coding in mind. Most notably, OpenAI has adapted its GPT-3 natural language model to create an autocomplete function that can prejudice lines of code. GitHub also has its own popular AI programming tool called Copilot. Neither of those programs however, have shown as much prowess competing against humans in solving complex competitive problems.
Though we’re still in the relatively early days of AI assisted code generation, the DeepMind researchers are confident AlphaCode’s recent successes will lead to useful applications for human programmers down the line. In addition to increasing general productivity, the researchers say AlphaCode could also “make programming more accessible to a new generation of developers.” At the highest level, researchers says AlphaCode could one day potentially lead to a cultural shift in programming where humans mainly exist to formulate problems which AI’s are then tasked to solve.
At the same time, some detractors in the AI space have called into question the efficacy of the core training models underpinning many advanced AI models. Just last month, a programmer named Matthew Butterick filed a first of its kind lawsuit against Microsoft-owned GitHub, arguing its Copilot AI assistant tool blatantly ignores or removes licenses presented by software engineers during its learning and testing phase. That liberal use of other programmers’ code, Butterick argues, amounts to “software piracy on an unprecedented scale.” The results of that lawsuit could play an important role in determining the ease with which AI developers, particularly those training their models on past humans’ code, can improve and advance their models.
本文由数字化转型网（www.szhzxw.cn）翻译而成，作者：Mack DeGeurin；翻译：数字化转型网郑亚茹；翻译审核：数字化转型网默然。
