The New Wave of Technology: AI Engineers on the Rise

I recently saw an article about the rise of AI engineers on latent.space, which aroused the author's interest. After further collecting and sorting out some of the latest developments and information about AI engineers. The author sorts it out into this article, hoping to be helpful to everyone, and to discuss and learn together.

With the continuous development of technology, the capabilities of the basic model are continuously enhanced and improved, and the supporting facilities of AI are becoming more and more abundant. AI-based products are penetrating into all aspects of our daily lives. Many industries, such as smart furniture, teaching models, medical diagnosis, and automatic driving, all have it. We are witnessing a historic change in the field of AI applications.



In this transformation, a new vocabulary is emerging: AI engineer (AI engineer). And its roles and responsibilities are changing compared to before. Previously, an AI-related task required a whole research team and took a lot of time to complete. In 2023, people can accomplish the same task in a non-busy afternoon simply by consulting Open AI's API documentation.



This is a discussion on Twitter. With the continuous development of the AI ​​field, different roles are gradually differentiated and the division of labor is refined. The position of the AI ​​engineer is on the right side of the API. In the figure, the position of the AI ​​engineer is to develop some applications and tools based on the API. , proxy, etc.

What is an AI engineer

Due to the current rapid development of this field, some meanings have not yet been determined. AI is also a very vague concept. The content of AI engineers is actually very wide. With the rapid development of AI technology, this position has become more and more important and prominent.



In order to focus on the content, try to define some AI engineers. Some time ago, there was also a Prompt Engineer, but the meaning of only Prompt Engineer is too narrow, and many things are not included.



Definition: An AI engineer is an engineer who specializes in designing, developing, and deploying AI models and applications . They typically have a background in software engineering, data science, and machine learning and are able to combine these skills to create efficient, scalable AI solutions.



Some things that can be done:

1.  Prompt Engineer (Prompt Engineer ): Responsible for debugging Prompt to make the model have a stable output. Perhaps with the development of the model, the requirements for Prompt will not be so high. To really take advantage of the power of this model, you have to be able to write code, not just a non-technical Prompt Engineer, in order to be more efficient. In other words, they have programming thinking and use natural language for programming.
2. Data processing: Understand some data processing methods, clean, preprocess, label, enhance data, and do some model fine-tuning.
3. Software engineering: Do some applications based on models, solve some real-world problems based on AI, application deployment and expansion, system stability, scalability design, performance optimization, etc. Understand the supporting tools around the AI ​​model.

4. Some algorithms: You can also go deep into the field of algorithms...to do more difficult AI model training

Focus on the application field of AI and solve some problems in the application.

ML Engineer vs AI Engineer

Most people still think of AI engineering as a form of machine learning or data engineering, so many people think that deep learning, machine learning, etc. need to be mastered. However, according to the observations of some foreign leaders, many outstanding AI-related engineers have not completed the study and research on deep learning, and they do not know PyTorch, or even the difference between data lakes and data warehouses, because AI engineers There is also no direct requirement for data engineering.



In some discussions abroad, some of the concerns of the two are different:

1. ML engineers are primarily concerned with the design, development, and deployment of machine learning models, and their work is usually more specific and granular. Handle practical issues related to models such as overfitting, imbalanced datasets, feature selection, etc.

2. AI engineers are used to implement AI solutions: build complete AI applications, such as chatbots, image recognition systems, etc. Not only to deal with machine learning, but also focus on other fields such as deep learning, reinforcement learning, etc.



On the overseas Indeed recruitment website, you can see ML, AL and other recruitment information. You can see that the growth rate of AI jobs and ChatGPT jobs is the highest, and it is expected that they will gradually surpass ML in the future. The X-axis is the year, and the Y-axis is the percentage of job posting types.



In the future, no one may suggest that you start learning AI engineering by reading "Attention is All You Need". It is definitely helpful to understand the basics and principles, but it is not a necessary condition. There are so many frameworks in development, and it is not necessary for each framework to go deep into the principle level. Sometimes you only need to use the product, and it is also very important to be able to do the application well.



Maybe in the short term for those with a good data science/ML background, AI engineers, Prompt engineers will feel inferior for a long time. Yet Big Brother's view is that pure demand and supply economics will prevail, and AI engineers will have their place.

▐   Hacker News discussion on AI/LLM engineers

https://news.ycombinator.com/item?id=36538423



On whether AI engineers are on the rise?



Take a positive view:

Practicality : AI/LLMs are already generating real revenue in areas such as generative text, art, code completion, and conversational AI, suggesting that their utility goes beyond the hype.

Engineering challenges : There are non-trivial engineering challenges in areas such as hint engineering, data preparation, model evaluation, and model integration into products. This justifies the "AI engineer" profession.

Job market : The current job market favors experienced AI/LLM engineers because talent is scarce. Companies are willing to pay high salaries.

Resources and tools : Improvements in resources and tools have made AI/ML more accessible to engineers than 5-10 years ago.

Skeptical view:

Job Title : The title "AI Engineer" is too vague. A specific job title like "LLM Engineer" or "Generation Engineer" is better.

Deep understanding : Many "AI engineers" only lightly integrate off-the-shelf models without deep understanding. This risks diluting its meaning.

AI Capabilities : AI capabilities are still limited. Overcommitment can lead to disappointment.

Legal issues : There are still unresolved copyright/IP legal issues for commercial AI applications.

Model Evaluation : Proper model evaluation and error handling is difficult and often overlooked by "AI engineers".

Data generation : Automatic data generation without human supervision runs the risk of bias and model degradation.

Why do AI engineers appear?



▐   The atmosphere of AI productization is hot

Most foreign start-up companies have a Slack discussion channel about AI #discuss-ai, these channels will change from informal groups to formal teams, such as Notion, Amptitude, and Replit have all done it. Plus independent hackers like Simon Willison, Pieter Levels (founder of Photo/InteriorAI), and Riley Goodside (now at Scale AI), engineers who spend their spare weekends working on ideas at AGI House and posting on /r/ Share tips on LocalLLaMA2. They are all taking the latest advances in AI and quickly turning them into real products used by millions of people.



There are already a large number of engineers working hard to productize AI. Over time, these software engineers will gradually specialize and use a unified title-AI engineer. AI engineer is expected to be the most in-demand engineering job title this decade. Although there are many researchers and PhDs in the AI ​​field, when it comes to actually developing and releasing AI products, there is a greater need for AI engineers than researchers. Practical and executive experience may be more important than academic research in practical applications of AI .

▐There   are certain challenges in productization

AI技术的应用和产品化看起来很有前景，但是在实际操作中，成功地评估、应用和产品化AI充满了还有很多细节的挑战：

1. 模型的多样性 ：从评估最大的GPT-4和Claude模型，到最小的开源Huggingface、LLaMA等模型，选择和评估合适的模型是一大挑战。
2. 工具的多样性 ：工具的选择范围从最受欢迎的链式、检索和向量搜索工具，如LangChain、LlamaIndex和Pinecone，到新兴的自主代理领域的工具，如Auto-GPT和BabyAGI。
3. 信息过载 ：每天发布的论文、模型和技术的数量正在指数级增长。以至于跟上这些信息几乎成了一份全职工作。

▐  市场的供需关系

1. 大公司的角色：像Microsoft、Google、Meta这样的大公司已经垄断了稀缺的研究人才，提供“AI研究即服务”API。你不能雇佣他们，但你可以利用他们的能力，全球大约有5000名LLM研究者，但有约5000万名软件工程师。供应限制意味着一个“中间”类别的AI工程师将应运而生。

2. GPU囤积：Stability AI通过强调他们的4000 GPU集群引发了初创公司的GPU竞赛。各种公司都在做GPU的储备竞赛。全球芯片短缺进一步加剧了这一趋势。大部分公司是训练不了模型的，只能去利用模型。

   

3. 快速验证想法：与其要求数据科学家/ML工程师在训练一个特定领域的模型之前进行繁琐的数据收集，不如让产品经理/软件工程师Prompt一个LLM，并构建/验证一个产品想法，然后再获取特定数据进行微调。使用LLM原型的工作流程可以使你的工作速度比传统的机器学习快10-100倍。因此，AI工程师能够以比传统方法低1000-10000倍的成本验证AI产品。

   

市场上没有那么多的人会训练基础模型，但是市场上需要很多人快速验证有关LLM的想法，并且有足够多的开发人员可以具备开发LLM应用的能力。



参考AI应用全景报告：https://ramsrigoutham.medium.com/the-landscape-of-generative-ai-landscape-reports-615a417b15d

软件工程演变中代码的角色？编程语言->自然语言

六年前，Andrej Karpathy写了一篇很有影响力的文章，描述了“软件2.0”。他比较 了两种软件开发方式：

1. 传统方式：程序员通过编写代码来明确指示计算机如何执行任务。
2. 新方式：通过神经网络和机器学习来“学习”如何执行任务，而不是通过明确的编程指令。这种方法的优势在于，它可以处理大量的数据并从中学习，而不需要人为地为每一个可能的情况编写代码。 类似于当前的算法推荐；

他最近又提到，现在最热门的“编程语言”实际上是英语，通过自然语言对话来完成特定的任务。

去年，随着人们开始使用GPT-3和Stable Diffusion，"Prompt Engineering"成为了一个流行的概念，如何通过提示来与这些模型交互和工作。“Prompt Engineer”这个称呼可能不太准确，因为这个角色不仅仅是给AI模型提供提示，还涉及到很多代码和基础设施的工作。或许还是AI工程师(AI Engineer)这个称呼更为合适。

有人说这种做法只是OpenAI的包装，只是使用技术而没有进行任何创新，还有担心LLM应用的安全隐私问题、提示注入和一些反向工程问题。



问题虽然存在，但毕竟这个是一个正在快速发展的领域，每个都可以有自己的观点，如果相信这个领域会快速发展，这些问题可以在前进的路上逐渐的去解决。还有当前LLM看似大，但也需要人类编写代码(或者通过Prompt)来指导和增强它的功能。



未来的编程不仅仅是由人类编写的代码，随着工程师越来越擅长利用AI，AI也将越来越多地参与到工程设计中。直到有一天，我们可能无法分辨出哪些是人类编写的，哪些是AI生成的代码。通过自然语言就可以构建出复杂的软件工程...

怎么成为AI工程师？

AI当前是一个快速发展的领域，所以要持续的学习跟随趋势关注行业的最新动态，推特上的大佬 @swyx 在一次AI的播客讨论上，提出了一些成为AI工程师的参考路径，大概如下。

播客地址：https://twitter.com/swyx/status/1674895620870651909

1. 基础教育 AI工程师需要对线性代数、微积分、概率和统计有了解，因为这些都是机器学习算法的基础。掌握至少一种编程语言，如Python或R，这是进行机器学习编程的基础。

2. 专业知识培养: 机器学习: 学习各种机器学习算法，如线性回归、决策树、神经网络等。深度学习: 了解CNN、RNN、Transformer等模型结构，以及如何使用框架如TensorFlow和PyTorch来实现它们。

3. 实践经验:

   项目实践: 手头上的项目经验很重要。从小项目开始，逐渐处理更复杂的问题。

   参与竞赛: Kaggle等平台提供了很多数据科学和机器学习的竞赛，通过这些竞赛，你可以学习到很多实践经验。

4. 信息获取：

   研究: 阅读最新的研究论文，了解最新的技术和趋势。

   网络: 加入AI和机器学习的社区，如OpenAI、Google AI等，与其他专家交流，分享知识。

   课程与认证：考虑获得与AI和机器学习相关的认证，如TensorFlow认证或AWS Machine Learning认证。

5. 确定方向：AI是一个快速发展的领域，总有新的技术和方法出现。保持好奇心，始终对新知识保持开放的态度。同时AI是一个可以广泛应用的领域，确定你最感兴趣的领域，深入研究。

不过以上感觉是针对研究型人员的成长路径，针对大多数人，想要往AI领域发展，怎么样学习成长，个人尝试按照自己的理解梳理一下：

1. 编程语言： 首先要懂一门编程语言、Python简单易上手，而且现在有AI加持；通过GPT辅助自己学习Python。

2. 基础原理： 了解大模型的原理，LLM是如何工作的，SD是如何工作的，对常见的AI领域的算法有一定了解，知道大致是干什么的。
   

3. 模型了解：了解市场上有哪些模型，各自的适用场景，都可以用来做什么；能都亲自体验一把

4. 提示词工程，学习怎么通过提示词让模型返回自己期望的结果。
   

5. LLM工具，通过工具提升自己对大模型的操作效率，利用相关工具能快速验证想法；
   

6. 软件工程，将LLM的能力产品化封装，并且能具备一定的部署运维能力，做一些实际的项目，积累实践经验
   

如果说想要往这个方向发展，最关键的部分，可能还是要有好奇心，保持持续的对技术的探索和学习，以及能够亲身实践对应的一些项目。



未来AI工程师的畅想？

1. 与其他领域的合作：也许未来的AI工程师可以类似当前的软件工程师，深入某个领域做领域建模，与领域专家合作，深入到各行各业，与医生、艺术家、建筑师等其他领域的专家合作，共同创建创新解决方案。例如，在医疗领域，AI工程师可以与医生合作，开发出更精确的诊断工具。

2. 自动化任务： 随着技术进步，大部分的任务都可以被AI自动化掉，AI工程师的角色将更多地转向创造、优化和维护复杂的系统，而不仅仅是编写代码。

3. AI的解释性： 工程师更加重视AI决策的透明度和解释性，确保AI的决策可以被人类理解和接受。

4. 道德和伦理：由于AI的普及和深入，对道德和伦理的考虑将成为AI工程师的基本素养。不仅要考虑技术的创新，还要确保其技术的应用不会对社会和个人造成伤害。

5. 具身AI，机器人：融合对心理学、生物学、神经科学、哲学等学科有一定的理解，开发出能理解人类智慧和情感，更加高效和人性化的AI系统，AI机器人。

总结

当下，随着大模型的不断发展和基础设施的完善，AI的技术越来越丰富，AI工程师这一新角色正在崛起，它会区别于以往的ML工程师，其重点不是设计和训练模型，通过Prompt提示，结合一些软件工程，一些大模型工具，然后构建AI相关的产品，解决实际问题。



最后也尝试去梳理了下成为AI工程师要学习成长的路径，期望能和大家一块进步和探索，在未来的世界中，利用新技术改变大家的生活。



Tomorrow belongs to those who embrace it today



团队介绍

大淘宝技术用户运营平台技术团队是一支最懂用户，技术驱动的年轻队伍，以用户为中心，通过技术创新提升用户全生命周期体验，持续为用户创造价值。

团队以创新为核心价值观之一，鼓励团队成员在工作中不断探索、实验和创新，以推动业界技术的进步和用户体验的提升。我们不仅关注当前业界领先的技术，更注重未来技术的预研和应用。团队成员会积极参与学术研究和技术社区，不断探索新的技术方向和解决方案。

团队立足体系化打造业界领先的用户增长基础设施，以媒体外投平台、ABTest平台、用户运营平台为代表的基础设施赋能阿里集团用户增长，日均处理数据量千亿规模、调用QPS千万级。在用户增长技术团队，我们提供“增长黑客”极客氛围和丰富的岗位选择，欢迎业界贤才加入。

¤ 拓展阅读 ¤

3DXR技术  |  终端技术  |  音视频技术

服务端技术  |  技术质量  |  数据算法

本文分享自微信公众号 - 大淘宝技术（AlibabaMTT）。
如有侵权，请联系 [email protected] 删除。
本文参与“OSC源创计划”，欢迎正在阅读的你也加入，一起分享。

微软官宣：Visual Studio for Mac 退役 中国开发者团队创建的编程语言：MoonBit（月兔） C++ 之父 Bjarne Stroustrup 分享人生建议 Linus 也反感乱七八糟的缩写，什么 TM 的叫 "GenPD" Rust 1.72.0 发布，未来支持版本最低为 Windows 10 文心一言面向全社会开放 WordPress 推出 “百年计划” 微软不讲武德，用“恶意弹窗”提示用户弃用 Google 高级、函数式、解释型、动态编程语言：Crumb 青语言 V1.0 正式发布

{{o.name}}

{{m.name}}