_ Engineer is obsolete? A new career promoted by AI bigwigs is born

Source: SenseAI

When most people are lamenting the huge waves, keen sailors have already set off to find the new continent. Recently, Karpathy, Tesla’s AI director and AI top streamer who returned to OpenAI not long ago, retweeted the latest episode of the Latent Space podcast, thinking that keeping up with the development of AI has become a full-time job. Only hint engineering is needed, and the demand will far exceed today's machine learning algorithm engineers. In this issue, we will continue to do further analysis on the capability portrait of AI talents that will be needed in the future.

Sense Thinking

We try to put forward more divergent deduction and deep thinking based on the content of the article, welcome to exchange.

**Model Technology vs. Product Landing: **While most people are lamenting the huge waves, keen sailors have set off to find new continents. Craftsmen who invented the compass and built ships played a role in the era, and next, sailors and workers expanded the territory.

Software 3.0 Era: SenseAI once again emphasizes the Agent network. This time, the machine is coming to us, and natural language will become the language package that completes most of the development needs, further liberating human imagination. Agents use interaction density and network collaboration to solve the execution, and what humans have to do is to propose bottom-level imagination, destroy and reconstruct.

This article has a total of 3255 words, and it takes about 9 minutes to read carefully

01. New job: AI Engineer

We are observing a generational applied AI "shift to the right" driven by the emergent capabilities and open source/API availability of the underlying models. A series of artificial intelligence tasks that took 5 years and a research team to complete in 2013 can now be completed with only API documentation and a free afternoon.

The API is universal: AI engineers can go left to optimize/host models, and research engineers can go right to build applications on top of the API, but their relative strengths and technical bases are clear.

However, the most difficult work is in the specific implementation details. At present, LLM still has the following challenges in terms of successful evaluation, application and productization:

**1. Model:**From evaluating the largest GPT-4 and Claude models, to the smallest open source Huggingface, LLaMA and other models.

2. Tools: From the most popular linking, retrieval and vector search tools such as LangChain, LlamaIndex, and Pinecone, to the emerging field of proxy tools such as Auto-GPT and BabyAGT.

3. News: The number of papers, models, and techniques published every day is growing exponentially with attention and capital, to the point where keeping a sniff of all these cutting-edge developments has become almost a full-time job .

The LLM creates a full-time job. Software engineering will incubate a new sub-discipline, focus on the application of AI, and effectively use the emerging technology stack, just like "site reliability engineer", "development operation and maintenance engineer", "data engineer ” and “Analytical Engineer” as up-and-coming.

**AI engineers, will rise to represent this type of role. **

Almost every startup has some form of AI discussion group. These groups will transition from informal groups to formal teams, as Amplitude, Replit and Notion have already done. Those thousands of software engineers working on productizing AI APIs and OSS models, whether on company time or on nights and weekends, on corporate Slack or standalone Discord, will be specialized and brought together under one title - AI engineer. This is likely to be the most in-demand engineering job in the next decade.

From the biggest companies like Microsoft and Google, to cutting-edge startups like Figma (acquired by Diagram), Vercel (RoomGPT by Hassan El Mghari) and Notion (Notion AI by Ivan Zhao and Simon Last), to independent hackers like Simon Willison, Pieter Levels (Photo/InteriorAI), and Riley Goodside (now at Scale AI). They earn $300,000 a year doing hint engineering at Anthropic and $900,000 building software at OpenAI. They spend free weekends working on ideas at AGI House and sharing tips on /r/LocalLLaMA2. What they all have in common is that they are turning advances in artificial intelligence into real products that are used by millions almost overnight.

**None have a Ph.D. When it comes to releasing AI products, you need engineers, not researchers. **

02, AI engineers will replace ML engineers

The demand for AI engineers will grow rapidly in the future. Currently on Indeed, ML engineers have 10 times the job opportunities of AI engineers, but the higher growth rate of AI leads to believe that this ratio will be reversed within 5 years.

Monthly Employment Trend Chart for HN Who's Hiring

All job titles are one-sided, but some are useful. We're wary and weary of the endless semantic debate between AI and ML, yet we're well aware that the conventional "software engineer" role is perfectly capable of building AI software. However, a recent question on Ask HN on how to break into AI engineering reveals a fundamental perception that still exists in the market:

Screenshot June 2023: Top voted answers to "How to Break into AI Engineering"

Most people still see AI engineering as a form of machine learning or data engineering, so they recommend the same tech stack. But you can be sure that none of the highly effective AI engineers mentioned above have done work equivalent to the Andrew Ng Coursera course, they don't know PyTorch, and they don't know the difference between a data lake and a data warehouse.

**Soon, no one will suggest starting AI engineering by reading Attention is All You Need, any more than you will learn to drive by reading drawings of a Ford Model T. Of course, understanding the fundamentals and history is always helpful, and it can really help you find innovations and efficiency/capability gains that haven't yet entered the general consciousness. But sometimes, you can use the products directly and know their qualities through experience.

Admittedly, the reversal of AI engineers and ML engineers will not happen overnight. People naturally want to flesh out their resumes, fill out market maps, and stand out by citing more authoritative in-depth topics. That said, prompt engineering and AI engineering will feel at a disadvantage over those with a solid data science/machine learning background for a long time. However, the economics of supply and demand will eventually win out, and the demand for AI engineers will far exceed that of ML engineers.

**03. Why now? **

1. The base model is a "few-shot learner", exhibiting the ability to learn even zero-shot transfer in context, which goes beyond the original intent of the model trainer. In other words, the people who created these models didn't fully know what they were capable of. Those who are not LLM researchers can discover and take advantage of these capabilities simply by spending more time interacting with the models and applying them to underappreciated areas of research (e.g., the application of Jasper to copywriting).

2. Microsoft, Google, Meta, and large fundamental model labs have monopolized scarce research talent, essentially providing "AI research-as-a-service" APIs. You can't hire them, but you can rent them - if you have software engineers on your team who know how to work with them. **There are about 5,000 LLM researchers worldwide, but there are about 50 million software engineers. Supply constraints determine that a "hub" AI engineer will emerge to meet market demand. **

**3. GPU Reserve. **Of course, OpenAI/Microsoft were the first to do this work, but Stability AI sparked a GPU arms race among startups by emphasizing their 4000 GPU cluster.

**4 Agile action directly from the product. **Instead of requiring data scientists/ML engineers to do a heavy lifting of data collection before training a domain specific model and then putting it into production, product managers/software engineers can prompt LLM first, and build/validate a product idea, then Then obtain specific data for fine-tuning.

Let's say the latter is 100x to 1000x more than the former, and by cuing LLM's "fire, prepare, aim" workflow gets you 10x to 100x faster than traditional machine learning. As a result, AI engineers will be able to validate AI products at 1,000 to 10,000 times lower cost. This is another Waterfall versus Agile development, and AI is Agile.

**5. Python → Java. **Data/AI has traditionally relied heavily on Python, and the first AI engineering tools like LangChain, LlamaIndex, and Guardrails also originated from the same community. However, there are now as many Java developers as there are Python developers, so tools now increasingly cater to this expanded user base, from LangChain.js and Transformers.js to Vercel's new AI SDK. The opportunity for market expansion is enormous.

**6. Generative AI vs Classifier ML. ** "Generative AI" as a term has fallen out of fashion, making way for other categories such as "Inference Engines", but in succinctly lays out the difference between existing MLOps tools and machine learning practitioners and what is best suited to use LLM and Still very useful when looking at the new and distinct role of the text-to-image generator. While existing machine learning research might focus on things like fraud risk, recommender systems, anomaly detection, and feature storage, AI engineers are building writing apps, personalized learning tools, natural language spreadsheets, and Factorio-like apps. Visual programming language.

Whenever there is a subgroup with a completely different background, speaking a different language, producing a completely different product, and using a completely different tool, they eventually split into separate groups.

04, 1+2=3: Programming in the era of software 3.0

6 years ago, Andrej Karpathy wrote a very influential article describing Software 2.0, contrasting the "classical stack" of traditional hand-written programming languages that model logic precisely, and the new "machine learning" neural networks that approximate logic. stack, enabling software to solve more problems than humans can model. This year, he noted in a follow-up article that the Hottest new programming language is English, finally filling the gray area he didn't mark in his original article.

Engineering became a meme last year, describing how work will change when people start taking advantage of GPT-3 and Stable Diffusion. People mock AI startups for calling them "OpenAI wrappers" and raise concerns that LLM applications are vulnerable to hint engineering and reverse hint engineering. Do barriers really exist? (Sense Says: Refer to our first article "Intelligent Systems: The Moat of Future AI Enterprises")

But one of the biggest themes in 2023 will be re-establishing the role of human-written code in coordinating and displacing LLM capabilities, from the over $200M Langchain, to the Nvidia-backed Voyager, demonstrating the clear importance of code generation and reuse. Engineering is both overhyped and here to stay, but the reemergence of the Software 1.0 paradigm in Software 3.0 applications is both a huge opportunity and an area of confusion, creating white space for numerous startups.

An investor who cannot do market research is not a good investor

Of course, this isn't just code written by humans. The stories of numerous projects (smol-developer, more broadly gpt-engineer, and the adventures of other code-generating agents like Codium AI, Codegen.ai, and Morph/Rift) suggest that they will increasingly be part of the AI engineer's toolkit . As human engineers learn to harness AI, AI will increasingly be involved in engineering.

Until someday in the far future we'll look up and no longer be able to tell the difference between the two.

References