8 ways in which LLMs will create new job opportunities

AI is enjoying a lot of hype these days. I am not a AI hater, but I am a hype hater. Getting to much better AI is more complicated that many believe.

I have enumerated below 8 points in which LLMs would create more job opportunities.

1. New startups: In the past, only rich companies could afford websites since it required expensive experts. Things like WordPress or Squarespace made websites cheaper, allowing more startups to exist. Many grew into big businesses that then hired more software engineers. Similarly, LLMs will allow non-technical entrepreneurs to quickly build products to test ideas. Most will fail, but some will find a niche, grow, and realize the LLM-generated code has limitations. At that point, they will hire technical talent, creating more jobs.
   Just as those wordpress enabled many new businesses to get online affordably, leading to more hiring of tech professionals down the line, I expect that LLMs would allow more non-technical entrepreneurs to launch startups without initially hiring developers. Once their businesses gain traction, these startups will likely need to hire software engineers to scale and improve upon the LLM-generated code. 

2. Automation creating more products: Consider an assembly line where a robot makes workers 5x more productive. Often only 1 worker is retained and the robot makes the same output. But with software, the LLM “robot” is cheap, there are no material costs, and more output can be sold. Thus, more features and products will likely be made, requiring more programmers. It may offset some layoffs.
   Likewise — LLMs making developers vastly more productive will incentivize companies to rapidly expand product lines and features rather than lay off engineers, since the “raw materials” of software have negligible marginal costs.

3. Customizing LLM models: LLMs are expensive to train. Only big companies can afford huge models. But techniques exist to customize generic models, which requires engineers without needing AI expertise. There will be work tailoring output for specific requirements.
   While few can afford to train massive LLM models from scratch, there is an emerging need to customize general models for specific use cases. I expect this model customization and fine-tuning work to create engineering jobs, even for those without expertise in machine learning.

4. Maintenance and monitoring: LLMs have lots of quirks needing ongoing tweaks. They can produce harmful output needing review. Monitoring systems to audit output and compliance will be needed.
   Humans will still be needed to specify requirements, evaluate outputs, and handle edge cases that LLMs struggle with. Rather than being replaced, software engineers will collaborate with and manage AI assistants. Their role will evolve to focus more on higher-level tasks.

5. Debugging LLM-created code and regressions: Despite advancements in AI, bugs will continue to persist in both LLMs and the code they generate. Debugging such issues requires understanding both the underlying programming constructs and the unique characteristics of AI-driven code. Addressing regression bugs — unforeseen consequences arising from changes in existing systems — may prove particularly challenging due to the complexity of integrating LLM-generated code within larger systems. Specialized expertise will be necessary to ensure seamless integration and minimize unintended impacts.

6. Defending against adversarial attacks: Adversarial attacks aim to manipulate or misdirect AI systems to achieve malicious outcomes. Preventing such attacks requires continuous monitoring, identification of vulnerabilities, and implementation of safeguards. Given the limited understanding of possible attack vectors for LLMs, staying ahead of potential threats poses a considerable challenge. Developing robust defenses against adversarial attacks represents an emerging area of opportunity for software engineers and cybersecurity specialists.

7. Teaching computers: Humans still vastly outperform LLMs at many cognitive tasks. Systems requiring complex reasoning, planning, etc. will still need human input. Engineers will be employed to structure knowledge and train computers.

8. LLMs will unlock entirely new domains and product categories that we cannot yet foresee, much like the internet and mobile computing enabled vast new industries to emerge. These uncharted territories will create fresh demand for software skills.

There is potential for job displacement but he net impact will be job creation in the field, drawing on historical parallels.

I encourage open debate but urges looking at evidence over hype or absolutist claims.

I see more upsides than downsides for software engineering careers from advanced AI development.

A few books on deep learning that I am reading:

Book 1

Book 2

Book 3