Process > LLMs

Process > LLMs

What makes a generalist?

It is easy to think of generalists as jack of all trades and masters of none. Generalists are the kind of people who can do two wildly different jobs from one week to the next because their skills, experience, and interests provide a broad foundation of knowledge.

When viewed through this lens, LLMs are generalists. They have consumed much of the internet and can offer confident-sounding answers to almost any question.

The temptation to treat LLMs as generalists has fuelled much of the sensationalist claims that entire workforces can be replaced with AI. When much of the work in a technology-focused role is simply creating and moving information around, why use a human when an LLM can respond so much faster, with much more confidence, and never complain?

I’d argue though that measuring generalists by the breadth of their knowledge is the wrong metric. Instead, it is better to measure generalists by their ability to deliver outcomes in the absence of well-defined processes.

These two ways of defining a generalist overlap to such a large extent when applied to people that the difference is mostly irrelevant. Take the most common representation of a generalist: the full-stack developer. A full-stack developer has an understanding of frontend and backend development, database management, cloud technologies, deployment strategies, and monitoring because they have done all of these things in the past and can replicate them in the future. It is expected that having and applying knowledge are two sides of the same coin. Seasoned full-stack developers can deliver meaningful outcomes with very little established process as they establish enough of a foundation on which to build the desired result.

However, these two definitions of a generalist can not be applied to LLMs. LLMs have a breadth of knowledge that no human could ever hope to attain. However, they can not apply that knowledge without rigorous processes.

If you ask an LLM to generate a web app for a shopping cart, it will do so and will produce something at least as functional as an average developer would produce. What an LLM will not do is infer your intention and also:

  • Implement CI/CD pipelines
  • Decide on deploying to a local server or a cloud platform
  • Add monitoring and alerting
  • Build supporting runbooks
  • Implement security scanning

In other words, an LLM will do what you ask it to do, but nothing more. Indeed, LLMs won’t be able to implement anything other than a text-based response without a serious amount of work to integrate it with all your platforms.

If you asked a human full-stack developer to produce the same web app, they would intrinsically understand the full scope of what was being requested and have the ability to deliver a complete solution.

When measured by their ability to deliver meaningful outcomes in the absence of robust and repeatable processes, LLMs are not generalists. They have amazing potential, but only when existing processes provide them with all the input they need, and can consume and implement their outputs.

The most successful implementations of LLMs will come from teams that have already done the hard work of building robust, repeatable, and trusted processes.

I expect though that many companies will presume that LLMs are generalists because they produce confident, and ever more accurate, answers to almost any question. This, in turn, leads to the idea that they now have an infinite supply of tireless robot employees ready and eager to turn high-level ideas into solutions as fast as a GPU will allow it.

It won’t work out well.


Completely agree Matt!

回复

要查看或添加评论,请登录

Matthew Casperson的更多文章

  • Reducing the "Time to Understand Customer" with AI

    Reducing the "Time to Understand Customer" with AI

    DevOps teams have long embraced the value of consistent metrics to measure their performance, with the DORA metrics…

  • Boiling the DevOps frog

    Boiling the DevOps frog

    These are the requirements for a random “junior full-stack developer” role I found advertised this morning. These…

  • Leetcode is awful

    Leetcode is awful

    We all know the familiar sitcom grocery store trope involving a pyramid of cans stacked in the middle of an aisle…

  • Replacing jobs with GenAI is the worst of DevOps all over again

    Replacing jobs with GenAI is the worst of DevOps all over again

    There is no doubt that DevOps called out some of the worst practices in IT departments. The inefficiencies of…

  • DevOps is a flat circle

    DevOps is a flat circle

    Moving from an engineering role into a highly technical sales role provides an amazing vantage point from which to…

  • (Almost) no one cares about your platform

    (Almost) no one cares about your platform

    I had the pleasure of attending, and presenting at, GitHub universe this year, and like most conferences, it was the…

  • Strangers deploying microservices

    Strangers deploying microservices

    I’ll admit I was skeptical when the GitHub team told us about the interactive sandbox sessions at GitHub Universe. If…

  • Deployment insights for everyone with LLMs

    Deployment insights for everyone with LLMs

    Logging levels are something that developers take for granted. I want to see WARN and above logs for my day to day…

    1 条评论
  • Octopus and Copilot as your own personal deployment firefighter

    Octopus and Copilot as your own personal deployment firefighter

    Imagine that production is down and the cost of lost business is adding up by the minute. This is a five alarm fire and…

    3 条评论
  • LLMs are not magic or scary

    LLMs are not magic or scary

    Tools like ChatGPT can seem like magic. It is now almost impossible to determine if text is written by a person or…

    1 条评论

社区洞察

其他会员也浏览了