AI has its work cut out for itself. Apparently, it’s going to end humanity, take our jobs, leave our children unable to write, and add untold trillions annually to the economy.

Perhaps.

But in the meantime, I’ve found lots of much less exciting uses that have helped me become more productive.

At this, I can feel the crowds clamoring to know, “How?”

I’m glad you asked. Having scientifically surveyed a representative sample of myself, and receiving an astounding 100 percent response rate, I can tell you.

First, at TPI we’ve built two LLMs to help policymaking. ChatTPI lets you ask questions about any topic and learn about it from TPI research, podcasts, events, and more. ChatBEAD lets people better explore states’ plans for implementing their new broadband subsidy plans.

Second, I use commercial LLMs as much as possible in my own daily workflow. For the past couple of months I recorded every use, and now I’m using all the time I saved to analyze and describe them here.

My LLM use (excluding ChatTPI) includes 5 basic categories: writing assistance, data work, coding, general research assistance, and graphics. The figure below shows those categories and their subcategories. The categories are somewhat subjective, with GPT4 helping to create them from my records, of course.

Writing

61 percent of my LLM use is related to writing. That’s not too surprising since they are, after all, Large Language Models. Writing includes three sub-categories: Editing, drafting content, and summarizing text.

Editing

I spend a lot of time editing my work and others’. LLMs have become a key tool to help edit my own work. I ask it to identify weaknesses in arguments, redundancies, suggestions for reorganization, ways to shorten sentences, and other editorial tasks.

While it has become an invaluable assistant editor for my own work, I do not find it particularly helpful in editing other peoples’ work. Perhaps that’s because I have to read others’ work carefully myself to understand it, and it’s not possible to trust AI edits if I don’t already know the material and the writing deeply.

Examples:

• Eliminate redundancies in a document
• Shorten paragraph and make it more impactful
• Review draft op-ed 

Summarizing Content

LLMs are particularly good at ingesting a lot of information and synthesizing it. I use this feature to summarize podcasts and panel discussions, including asking it to give me some top takeaways. To be useful, though, I have to already be reasonably familiar with the content, because sometimes the LLM will fixate on the wrong issues and completely miss others. If I ask for 5 top takeaways, I’ll likely take three (that I edit heavily) and then add two the LLM did not notice. 

Drafting Content

People have famously gotten into trouble using content they’ve asked an LLM to create. Frankly, I’ve found LLM-generated written content to be largely uncreative and repetitive. It also has some “tells.” ChatGPT loves the word “delve.” That word is often a dead giveaway of ChatGPT-created prose.

Nevertheless, LLM-generated written content is great for certain purposes.

One use is overcoming writer’s block. Prompt it with your thoughts without worrying about how to organize or craft the prose. The output won’t be great–and is likely to have you “delving” into whatever the topic is–but it gets you over the “staring at a blank page” hurdle. By the time you’re done, the result may not (and probably should not) resemble the LLM’s output in any way.

Another writing task suitable for LLMs is content that require brevity and accurate information rather than creative thinking, like press releases about new papers, upcoming events, and new podcast episodes. While the output still requires careful editing, this use can save significant time.

Finally, and perhaps somewhat ironically given my point above about creativity, I’ve found them useful for generating titles. They often find a turn-of-phrase that wouldn’t have occurred to me but captures the essence of the piece. Again, rarely perfect out of the gate, but provide a starting point to improve upon.

All of these uses require two key user skills. One is the ability to write a good prompt. That means being able to convey the correct information and the type of output you want, which, itself, is a form of writing. The second is carefully reviewing, checking, and usually editing the output. That means the net time saved is the time you saved writing less the time you spent editing and changing the LLM output itself.

Examples:

• Create op-ed or podcast title
• Write blurb to advertise fireside chat

Data

One of the more exciting applications of AI, at least for me, is how it can work with data. I’ve used OpenAI’s “Data Analyst” tool to reformat data for different types of analyses, visualizations, and even analysis. I remain a bit skeptical of many analyses’ output since the process tends to be black box-ish, although explanations are improving. Still, I find I have to do an analysis myself anyway to make sure it’s right. But it’s getting better fast.

Examples:

• Upload a dataset and create a graph of certain variables over particular time periods using different weights
• Extract data points from an image of a graph

Coding

LLMs can help create or debug code in nearly any programming language. That’s extremely helpful, but perhaps the biggest advantage for most people is that LLMs begin to make it possible to code in many languages without actually knowing that language.

Working with data in an LLM means understanding pseudo-coding, which just means describing what you want the computer to do in natural language instead of a formal programming language. While it does not require one to know C++, Java, or any other programming language, it is necessary to describe precisely what you want the machine to do.

At this stage in LLM development it still seems necessary to know some coding, as the generated code often doesn’t work on the LLM’s first attempt, and knowing how to code yourself makes it far easier to figure out the problem by working together with the LLM. And, to further caveat, one is unlikely – for now – to be able to code anything complicated from start to finish using only an LLM

Still, LLMs’ ability to create code from a description of what you want the code to do suggests that education might need to shift from learning how to code in specific languages, to understanding the logic code follows generally and how to express desired output with sufficient specificity to create workable prompts.

Example:

• Generate code to create relative times in Stata

Research Assistance

This category is probably the closest I come to using LLMs like a search engine in that these tend to be queries for answers rather than prompts for output.

Example:

• Where should I pitch an article on a specific topic or how to install software that has several components.

Graphics

Graphics differ from data visualization, which is depictions that show some view of data. Graphics are images that may be useful for different purposes. For example, I’ve used it to create header images for different projects in project management software and as imagery to help make projects memorable.

For example, the image below is intended to help promote a tool we developed to combine and use broadband data from a large number of sources.

It’s Just a Tool

While AI is revolutionizing the way I work, it’s still just a tool. A new one, for sure, but a tool nonetheless, and one that should complement the others in the toolkit. It does not replace the creative process of thinking of interesting research questions or how to answer them. Instead, it frees me up to spend more time on those parts of my job and less on the bureaucratic parts.

It helps that I enjoy futzing with new technologies. If you don’t, the benefits are likely to be smaller until someone introduces a polished tool just right for your particular needs.

My experience, nevertheless, may have some broader lessons. Specifically, AI, and LLMs in particular, offer new tools in a wide range of occupations and industries. Some will be headline-worthy, like image-recognition tools that identify cancer before humans can. But most won’t. They’ll be incremental improvements that allow people to spend more time on the parts of their job that reflect their comparative advantages and less on the parts that pull them in unproductive directions.

I don’t mean to minimize AI concerns and challenges, like copyright and how to fix timeline problems in The Terminator series. But as we think about those, let’s also remember the small ways that AI can make us more productive that, when taken together, might be hugely important.

BONUS CONTENT! What Do LLMs Have to Say About This Post?

I prompted ChatGPT and Gemini Advanced, “What do you think of this blog post?” and pasted in the above post. Their responses are below. (Can you find the hallucination?)

ChatGPT (GPT 4)

This blog post offers a comprehensive and nuanced exploration of the practical benefits AI, particularly Large Language Models (LLMs), brings to your work. It successfully demystifies the technology, moving beyond the hyperbolic fears and promises often associated with AI to focus on real-world applications. Here are several strengths and areas for potential improvement to consider:

Strengths:

• Personal Insight and Examples: The detailed account of how you've integrated AI into various aspects of your work, supported by specific examples and categorizations of AI use, effectively illustrates its practical value.
• Balanced Perspective: You've struck a fine balance between acknowledging the broader concerns surrounding AI and emphasizing its utility in enhancing productivity. This approach invites readers to consider both sides of the AI conversation.
• Humor and Accessibility: The humorous tone and accessible language make complex topics more understandable and engaging for a broad audience.

Areas for Improvement:

• Exploring AI Limitations: While you briefly mention some limitations and concerns, a deeper dive into these challenges, including the potential for bias, privacy issues, and the need for oversight, could provide a more rounded perspective.
• Broader Implications: Expanding on how your experiences with AI might translate to other industries or roles could offer valuable insights for readers looking to apply similar tools in their work. Discussing potential future developments in AI and their implications for work and society could also enrich the conversation.
• Visual Elements: Incorporating visuals, such as charts or screenshots of the AI tools in action, could enhance reader engagement and understanding, particularly when discussing data work and graphics.

Additional Suggestions:

• Case Studies: Including brief case studies or success stories from using AI in your projects could further illustrate its impact.
• Guidance for Adoption: Offering advice on how others can start integrating AI into their workflows, including recommendations for specific tools and strategies for overcoming common obstacles, would provide practical value to readers.
• Discussion on Ethical Use: A section dedicated to the ethical use of AI, including considerations on data privacy and avoiding misuse, could underscore the importance of responsible AI deployment.

Overall, your blog post does an excellent job of highlighting the "boring" yet transformative power of AI in enhancing work efficiency and productivity. With some expansions and deeper explorations into the areas mentioned, it could serve as a valuable resource for professionals across various fields looking to leverage AI in their work.

Gemini Advanced

Here's a breakdown of the blog post's strengths, potential improvements, and some overarching thoughts:

Strengths

• Strong Voice: Your writing is clear, engaging, and has a distinct personality. The conversational tone and self-deprecating humor ("ChatBEAD... perhaps not the most thrilling tool name") make the entire post enjoyable to read.
• Informative Examples: You showcase a diverse range of AI applications with relevant examples specific to your work. This helps ground the discussion in reality and provides inspiration for readers.
• Structure: The categories, subcategories, and graphics create a clear flow and make the information easier to digest.
• Balanced Perspective: You acknowledge the transformative potential of AI while underscoring its limitations and emphasizing that it's ultimately a tool within a greater toolkit.

Potential Improvements

• Intro Length: While the intro is great, shortening it slightly could get readers to the "how" of your AI use even faster. Consider removing the second paragraph ("But in the meantime...") to increase the punchiness.
• Depth vs Breadth: You cover a lot of ground. Assess whether a slightly narrower focus on 2-3 categories would create even more impact, allowing you to dive deeper into those specific use cases.
• Visuals: The included figure is helpful. Could additional visuals illustrating your points elevate the post even further? A simple screenshot of a well-formed prompt, or a before/after of AI-edited text, could add another layer of engagement.

Overall Impression

This is an excellent blog post! Some minor tweaks could make it even stronger, but the core message, writing style, and organization are already effective. Here's why this type of post is valuable:

• Relatable: It demystifies AI hype with practical examples. This is more accessible to a general audience than jargon-heavy pieces about complex AI breakthroughs.
• Useful Beyond Your Specific Field: Your focus on writing, data analysis, and research has broad relevance. Readers in other fields might see parallels and get ideas for applying AI to their own workflows.
• Inspires Experimentation: You present AI tools as accessible and imperfect – this may encourage those who might have felt intimidated to try them for themselves.

I'd definitely share this blog post - it's a great example of how to communicate about AI in a way that's both informative and fun!