LLMs, in their current form, are mid. They are better then bad, but not as great as perceived.
If you use LLMs in your area of expertise, you likely ran into examples of their “midness.” You must escape the tech demos that make you say “wow, AI is going to take over the world.” The areas that matter are where people actually use them. In these areas, a lot of their output is just ok.
I see this is writing. The writing LLMs produce is functionally unusable for writing people actually want to read. It creates writing passable when forcing someone to read it (like a high school essay), but we live in the real world, and this doesn’t cut it. For real-world writing, LLMs miss a bunch of aspects:
- They’re overly wordy and cliche.
- They rely on worn-out examples.
- They don’t structure paragraphs or sentences well, most are too long and uniform.
- Their responses are slow. I’m functionally faster than an LLM because I write a sentence and I’m done. For the LLM, I wait for the full text, read, understand, and edit it to complete a section. This takes much longer in reality.
Although I’m not an expert in other domains, I’m sure you can create a similar list for those too.
It’s ok to be mid
Contrary to how it might seem, this isn’t a post about how LLMs suck. The ability for LLMs to be mid is a transformative innovation. It turns out there are a ton of areas that only need mid, the world runs of people doing a “good enough” job.
For example, I regularly write small apps in languages I’m unfamiliar with as examples. They don’t require perfect code. LLMs help me write a lot of mid code and this dramatically increases my productivity. I could read the docs and do tutorials to learn what I need, but it’s a lot faster to use LLMs to get me there.
An LLM is like a friend who studied a subject in university 7 years ago and hasn’t used the information since. They can help you out with the basics, but for in-depth thinking or the latest developments, they’re not going to be very good.
If you multiply this across every industry every person is unfamiliar with, you can see the boost mid responses provide. When people faced a task in an unfamiliar area before LLMs, they needed to learn enough to become mid themselves. Using LLMs saves a lot of time and energy.
The proliferation of mid
Because mid functionality has a large impact in many areas, LLMs are proliferating. There is a massive wave of “LLM wrapper” startups (and discourse). Importantly, many of them are proving to be solid businesses.
- The popular AI writing company, Jasper, had revenues of $75M in 2022.
- Copy.ai hit $10M in ARR in October 2022.
- GitHub Copilot does an estimated $48M in revenue a year. They are proving there is a market for mid.
These are just the most obvious implementations too. They will proliferate into other industries where “mid” work is passable. For example, instead of needing to read research or reports, an LLM can read them for you and give you a summary. Instead of writing and posting job descriptions, let an LLM do it for you.
A big advantage of LLMs as new, exciting technology is that there are many companies (and individuals) with the budget to try them out. They created a shock and opportunity to capitalize on. Many companies without money for B2B SaaS tools with similar functionality 2 years ago suddenly have a budget to try out LLM wrappers (same goes for VC investment).
This doesn’t mean software adoption is easy. Much of the economy still runs on no technology or old technology. LLMs aren’t going to magically solve the fact people are stuck in their ways, don’t see the benefits, or must fight through review, procurement, and onboarding processes.
The benefits of mid LLMs are huge, but people are still stubborn, and LLM-related companies still run into these issues. As transformative as they are, their midness isn’t enough to break human nature.
What does a world of mid LLMs mean to you?
My basic career plan is to keep writing and coding for the rest of my life. The first week after ChatGPT’s release, I reconsidered everything. The benchmarks and demos coming out were incredible. I calmed down when I used it and realized it was mid (and it has only gotten more mid since then).
Importantly for me, being an interesting and informative writer is more valuable when the majority of content is mid. For coding, planning, structuring, and writing a high-quality app is more valuable for the same reason. As long as I continue to develop as a writer and software developer, and leverage the tools when I can, my plan doesn’t need to change.
As mid LLMs swallow what they can, specialization becomes more valuable. This is because systems get complicated and standards become higher. Strategy, design, structure, and focus are all things mid LLMs don’t do well. Simply, taste remains important.
The other benefit of LLMs is that they help your productivity, especially as an expert. You can get the stuff you don’t want to do done faster. You can focus on what you are good at. Tool wrangling for your domain will continue to be a valuable skill.
Mid LLMs push people towards specialization which requires a higher work quality to succeed. This is a sign towards a post-scarcity world, and also a good sign. Increasing specialization is good for people who want to pursue their passions and inclinations. The greatest specialization requires curiosity, and I’m excited for more people to have the opportunity to discover this thanks to mid LLMs.
Let me know what you think on Twitter.
Ian Vanagas 