Getting locked in

Like many people who work in bioinformatics, my background was in biology and had nothing to do with computers. As I worked longer in this field, my responsibilities grew from writing code and performing analyses to setting up and maintaining servers (and the many unofficial tasks that an IT help desk are supposed to do, like connecting to printers, email servers, and network drives). This is fine because I am genuinely interested in computers and I like helping people when I can.

As I started to learn more about computers and the management of them, I would have these moments where I go "Ahhhhh that's the word/concept for it!". An example is technical debt, which I believe is widely experienced in academic labs, where the PhD student and/or postdoc is rushing to generate results for their thesis and/or manuscript and forgot to document their work or write better code. Related to code, even since I learned what spaghetti code meant, I started to refer to my own code as spaghetti code. Another is bus factor, which I constantly tell people about because I find it morbidly funny. And then there's the one that everyone has experienced if they have tried to install bioinformatics tools: dependency hell (though there is definitely much less of this nowadays!). There's a lot more of these terms but for now I wanted to write specifically about vendor lock-in.

I believe everyone is familiar with vendor lock-in but like my past self, they didn't know the word for it or perhaps didn't realise that they were locked in. It is simply when someone, like a company, creates something, like software or a tool, that makes you fully dependent on that something. One example of creating a vendor lock-in situation is to create files that only your software can properly process or process at all. There are of course legitimate reasons for creating proprietary formats but sadly they get abused by large tech companies. And of course once you are fully dependent, you are at the mercy of the tool provider. Ever since I learned about vendor lock-in, it has changed the way I select tools (and to a certain extent, the way I work) and most of the time I simply check whether a tool is open-source or not. If I have to use something proprietary, I make sure that there's an alternative in case I need to switch. A good rule of thumb is before you adopt a tool, ask how hard it would be to leave it.

The day I tested out ChatGPT for the first time, I was amazed and wanted to use it to learn more about everything! To this day, that is still my number one use case for LLMs: studying. But I had realised that despite having the word "open" in its name, OpenAI was still a company after all and didn't disclose much about their models right from the start. Since I didn't want to get locked in, I looked for open models and for ways to run these models myself. The first alternative I found and tried was GPT4All, which I couldn't get to work, so I gave up. Some time later, I found Jan and actually didn't expect it to work at all but it did and with minimal setting up! Then one day DeepSeek was all the buzz and of course I wanted to test it out. I found a tool called Ollama, which has become my preferred tool for running open models, and it was super easy to set up and run DeepSeek (and other open models). I own very modest hardware, so the output is quite slow for the larger open models, in case you were wondering.

What spurred this blog post was a very recent 404 Media article titled The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI. If you don't know 404 Media, I highly recommend them for independent and unbiased reporting on tech! If you follow AI closely, you will know that the two biggest players, OpenAI and Anthropic, basically lose money each time someone uses their service. This is obvious for users on free accounts but even paying customers don't cover the cost of maintaining and running LLMs. You may wonder how this is possible because isn't the sole purpose of businesses to make money? And of course this reminds me of the Bad Money episode of Silicon Valley (a parody show that hits too close to home):

It's not about how much you earn, it's about what you're worth. And who's worth the most? Companies that lose money. Pinterest, Snapchat, no revenue.

Perhaps this is how blitzscaling is supposed to work? And perhaps this is all fine because AI is the future and the money will come later? Well, it seems like you can't continually burn money because prices for tokens are going up, which is what the 404 Media described as the Tokenpocalypse. (A token is simply how the LLMs represent text and essentially it's the input and output of a LLM.) Perhaps these AI companies have data to support that now is a good time to raise prices because they have gotten enough people locked into their closed frontier models and/or tools? Whatever the reason for the price increase, if you have become fully dependent, then you'll have to pay up to continue using their services as you did before or pay the price to switch.

As I wrote near the start of this post, my biggest use case for LLMs is for studying. I need to understand computing, statistics, mathematics, biology, and more for my work. Wikipedia is great but quite often I struggle with the maths and stats and I find that LLMs are great for elaborating on things I don't fully understand. When I do use LLMs to help with my coding, I am deliberately slow, because I always look through each line of code and ask it to explain every piece of unfamiliar code. Because just like how I don't like to be locked in with tools, I don't want to be locked in with AI as well.