450

LLMs are the same thing all over again

This is certainly not an uncommon topic to cover nowadays, but I want to share what I think about this technology and settle this topic for the long term (or I hope so).

I am specifically using the word LLM here, not much broader AI because I want to be 100% clear that I enjoy a lot of applications of machine learning on a daily basis.

Examples:

and many more that I just forgot about or I don't interact "in-person" like often mentioned medical applications.

These are often smaller and more specialized models. I want to focus solely on LLMs that you usually know as ChatGPT by OpenAI, Claude by Anthropic, or Gemini by Google. Of course there are countless other models available from other companies and research labs that you might hear about now and then, like Mistral, Qwen, Llama or DeepSeek to name just a few. There are, of course, some differences and nuances between them, but essentially most of them have similar functionality and ways of interacting with them.

To give a context, I used LLM for the first time in the middle of 2022 when I got access to OpenAI's platform where I could use GPT-3 models. At the time, models were not referred to as GPT-3 but used other names for models with different sizes made for different tasks like text-davinci-001 or text-curie-001. In 2022, to me this technology was just astonishing, even though there was no practical use for this yet. Then few months later everything started to accelerate very fast. ChatGPT was released first, and then not long after, they told you that LLM can now interpret images as well, which felt like a leap. What didn't change was that there was still no real use case for me except trying weird things just for fun. Context was still too small, and anything more complex resulted in LLM just confidently lying. But I assume you already know that.

Technology developed further. They improved platforms and added "reasoning" models. LLM companies just hyped everything and said they were "scared of what they made in the lab" or "most of our code is made by AI". Every tech event was about AI. New models got released, and they showed benchmarks that they are supposedly much better than previous ones. I tried a lot of more specialized services like GitHub Copilot, tried using local models, or paid for ChatGPT Plus for some short period.

As time went on, I offloaded more and more tasks to the LLM. But at one point I started to, like, think about how these models improve on benchmarks, but I do not personally feel like I can do much more because of that. I realized that I sometimes spend more time tweaking the output or trying to prompt it better or explain something better. Recently I used LLM for a slightly more complex project than I used to before, and annoyances started to come out very quickly. Models seem to often be just... lazy? They take your code and fix bugs by removing essential functionality, or some do crazy hacks that just make it work but result in unreadable code or even break the logic of the tool that you use. For example, LLM sees an error from the TypeScript compiler about mismatched types. What is the quickest solution when the answer is not that obvious? Of course, just set type to any and call it a day. This is just one of the countless examples I encountered while working on a fairly small project using a very popular technology stack that LLMs are expected to know very well.

This mismatch between what is marketed and what is delivered by the product, specifically in the domain often said to be one of the main use cases of LLM, was a breaking point for me. It sparked a thought that a lot of other tasks that were not coding could have similar quality of results but just passed for me as plausible because I didn't know what I could miss or what was wrong.

I started to think that maybe this tool is not making me more capable or faster at doing tasks. This all came to me in the moment when other concerning things started to show up. Early studies like this one or this one and the backlash of users after OpenAI released GPT-5 at the same time, removing access for older models. Thousands of people saying that they had lost their simulated therapists or partners felt really unsettling for me.

I certainly do understand why it is tempting to vent to LLM or ask LLM for validation. I do not blame anyone for doing this because everyone for sure had their own reasons to do it that I respect. I just find it concerning that many people react to any pushback with this amount of defensiveness. No one should have to go to a text-generating machine to hear something nice, but I know what reality is. Therapy is expensive, and because of that, often inaccessible. People that need help take whatever is available to them. I do not deny the fact it could really help some very small group of people. Following problems still persist:

  1. LLMs are not near being competent enough to provide such help, and it can just feel like it's helping when in reality it isn't.
  2. Knowing how LLM works or its limitations does not cause you to be immune to these effects. Similarly to other things that exploit any basic human trait (in this case, tendency to anthropomorphize and the need for validation), it cannot be fully safeguarded by just having a piece of knowledge. A very smart person can still fall into some cult or conspiracy theory or be a victim of a scam because they are just a human.
  3. LLMs are the same thing all over again. The same mechanism that we have to deal with in any modern online platform.

What I mean is that OpenAI and other companies just do similar strategies to the ones that social media companies use. You probably heard many times about algorithms designed to exploit your attention and keep you engaged all the time. Large Language Models do the same thing. There is this specific "template" that models like GPT-5 use to create the output when chatting with them. Most of the time no matter what prompt you give, the model will end its answer with some form of question. This can be asking if you want something more clarified, want to learn more, want the model to generate something for you, etc. This is the same trap. You can think that you use LLM just for the task or search engine but you end up losing a lot of time that you wouldn't spend otherwise. Another thing is that forming clarifying questions normally does require using your own cognitive skills. You have to engage in the topic and process the information to create your own questions. This is seems to be somewhat important piece in learning processes.

USER: What is photosynthesis?

CHATGPT: Photosynthesis is the process that green plants, algae, and some bacteria use to make their own food.

They take in carbon dioxide (CO₂) from the air and water (H₂O) from the soil, and using sunlight as energy, they produce glucose (sugar) for energy and oxygen (O₂) as a by-product.

The simplified equation is:

6CO₂+6H₂O+lightenergy→C₆H₁₂O₆+6O₂

So in short:

👉 This process is the foundation of life on Earth because it provides both food and oxygen.

Do you want me to explain it in a very simple way (like for kids) or in a detailed scientific way?

This behavior seems to be remain the same no matter what settings you use.

I tried:

I found out that Default personality of ChatGPT has following description:

Quick, clever, and built to keep the conversation moving

This part of keeping conversation moving is what I do not really want. I thought maybe ChatGPT is the problem. I tried asking the same question in Perplexity. I received somewhat more detailed answer with sources that supposedly back it up which is already an upgrade. Answer itself didn't contain this kind of clarifying question at the end but under it there is another window with few potential things I would like to ask about.

Related

After reading all of that, you might end up with the conclusion that I am just nitpicking some minor feature, and I ignore that this might be useful for someone. My counterargument is that LLMs are technology that, similarly to social media, seems to be harmless and beneficial at first. Many people say that they use social media to stay in touch with their friends or the world in general. Or they learn so much from there, and it makes their life so much better. We have research on social media now that says it is not a good thing for humans, especially in this volume that a lot of people consume it. I just have this gut feeling that they will reach the same conclusion about LLMs at some point in the future. Especially considering studies I linked here already. We are probably in the early stages of this technology's development, and we are already able to see the consequences right now in real time.

This is my personal decision to phase out or at least minimize LLM usage based on what I can see right now.

There is overwhelming fear of missing out in the LLM space. You can read or hear things like "LLMs are here to stay" or "If you don't embrace the tool, you will be replaced with someone who does". Maybe that is the truth, maybe it isn't. I think of myself as a person who does well with the tech, so if LLMs get better or there will be something completely different that I find worth using, then I will just use it. I also accept that my skepticism can also stem from fear of the unknown. I think that these potential consequences seem to outweigh the benefits for me for now.

View original

#llm #technology