“There’s no reason to use GPT-4, Claude3. Plus, it’s free.”

I am an open-source admirer, but I don’t think the advantages below will last very long.

“There’s no reason to use GPT-4, Claude3. Plus, it’s free.”

Developers are not only more expensive, but starting to manipulate open sources means that I take full responsibility. Some want to have responsibility and authority, while others do not, efficiency and cost decline due to accumulation cannot be ignored. LLAMA3 lowered the cost by paying for training on Facebook and passing only fine tuning to individuals and businesses, but sooner or later, the applications that applied them will be fragmented and the cost of creating integrated applications may be unexpectedly high.

It’s not that either side is wrong, but that’s what software is originally like, and each has different strengths and weaknesses.

There are inexpensive, inexpensive, and sovereign NAS solutions, but we mostly use OneDrive or iCloud. Upgrade of ChatGPT5 will be easy, but upgrade of custom LLAMA4 implantation software will be quite a hassle.

However, as an open-source admirer, LLAMA3’s advantage is its “degree of freedom” and its expansion power in the market. I loved this advantage, and all the people in the company did the crazy thing of purchasing and using Andro One with the earliest version of Android, Cupcake (1.5). Freedom is good, but it will not be free to get the perfect experience you want, and I think the release of ChatGPT5 will turn the mood upside down again.

More than that, I think LangChain is an object. The performance of the AI engine is enough, and it seems that we have moved to a point where how to build and use the SW ecosystem well becomes more important. In that sense, it is also true that the open source is at a pretty good point. The reason for using Meta LLama so far. I use it because the source is public. I use it for security reasons. For these two reasons, I have used LLama until now. In the future, I will say that it is used for excellent functions in addition to two reasons.
Until now, the reason why I used GPT-4, Claude3 is that I don’t have a source, I don’t have security, but I spent a lot of money and wrote it because it has a good function.

However, it won’t have to be like that anymore. There’s no reason to use GPT-4, Claude3. Plus, it’s free.

Without GPT-5 in the coming months, the LLM’s axis is expected to suddenly tilt toward LLama. This has been a hope for Korea. This is because LLama can fine tune into Hangeul data at all costs, even though Hangeul is not working well.

I turned LLama38B around on my office laptop. It’s a little slow, but it works the same way. The result was surprising. LLama3 had excellent reasoning ability. They are good at code generation. They also have excellent mathematical ability. LLama38B is very equal to or slightly better than the reasoning ability of other large foundation models. The reason this is very important is that asking and answering chatbot-type questions works well if you fine tune a lot of Korean data. However, reasoning is possible only when the original foundation model is good. Because fine tuning does not improve. If reasoning is good, it also means that you can create AI Agents based on LLama in the future. So far, almost all of them have made AI Agents based on GPT-4. So it means that LLama3’s AI Agent function is expected in the future.

It is believed that this will certainly advance the era of On Device AI in the future. This is because the LLama38B is better than all 8B models released so far. In the future, the fine-tuned version of LLama38B may be uploaded to smartphones, PCs, notebooks, and automobiles. Korea has done a lot of fine-tuning with LLama, so LLama is a very familiar foundation model. This is because all LLMs in the Mistral and Solar families are LLama-based.

I’m really looking forward to the upcoming LLama 400B, fine-tuned LLM, etc.