If you blindfold yourself, spin around 50 times, and throw a stone, you'll inevitably hit an AI-based model. The hype surrounding this technology has left any sense of control behind, but beyond doubts about its quality and usefulness, the truth is that it has become much simpler to install in conventional environments. In fact, the Ollama project reduces the whole process to an initial download, a couple of PowerShell commands, and a few minutes of your time.
You might be wondering why I'm not covering AI-related news as before. The truth is we're going through a 'Me Too' phase that borders on the exasperating. The number of projects abusing the 'now with AI' label or similar is absurd, and in the vast majority of cases, their results are disappointing.
Add to that the immediate availability of the most important models, like ChatGPT (which removed the login requirement in its free version), Google Gemini, and Copilot (which Microsoft offers even for going to the bathroom). However, today I present a small act of rebellion, in the form of local, private, and offline models. Obviously, you'll need some firepower for them to run smoothly, but the software requirements are very simple. The first step is to download Ollama.
Ollama: An Effective Way to Run Offline AI Models
Once you've downloaded and installed Ollama, you'll notice a small icon in the system tray. This only indicates the environment is running. You can also go to http://localhost:11434/ and verify that the phrase 'Ollama is running' appears. Now, Ollama has no interface, but it doesn't need one: the 'chat' itself works under PowerShell.
I recommend opening PowerShell with elevated privileges, since Ollama will need to download models to ensure local execution. The full list is here, but Llama2 and Mistral are two interesting options to start. For example, running the command 'ollama run llama2' will automatically get a copy of Llama2. The average is about 4 gigabytes per model, so you'll need some patience and enough space on your drive. The model will be ready to use when you see the cursor, accompanied by the expression 'Send a message (/? for help)'.
So, how well does Ollama work? I admit the environment is very good, but the final quality of the sessions will always be tied to what each model can or cannot do. Both Llama2 and Mistral got off to a good start, however, in less than ten minutes they were already delusional with incorrect information. Mistral conjured up a U.S. project (Project ALFRED) based on RTGs in the mid-'60s that I couldn't find anything about. In contrast, Llama2 gave me two different dates for the general elections in Mexico... and invented candidate names. ¯\_(ツ)_/¯
In summary, as an experimentation tool, Ollama is solid. The models, however, are another story. Link below!
Official site and download: Click here
