Original video in Chinese.
Key Takeaway
- Brave browser connects with local large language models such as Ollama to deliver built-in AI features, making web browsing more efficient and convenient.
- Brave’s AI features support direct conversation, summarizing selected text, and using the current webpage content as reference for the large model.
- The article emphasizes the necessity of connecting browsers with AI and argues that Brave is on the right track overall, but future follow-up from big players like Chrome and Gemini will bring competition.
- Although Brave’s AI features are still in testing, its local deployment and combination with open-source large language models give users more choices and freedom.
Full Content
All software should connect with AI. For example, the browser we use most often.
When you’re surfing the web and suddenly want to ask something or generate something, having to deliberately open another webpage (like ChatGPT) or switch to another app (like Obsidian) is especially troublesome and kind of annoying.
So Brave browser came along. This product has probably been around for several years. I won’t even mention the usual features like ad blocking. What attracts me most is that it can connect with local large language models, for example by running Qwen on your own computer, and then implementing the kinds of scenarios I just mentioned.
If you want to give it a try, remember to download the Nightly version. This is the test version, updated every night. After downloading, just install it step by step. Once that’s done, go to the settings page and click Leo—this is the name they gave their AI Assistant. Then, in the settings for adding a model, link it to Ollama.
Ollama is currently a very mainstream tool for running open-source large language models locally. You can go to their official website to download the software, and then use a single command in the terminal to download the large language model you want, such as qwen2, which I use most often.
Once the software and model are ready, open Ollama and you’ll see that it lives in the status bar. Whenever it needs to call the large model, it will start up. The advantage is that it can stay running all the time; the downside is that the first launch may take around 10 seconds or so, because it has to load the large model. Once it’s up and running, it’s extremely fast—really much nicer than the cloud.
It’s very simple to connect Brave with Ollama. If, like me, you haven’t made any configuration adjustments to Ollama, just follow the prompts and fill in the address and the name of the model you want to run. As long as the name matches, there shouldn’t basically be any problem.
Back on the frontend interaction page, you can open Leo AI from the sidebar. The official version provides large models, but here we want to use a local one, so choose the Qwen model we just set up. Direct conversation is completely fine. If you want to combine it with the webpage you’re currently browsing, there are two ways:
First, you can select a portion of text on the webpage, and then right-click to see the official preset AI functions, such as summarization.
Second, after checking the button above the input box, the current webpage you’re browsing will automatically be provided to the large model as reference for the answer.
However, I found that this is not the same processing method as the RAG approach we usually use. Because I tried it, and if it’s a webpage with a long article, it will prompt that only part has been read, which shows that its method is very simple and crude: it just sends all the content as context. Only when it exceeds the context window limit will it say it hasn’t finished reading.
But then again, if it really used the RAG approach, you’d also have to deal with Embedding and so on, which would make it especially heavy and unsuitable for the web browsing scenario. Because users will keep opening new webpages and switching back and forth.
At present, Brave browser’s AI features are still in the testing phase. As I said earlier, I think connecting browsers and AI is extremely necessary. Brave’s overall direction is fine. However, big companies will definitely follow, such as Chrome, which will absolutely implement the same kind of functionality through Gemini. For a product like Brave, one way to survive is to give users more freedom. I especially hope they can add more customization features later, and at the very least, put out the model prompts first. That would definitely increase daily usage frequency.
OK, that’s it for this issue. If you want to find me, come to the newtype community; I’m there. See you next time!