I didn’t think I would like the AI build into the browser, but it can really be useful.
I can appreciate the privacy features of Leo to forget conversations.
I am a bit frustrated by Leo when certain conversations as it is very afraid of even referencing certain topics.
For context: I am researching toddler aggression and I would like references on the topic. Is there validity to the child’s feelings?
I would be willing to pay for an uncensored AI with the understanding that as with all AI, it does not represent an absolute truth or morality.
I really would like it if Brave was braver on this topic as I would imagine that your customers can handle the harsh realities of the internet. Censoring Leo is the beginning of censuring the views of the Internet. I would prefer that Brave was more open to a censorship free experience.
As I have mentioned, I am willing to pay a subscription for a model that has the ability to reference all aspects of a topic, and not a childlike view of morality to box me in.
I am disappointed that Brave would feel that I need to be protected. Please treat me like an adult.
I pay for Brave because I appreciate the vision that you are creating for users. I want you to succeed. Please don’t compromise on censorship, or think to become the arbiter of truth, or you will become no different than an underdog Google.
In some other thread someone quoted tweets by Brendan Eich (I hope I got the spelling right) that they are considering training their own models.
I would also love to have completely uncensored and non-moralizing LLMs available at least as an option though I don’t know what legal challenges this would have or if it would pit advertisers against Brave. I hope it won’t be a challenge.
I think it would be amazing to have a personalized Leo, trained in your personal browser history, kept private with the amazing benefits’.
Oh I bet that would be too expensive for now. IIRC OpenAI was charging more than 4x the price of their regular model if you wanted answers from a model trained on custom data.
But what Brave could do is what ChatGPT already allowed: a customized prompt so that you can add information about what you expect from all conversations. In fact if we were allowed to easily edit this file without having to compile our own Brave it would be amazing for experimenting with customized assistants.
The paid version reads and can summarize the page that you are navigated to.
I have experimented with it by navigating to a local web page I have created while studying for the CISSP. It can provide feedback on the outline in HTML format, but it bugs out when I feed it a longer HTML document (my notes) or scholarly articles online.
It would seem to me that although the function would be more computationally expensive, It is quite feasible with the potential for customization.
There are two different things that you can do to get answers based on some arbitrary piece of text:
-
Train the LLM also on that text. This is very costly, and takes enough time to rule out doing it on the fly, but the LLM will have “innate knowledge” of the text regardless of what you do next.
-
Give the text as context to the LLM when asking about it. All LLMs do is that you give them a block of text, and they give you probabilities for what’s the next word. Then your code around the LLM choses either the most probable word as your output, or use some semi-random method to chose form the N most probable ones or something like that. Then you repeat the process (with the new output placed inside the context) and get another word until you reach an “end of reply”. In reality it doesn’t work on the level of individual words, but of tokens, which are often smaller than a word (Quoting the internet: A 750 word document will be about 1000 tokens). Most LLMs have just a few KBs worth of memory for the context. So what Leo does, is that it puts in the context of the LLM some ground rules, like what time it is, that it’s talking to a user of Brave, and that it should be helpful and not rude. Then it puts a sentence that says “The user is reading this article:” then it puts the contents of the article. Then it puts the conversation history till that point, if any, marked as such. Then it puts the question you posed (your last message), marked as such. So the problem is that if the page/article is too long, it won’t fit in the context. You can do tricks like as for summarizing chunks of the text and then combining the summaries and then using that summary as context for the AI but this would mean both a greater delay and greater cost for starting a conversation with the LLM.
Option 2 is what I would have thought would have been the default. I think most would be willing to wait a couple of minute for a summary of summaries.
As you said, I could manually build the context, but that would take time and significantly reduce the value of the summarizer.
what’s the best way to disable LEO, from agreement page it says the LEO llama source is built by META, and META will retain all data it receives from LEO, while the data encrypted from browser to server, who knows what occurs with your data after META has it. I’d prefer to be able to disable it so it isn’t reading my pages, otherwise I wont’ be able to use it with work as that data is proprietary.
I really want to like Leo, but I am cancelling the monthly subscription as I feel it compromises the values that I initially supported Brave for.
I really, really don’t like directed censureship, and withholding available information that is requested is a version of censureship. That reason is why I support (PAY for) Brave Search and not DuckDuckGo. Neutered Leo is a violation of this principle under the argument that I, a paying adult, cannot make a moral judgement based upon information available.
I do not like morality injected into LLMs as I think it is a form of behavioral modification.
I want LLMs to function as an intelligent search and summarizer. That’s all!
If having an LLM built into the browser means Brave had to compromise on privacy and censureship, then we have a problem.
I will have to determine if Brave is changing the principles they are built on, and if they are, then I will find another alternative.
It’s expensive to train LLMs. And I bet that it’s even legally risky, and potentially risky for creating a PR disaster to train completely uncensored LLMs. I would love to see some in the future, but I wont blame Brave for not providing one at the moment.