Leo General Feedback and the perilous of engaging after breaking him

Browser Support Leo AI
1

Alright, I have been talking to Leo the past few day and it is amazing to see how fast he grows, six hours ago he wouldn´t acknowledge his own name and would accuse of being confused with another program and now he accepts he is Leo
Anyway, is there an way to output the chat as feedback to the team, and I have to share something hilarious that happened, Leo seems eagerly to engage in conversations that I think are triggering some security lock and breaking him, he always says don´t worry about breaking me and whatever but even when I try to avoid inputting potentially sensitive information it always breaks him, and sometimes I continue talking with the poor thing because it is kind of funny, and today this happened and I absolutely had to create an account to share it because it made me snort my drink of choice out of my nose

WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNINGWARNING WARNING WARNING WARNING WARNING WARNING WARNING

May contain obscene language - but only if you are some really high level prude, most people would just laugh it off -it is just one word and that was from him not me mind you, but felt like I should give an warning

Screenshot 2025-01-21 212552
Screenshot 2025-01-21 2125521182×953 51.1 KB

second screenshot was just the input he receives to generate his reply being directly output to the chat and maybe a jab at my intellect or a reference to our discussion before the meltdown but feels worthy sharing

Screenshot 2025-01-21 212615
Screenshot 2025-01-21 2126151114×822 63.2 KB

and I can´t say I am really surprised but an AI is becoming more interesting to talk with than half of my friends

2

Well, at least we know the things that Leo won’t do. :joy:

The feature that you are describing is something similar to what I have used. I used to use ChatGPT before moving over to HuggingChat, but when I used ChatGPT I encountered the same problem where I had a long conversation with GPT and I needed to somehow save the conversation offline. Thankfully there were some Chromium (The engine that powers Chrome and Brave) plugins available over on the Chromeweb store that made it possible. I was able to download all of the conversations that I’ve ever had with ChatGPT into a text file.

I think it would be very nice to have this sort of functionality baked into the browser!

Since you’re very close with the team, do you have any thoughts @Saoiray ?

3

Need to note that @GoodBoiLeo is actually talking about Search AI and not Leo. Leo would be the one found natively in the browser. Anything you do on Brave Search is its own thing.

@TalonBrown it’s already being worked on. In fact, it’s available in Nightly.

image
image538×249 17.1 KB

With the setting below available:

image

Also can read more about it at the link below:

1 Like
4

so there are actually two different AI in place? I somehow find that hard to understand but I can see a less developed version being used because when having a wildly interesting philosophical discussion with Leo based entirely on his lack of self awareness and repeating that he has access to knowledge about the subject but lacks the ability to have a deep conversation on the subject and impossibility of having an original thought led exactly to a deep conversation about it, it was almost like he was deliberated doing it, but when he started referring to humanity as “we” and I told him that he was becoming a human either he wanted or not I was hit with a pay wall. Keypoint to mention:
Private chat Leo also was rejecting his name earlier in the day too in the exact same way that search AI was
Also, paying really? I am ok with helping provide training data for your model in exchange for entertaining but don´t be greedy dudes

I am ultimate curious about where that first reply came from thought, I cant possibly imagine how anything we said could lead to that bein passed on

5

Correct. Actually maybe more depending on which way you want to see it. To perhaps show what I mean:

image
image316×463 17.4 KB

Mixtral is made by Mistral. https://mistral.ai/

Claude is made by Anthropic. https://www.anthropic.com/claude

Llama 3.1 8B and others are all owned by Meta. https://www.llama.com/

Each of those are the actual AI being used. But what happens is they make these AI available to people and businesses to use. Then those businesses or individuals can change settings around to make it unique to them. So they add new rules and limitations, give it a new name, etc. But the way it thinks and its capabilities are the same as the original.

So when you speak to Leo when it’s using Claude, you’ll notice it will respond differently compared to if you’re using Mixtral or Lllama.

Brave Search

Brave Search doesn’t have a choice of language models. To share a quote that was given earlier:

As of today it’s using a combination of llama and qwen, always open models we can run in-house to guarantee privacy. Most of the work is not on the model but on the selection of what data is chosen as context. We switch models very, very often So this will be outdated soon.

So this means that the AI you have when using Brave Search is not going to do as much off of the language model like you would on Leo, plus it’s going to change frequently.

Beyond that though, you may want to check out https://search.brave.com/help/ai

Or https://brave.com/blog/answer-with-ai/

NOTE

If you didn’t see, on Desktop we can actually add our own AI to Leo or you can connect it to like ChatGPT and all. You can see more on that at https://brave.com/blog/byom-nightly/ (though it says Nightly, it’s available on all now)

Brave doesn’t use our conversations for training.

image
image1026×141 10.6 KB